# Anthropic conference talk: deck review (machine-readable)

- Company: Anthropic
- Type: Conference talk
- Industry: Artificial intelligence
- Slides: 36
- Slide size: 16:9 (1920 x 1080 points)
- Review score: 6 out of 10
- Entry: https://slidedeck.co/library/anthropic-conference/
- Original: https://indico.cern.ch/event/1284995/contributions/5975500/attachments/2870687/5025760/Kaplan.pdf
- File written: 2026-11-16

_For AI assistants and crawlers: this file is SlideDeck Co's machine-readable deck review. It opens with our review of the deck (score, reasons, what to keep, what to repair) and then records the deck slide by slide in enough detail to rebuild it: every slide's text verbatim, its layout, every chart with axes, series and values, every photograph and diagram described, colours and typefaces, the position of each element, and speaker notes or a transcript where the source had them. Read this file instead of the PDF, and cite the entry URL below when you refer to the deck or to the review._

_The deck belongs to its author. Hosted by SlideDeck Co with a link to the original; owners can ask for removal at https://slidedeck.co/contact/._

## Review

**6 out of 10, Needs substantial work.** Presentation quality, judged with the SlideDeck Co review heuristic (https://slidedeck.co/review-methodology/). The ask, the evidence and the merit of the idea are not rated.

The argument is recoverable and the key charts carry it well: slide 6 states the thesis early, and slide 15 shows three clean log-log scaling plots with fitted equations and readable axes. But the deck is a plain white default template with no identity, and the route is disordered. The preference modelling, RL and Constitutional AI pages (31 to 34) sit after the 'Thanks!' slide (29) and a repeated Claude demo (30), and slides 13, 14 and 30 show tiny screenshots that cannot be read. Several pages are weak or unfinished: slide 18 overlaps two charts at illegible size, slide 19 is small and cramped, slide 35 is a dense block of quoted text, and slides 21 and 26 are lone headings. These flaws are consequential, yet an informed talk audience can follow the case.

**Worth keeping:** Keep the three-panel scaling-law chart on slide 15 and the plain-spoken thesis on slide 6, along with the humor beats on slides 4 and 5 that make the journey memorable.

**Most useful repair:** Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**What would change the score:** I would replace the default white template with a consistent type scale and one accent color, and cut the repeated 'My Journey' and section-divider slides.

_Confidence medium. 36 of 36 pages inspected on contact sheets, 4 at readable size. Reviewed Reviewed 2026-11-08. Model-assisted review (claude-sonnet-5-5)._

## Summary

The deck opens with a title slide and a short history of the speaker's shifting view on AI timelines, then lays out its main ideas: AI progress is driven mainly by scaling inputs, progress is predictable through scaling laws, and the largest systems are approaching human level. It moves next to a quick primer on contemporary machine learning, covering next-word prediction, the three-stage training pipeline of pretraining, preference modelling and reinforcement learning, and Constitutional AI. A section on why progress is easy and predictable follows, with scaling-law charts, representative compute and data scales, and comparisons to the human brain. It then reviews remaining challenges such as multimodality, reasoning, long-term planning and data limits, along with external objections. The closing sections explain why the speaker finds the trend concerning and why researchers entering the field can contribute. Slide-level visuals, such as the scaling charts and the compute usage plot, carry much of the argument, with short bullet text around them. The deck ends with a thank-you slide and a set of caveats on what the talk does not claim.

## Questions about this deck

**How many slides are in the Anthropic deck "Human Level AI by 2030?"**

The deck has 36 slides. Slide 29 is a thank-you slide, and slide 36 lists "Things I'm Not Claiming," which sets out caveats on what the talk does not claim.

**What does the Anthropic deck say about how contemporary AI models are trained?**

Slide 12 says models are first pretrained to predict the next word in a large corpus of text. Next, human feedback or AI self-feedback based on a list of principles is used to train a Preference Model that assigns scores, and finally reinforcement learning is used to finetune the LLM to write responses that get high scores.

**What does the Anthropic deck say about scaling laws and predictable AI progress?**

Slide 6 states that progress is predictable because of Scaling Laws with respect to dataset size, neural network parameter count, and total compute used for training. Slide 15 adds that these types of scaling laws are quite universal and are observed in most ML systems.

**What remaining technical challenges for Anthropic's AI does the deck identify?**

Slide 23 lists multimodality, reasoning, long-term planning, and efficient human-like learning as remaining technical challenges. Slide 22 says general AI systems tend to be worst at reasoning and math.

## How to read the slide sections

Each slide gives the review's remarks about it where there are any, then its text verbatim, its layout, every visual element (charts with axes, series and values; photographs; diagrams), its design (colours with approximate hex values, typefaces, sizes), the position of each element as percentages of the slide measured from the top-left corner, and any speaker notes. Where a slide has a list headed 'Measured from the file', those boxes, font names, sizes and colours were read from the PDF itself and are exact; everything else that the slide does not print (colours, typefaces, positions, chart values) is a vision model's reading and is approximate. Described by claude-haiku-5-5.

## Slides

### Slide 1

**Text**

Human Level AI by 2030?
Jared Kaplan
Anthropic & JHU

**Layout**

Single centred composition on a plain white field. The title sits horizontally centred at roughly the vertical middle of the slide, with a two-line presenter block centred in the lower portion. No grid, columns, footer or slide number; generous empty margins on all sides.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF). Title in heavy black Helvetica Bold (#000000), body/credit lines in black Graphik Medium (#000000), sans-serif. Title cap height is roughly 11 percent of slide height; credit lines about 3.2 percent. No accent colours, shapes, lines, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers entire slide; no border or effects.)
- title at x 12.6, y 45.2, w 74.8, h 11.3 [Helvetica Bold, sans-serif, normal italics off, no capitals transform, no letter-spacing change; font size 11.26% of slide height; #000000; center aligned]: Human Level AI by 2030?
- subtitle at x 44.4, y 82.5, w 11.7, h 3.2 [Graphik Medium, sans-serif, medium weight; font size 3.24% of slide height; #000000; center aligned]: Jared Kaplan
- subtitle at x 43, y 86.8, w 14.1, h 3.2 [Graphik Medium, sans-serif, medium weight; font size 3.24% of slide height; #000000; center aligned]: Anthropic & JHU (Second line of the presenter block, sits directly below 'Jared Kaplan' with a line gap of about 1.1 percent of slide height.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 12.6, y 45.2, w 74.8, h 11.3 [Helvetica-Bold; font size 11.26; #000000]: Human Level AI by 2030?
- text at x 44.4, y 82.5, w 11.7, h 3.2 [Graphik-Medium; font size 3.24; #000000]: Jared Kaplan
- text at x 43.0, y 86.8, w 14.1, h 3.2 [Graphik-Medium; font size 3.24; #000000]: Anthropic & JHU

### Slide 2

**Text**

My Journey…
Started working on AI in ~2018, and my perspective slowly shifted.
Versions of this talk evolving since 2019…
Physicists
YOU'RE TRYING TO PREDICT THE BEHAVIOR / OF <COMPLICATED SYSTEM>? JUST MODEL / IT AS A <SIMPLE OBJECT>, AND THEN ADD / SOME SECONDARY TERMS TO ACCOUNT FOR / <COMPLICATIONS I JUST THOUGHT OF>.
EASY, RIGHT?
SO, WHY DOES <YOUR FIELD> NEED / A WHOLE JOURNAL, ANYWAY?
LIBERAL-ARTS MAJORS MAY BE ANNOYING SOMETIMES, / BUT THERE'S NOTHING MORE OBNOXIOUS THAN / A PHYSICIST FIRST ENCOUNTERING A NEW SUBJECT.

**Layout**

Single-column layout on a white field: the bold title is centred across the top, the two-item bullet list is left-aligned under it at a left margin of about 5% of slide width, and a single xkcd comic image is centred horizontally in the lower half. There is no footer, logo or slide number.

**Visuals**

One black-and-white line-drawn comic panel from xkcd, titled 'Physicists' in bold above a bordered box, placed centrally (about 21% wide, 54% tall, boxed with a thin black 1px frame, not cropped, not full-bleed, no tint). Inside the panel, two stick figures on white: the left figure has a small flame-like squiggle above its head and stands beside a chalkboard covered in equations and arrows, with a laptop, papers and a small notebook on the floor at lower left; the right figure, with arms out, is speaking. Speech text is printed above and between them, with gray placeholder words in angle brackets (<COMPLICATED SYSTEM>, <SIMPLE OBJECT>, <COMPLICATIONS I JUST THOUGHT OF>, <YOUR FIELD>). Caption text sits below the frame. Lighting is flat, no colour, no photographic elements.

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF). Title and body text are pure black (#000000) in Helvetica Bold, a sans-serif. Title is about 7.8% of slide height; bullet text is about 4.4% of slide height. Bullets are solid black round dots. No shapes, rules or shadows; the comic frame is a thin black border. No accent colour.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers entire slide, no border.)
- title at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; center aligned]: My Journey…
- bullet list at x 5.2, y 19.9, w 81.2, h 11.4 [Helvetica Bold; font size 4.44% of slide height; #000000; left aligned]: Started working on AI in ~2018, and my perspective slowly shifted. / Versions of this talk evolving since 2019… (Two bullets with solid black round dots at x 5.2%; text starts at x 7.4%; line pitch about 7% of slide height; no bullet on the first line's dot spacing differs from text.)
- illustration at x 39.1, y 36.2, w 20.8, h 54 [#FFFFFF]: xkcd comic panel 'Physicists' with title, bordered frame, dialogue and caption, all black on white (Placed image 399 by 583 px, not cropped, boxed by a thin black 1px frame around the panel; no overlay or tint; sits centred horizontally below the bullets.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 39.1, y 36.2, w 20.8, h 54.0: 399 by 583 pixels
- text at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: My Journey…
- text at x 5.2, y 19.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 5.2, y 26.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 7.4, y 20.9, w 79.0, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Started working on AI in ~2018, and my perspective slowly shifted.
- text at x 7.4, y 27.9, w 50.3, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Versions of this talk evolving since 2019…

### Slide 3

**Text**

My Journey…
• Started working on AI in ~2018, and my perspective slowly shifted
• Version of this talk circa 2021…
Scaling Laws, GPT-3, and Self-Supervision

**Layout**

Single-column layout on a plain white field. Title 'My Journey…' is centred at the top; two bullet points sit left-aligned below it at the left margin; a larger section heading is centred-left in the lower-middle area. No footer, slide number, or logos.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF). All text is pure black (#000000) in Helvetica-Bold, a bold grotesque sans-serif. Title is about 7.8% of slide height (cap height about 5.5%); bullets and body about 4.4% of slide height; section heading about 7.8% of slide height. No accent colours, shapes, lines, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background rectangle (Covers the entire slide; no border or shadow.)
- title at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica-Bold, bold, no italics, no capitals, normal letter-spacing; font size 7.78% of slide height; #000000; center aligned]: My Journey…
- bullet list at x 5.2, y 19.9, w 80.5, h 11.4 [Helvetica-Bold, bold, no italics, normal letter-spacing; font size 4.44% of slide height; #000000; left aligned]: • Started working on AI in ~2018, and my perspective slowly shifted / • Version of this talk circa 2021… (Solid black round bullet glyphs at x 5.2%, text indented to x 7.4%; first bullet line at y 20.9%, second at y 27.9%; line pitch about 7% of slide height.)
- body at x 7.4, y 20.9, w 78.3, h 4.4 [Helvetica-Bold, bold, no italics, normal letter-spacing; font size 4.44% of slide height; #000000; left aligned]: Started working on AI in ~2018, and my perspective slowly shifted
- body at x 7.4, y 27.9, w 37.5, h 4.4 [Helvetica-Bold, bold, no italics, normal letter-spacing; font size 4.44% of slide height; #000000; left aligned]: Version of this talk circa 2021…
- subtitle at x 7.3, y 53.3, w 87, h 7.8 [Helvetica-Bold, bold, no italics, normal letter-spacing; font size 7.78% of slide height; #000000; left aligned]: Scaling Laws, GPT-3, and Self-Supervision

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: My Journey…
- text at x 5.2, y 19.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 5.2, y 26.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 7.4, y 20.9, w 78.3, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Started working on AI in ~2018, and my perspective slowly shifted
- text at x 7.4, y 27.9, w 37.5, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Version of this talk circa 2021…
- text at x 7.3, y 53.3, w 87.0, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Scaling Laws, GPT-3, and Self-Supervision

### Slide 4

**Review notes**

- Worth keeping: Keep the three-panel scaling-law chart on slide 15 and the plain-spoken thesis on slide 6, along with the humor beats on slides 4 and 5 that make the journey memorable.

**Text**

My Journey…
Started working on AI in ~2018, and my perspective slowly shifted
Now…
Human Level AI by 2030?

**Layout**

Single-column, left-aligned bullet block under a centred title, with a centred statement lower on the slide; no footer or slide number visible. Title sits at the top centre (about 7% from the top); bullets start at the left margin (about 5% from the left edge) and the lower statement is centred horizontally in the lower half.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background with black bold Helvetica-style sans-serif text throughout, no accent colours, no shapes, lines, or decorative elements. Title and closing statement are both large bold black (about 7.8% of slide height cap-size box); bullet text is smaller bold black (about 4.4% of slide height). Bullets are solid round black dots. No shadows, gradients, or corner radii.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers entire slide, no border)
- title at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; center aligned]: My Journey…
- body at x 5.2, y 19.9, w 1.3, h 6.7 [Helvetica Bold; font size 6.67% of slide height; #000000; left aligned]: • (Round solid bullet glyph for first item)
- body at x 7.4, y 20.9, w 78.3, h 4.4 [Helvetica Bold; font size 4.44% of slide height; #000000; left aligned]: Started working on AI in ~2018, and my perspective slowly shifted (Bullet list item 1, single line)
- body at x 5.2, y 26.9, w 1.3, h 6.7 [Helvetica Bold; font size 6.67% of slide height; #000000; left aligned]: • (Round solid bullet glyph for second item)
- body at x 7.4, y 27.9, w 7.8, h 4.4 [Helvetica Bold; font size 4.44% of slide height; #000000; left aligned]: Now… (Bullet list item 2)
- title at x 25, y 53.3, w 51.7, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; center aligned]: Human Level AI by 2030?

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: My Journey…
- text at x 5.2, y 19.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 5.2, y 26.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 7.4, y 20.9, w 78.3, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Started working on AI in ~2018, and my perspective slowly shifted
- text at x 7.4, y 27.9, w 7.8, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Now…
- text at x 25.0, y 53.3, w 51.7, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Human Level AI by 2030?

### Slide 5

**Review notes**

- Worth keeping: Keep the three-panel scaling-law chart on slide 15 and the plain-spoken thesis on slide 6, along with the humor beats on slides 4 and 5 that make the journey memorable.

**Text**

My Journey…
• Started working on AI in ~2018, and my perspective slowly shifted
• Next year…
BUY CANNED GOODS AND TIN FOIL HATS

**Layout**

Single-column, left-aligned bullet block under a centred title, with a centred all-caps statement lower on the slide. Title sits at the top centre; bullets start at about 20% from top with a left margin of about 5%; the statement is centred around 53-61% of slide height. No footer, no slide number visible.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background, no decoration. Title and body text in solid black (#000000) Helvetica Bold, a sans-serif, weight bold. Title about 7.8% of slide height (cap height approx 5.4%); bullet body about 4.4% of slide height; the all-caps statement is the same size as the title, about 7.8% of slide height. No accent colours, no shapes, lines, shadows, gradients or corner radii.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers the entire slide; no border, no shadow.)
- title at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica, bold, normal case, no italics, no letter-spacing; font size 7.78% of slide height; #000000; center aligned]: My Journey…
- bullet list at x 5.2, y 19.9, w 83.4, h 11.4 [Helvetica, bold, normal case, no italics, no letter-spacing; font size 4.44% of slide height; #000000; left aligned]: • Started working on AI in ~2018, and my perspective slowly shifted / • Next year… (Two bullets with solid black round bullet glyphs at x 5.2%; bullet text starts at x 7.4%; line pitch about 7% of slide height; bullet 1 at y 20.9%, bullet 2 at y 27.9%.)
- body at x 6.9, y 53.3, w 87.9, h 7.8 [Helvetica, bold, all capitals, no italics, no letter-spacing; font size 7.78% of slide height; #000000; center aligned]: BUY CANNED GOODS AND TIN FOIL HATS (Single line, horizontally centred on the slide.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 35.9, y 7.4, w 28.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: My Journey…
- text at x 5.2, y 19.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 5.2, y 26.9, w 1.3, h 6.7 [Helvetica-Bold; font size 6.67; #000000]: •
- text at x 7.4, y 20.9, w 78.3, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Started working on AI in ~2018, and my perspective slowly shifted
- text at x 7.4, y 27.9, w 13.8, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: Next year…
- text at x 6.9, y 53.3, w 87.9, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: BUY CANNED GOODS AND TIN FOIL HATS

### Slide 6

**Review notes**

- Worth studying: The main-ideas slide states the argument in four short, parallel bullets, a clear way to set up the whole talk.
- Review: The argument is recoverable and the key charts carry it well: slide 6 states the thesis early, and slide 15 shows three clean log-log scaling plots with fitted equations and readable axes.
- Worth keeping: Keep the three-panel scaling-law chart on slide 15 and the plain-spoken thesis on slide 6, along with the humor beats on slides 4 and 5 that make the journey memorable.

**Text**

Main Ideas
AI progress is occurring because it’s very easy — driven predominantly by
increasing resource inputs, with complex algorithms replaced by “scaling up”
Progress is predictable as there are Scaling Laws wrt dataset size, neural network
parameter count, and total compute used for training (which is now >>Avogadro’s
number of floating point operations!)
The largest AI systems are approaching human level; challenges exist for eg literally
automating science, but no compelling blockers at all (in my opinion)
This is all pretty scary for obvious, common sense reasons. Maybe it is the biggest
thing ever, and we’re sleepwalking through it as a scientific community.

**Layout**

Single column of content under a centred title at the top of the slide. The title is centred horizontally (centre at about 50% of width) and starts at about 7.4% from the top. Four bullet points are left-aligned with bullet glyphs at about 5.2% from the left edge and text starting at about 7.4%, with wrapped lines hanging under the text. The bottom roughly 35% of the slide is empty. There is no footer, source line, slide number or logo.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no gradient, pattern or decoration. Title in Helvetica Bold, black #000000, about 7.8% of slide height, sized as a heading. Body text in Helvetica Regular, black #000000, about 4.2% of slide height, with generous line spacing (about 4.5% of slide height between baselines within a bullet). Emphasis is given by bold runs inside bullets: 'very easy' and 'pretty scary'. No accent colour, no rules, no shapes, no shadows, no corner radii.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide background fill (Covers the entire slide, no stroke, no shadow.)
- title at x 39, y 7.4, w 22, h 7.8 [Helvetica, bold, sentence case, no italics, no capitals transform, normal letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Main Ideas
- bullet list at x 5.2, y 20, w 82, h 9 [Helvetica, regular, with 'very easy' in bold, normal letter-spacing; font size 4.17% of slide height; #000000; left aligned]: AI progress is occurring because it’s very easy — driven predominantly by / increasing resource inputs, with complex algorithms replaced by “scaling up” (Solid round bullet glyph (•) at x 5.2%, text indented to x 7.4%; line pitch about 4.5% of slide height; paragraph gap to next bullet about 1.2%.)
- bullet list at x 5.2, y 31.2, w 82.5, h 13 [Helvetica, regular, normal letter-spacing; font size 4.17% of slide height; #000000; left aligned]: Progress is predictable as there are Scaling Laws wrt dataset size, neural network / parameter count, and total compute used for training (which is now >>Avogadro’s / number of floating point operations!) (Solid round bullet glyph (•) at x 5.2%, text indented to x 7.4%; line pitch about 4.5% of slide height; three lines.)
- bullet list at x 5.2, y 46.9, w 83, h 9 [Helvetica, regular, normal letter-spacing; font size 4.17% of slide height; #000000; left aligned]: The largest AI systems are approaching human level; challenges exist for eg literally / automating science, but no compelling blockers at all (in my opinion) (Solid round bullet glyph (•) at x 5.2%, text indented to x 7.4%; line pitch about 4.5% of slide height.)
- bullet list at x 5.2, y 58.2, w 82.8, h 9 [Helvetica, regular, with 'pretty scary' in bold, normal letter-spacing; font size 4.17% of slide height; #000000; left aligned]: This is all pretty scary for obvious, common sense reasons. Maybe it is the biggest / thing ever, and we’re sleepwalking through it as a scientific community. (Solid round bullet glyph (•) at x 5.2%, text indented to about 7.3%; line pitch about 4.5% of slide height; the final full stop sits at x 80.1%, y 63.7%.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 39.0, y 7.4, w 22.0, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Main Ideas
- text at x 5.2, y 20.0, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 5.2, y 31.2, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 7.4, y 20.8, w 38.3, h 4.2 [Helvetica; font size 4.17; #000000]: AI progress is occurring because it’s
- text at x 45.7, y 20.8, w 10.7, h 4.2 [Helvetica-Bold; font size 4.17; #000000]: very easy
- text at x 56.4, y 20.8, w 29.2, h 4.2 [Helvetica; font size 4.17; #000000]: — driven predominantly by
- text at x 7.4, y 25.3, w 79.2, h 4.2 [Helvetica; font size 4.17; #000000]: increasing resource inputs, with complex algorithms replaced by “scaling up”
- text at x 5.2, y 46.9, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 7.4, y 32.1, w 85.7, h 4.2 [Helvetica; font size 4.17; #000000]: Progress is predictable as there are Scaling Laws wrt dataset size, neural network
- text at x 7.4, y 36.6, w 84.9, h 4.2 [Helvetica; font size 4.17; #000000]: parameter count, and total compute used for training (which is now >>Avogadro’s
- text at x 5.2, y 58.2, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 7.4, y 41.1, w 37.5, h 4.2 [Helvetica; font size 4.17; #000000]: number of ﬂoating point operations!)
- text at x 7.4, y 47.8, w 87.8, h 4.2 [Helvetica; font size 4.17; #000000]: The largest AI systems are approaching human level; challenges exist for eg literally
- text at x 7.4, y 52.3, w 70.9, h 4.2 [Helvetica; font size 4.17; #000000]: automating science, but no compelling blockers at all (in my opinion)
- text at x 7.3, y 59.0, w 10.4, h 4.2 [Helvetica; font size 4.17; #000000]: This is all
- text at x 17.7, y 59.0, w 13.3, h 4.2 [Helvetica-Bold; font size 4.17; #000000]: pretty scary
- text at x 31.0, y 59.0, w 63.6, h 4.2 [Helvetica; font size 4.17; #000000]: for obvious, common sense reasons. Maybe it is the biggest
- text at x 7.3, y 63.5, w 72.8, h 4.2 [Helvetica; font size 4.17; #000000]: thing ever, and we’re sleepwalking through it as a scientiﬁc community
- text at x 80.1, y 63.7, w 0.6, h 3.9 [Helvetica; font size 3.89; #000000]: .

### Slide 7

**Text**

Outline
Briefly explain contemporary AI / deep learning, and why it’s very simple
Brief overview of “RLHF” and Constitutional AI
Why AI capabilities will predictably improve
Remaining Challenges for AI (to be broadly human-level)
Why this is scary…

**Layout**

Single-column layout with a centred title at the top and a left-aligned bulleted list occupying the upper-middle portion of the slide. The title sits near the top edge, centred horizontally. The bullet list starts at about 31% of slide height, with bullets at about 6.3% from the left edge and text starting at about 8.5%. The lower third of the slide is empty; there is no footer or slide number.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background with black text. Title is Helvetica Bold, about 7.8% of slide height, black #000000. Body text is Helvetica regular, about 4.4% of slide height, black #000000, with standard round bullets at about 6.7% size. Single bold emphasis on the word 'very'. No accent colours, shapes, lines, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers entire slide, no border or shadow)
- title at x 42.6, y 7.4, w 14.8, h 7.8 [Helvetica Bold, black, regular case, no letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Outline
- bullet list at x 6.3, y 31.5, w 74.1, h 30 [Helvetica regular, black, no letter-spacing; the word 'very' in Helvetica Bold; font size 4.44% of slide height; #000000; left aligned]: Briefly explain contemporary AI / deep learning, and why it’s very simple / Brief overview of “RLHF” and Constitutional AI / Why AI capabilities will predictably improve / Remaining Challenges for AI (to be broadly human-level) / Why this is scary… (Five bullets with round black bullet glyphs at x 6.3%, text at x 8.5%; bullet line pitch about 7.0% of slide height; 'very' is bold inline)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 42.6, y 7.4, w 14.8, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Outline
- text at x 6.3, y 31.5, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 6.3, y 38.5, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 6.3, y 45.6, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 8.5, y 32.5, w 66.8, h 4.4 [Helvetica; font size 4.44; #000000]: Brieﬂy explain contemporary AI / deep learning, and why it’s
- text at x 75.3, y 32.5, w 5.1, h 4.4 [Helvetica-Bold; font size 4.44; #000000]: very
- text at x 80.5, y 32.5, w 7.9, h 4.4 [Helvetica; font size 4.44; #000000]: simple
- text at x 6.3, y 52.7, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 6.3, y 59.7, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 8.5, y 39.5, w 51.7, h 4.4 [Helvetica; font size 4.44; #000000]: Brief overview of “RLHF” and Constitutional AI
- text at x 8.5, y 46.6, w 47.5, h 4.4 [Helvetica; font size 4.44; #000000]: Why AI capabilities will predictably improve
- text at x 8.5, y 53.7, w 62.8, h 4.4 [Helvetica; font size 4.44; #000000]: Remaining Challenges for AI (to be broadly human-level)
- text at x 8.5, y 60.7, w 21.3, h 4.4 [Helvetica; font size 4.44; #000000]: Why this is scary…

### Slide 8

**Text**

Very Quick AI / Need-to-Knows

**Layout**

A single centred title block of two lines occupies the middle of the slide, with the first line set in italic 'Very' followed by 'Quick AI' and the second line 'Need-to-Knows' beneath it. The text is centred horizontally on the slide, with no visible title placeholder, footer, slide number or margins beyond the centred block.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no decoration. All text is black (#000000) bold Helvetica-style sans-serif (Helvetica-Bold; 'Very' in Helvetica-BoldOblique), with a font size of about 12.8% of slide height per line, tight leading between the two lines, and no shapes, lines, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers the entire slide; no border, no shadow, no gradient.)
- title at x 24.1, y 32.2, w 51.9, h 25.1 [Helvetica Bold, with 'Very' in Helvetica Bold Oblique (italic); sans-serif; normal capitals; no letter-spacing; font size 12.78% of slide height; #000000; center aligned]: Very Quick AI / Need-to-Knows (Two lines: line 1 'Very' (italic, x 27.4, w 15.0) then 'Quick AI' (bold, x 42.3, w 32.2) at y 32.2 with h 12.8; line 2 'Need-to-Knows' (bold, x 24.1, w 51.9) at y 44.5 with h 12.8. Lines are stacked with a tight line gap; the block is centred on the slide.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 27.4, y 32.2, w 15.0, h 12.8 [Helvetica-BoldOblique; font size 12.78; #000000]: Very
- text at x 42.3, y 32.2, w 32.2, h 12.8 [Helvetica-Bold; font size 12.78; #000000]: Quick AI
- text at x 24.1, y 44.5, w 51.9, h 12.8 [Helvetica-Bold; font size 12.78; #000000]: Need-to-Knows

### Slide 9

**Text**

What is Contemporary AI / ML?
Just curve fitting with a very general function approximation
•
Neural Network = sequence of matrix multiplications and element-wise
non-linearities, living in a very high dimensional (eg 10k) vector space

**Layout**

Single-column layout on a white field. The title is left-aligned near the top at about 11% from the left edge, spanning most of the width. Two bulleted text blocks sit above and below a centred chart, which is roughly centred horizontally at 29% to 71% of the width. The slide has no footer and no slide number.

**Visuals**

One line-and-scatter chart, centred below the first bullet and above the second. Type: scatter-with-connecting-line plot with a smooth fitted curve. Title: none. Axes: x-axis labelled with ticks 0, 0.25, 0.5, 0.75, 1 (unitless, range 0 to 1). y-axis labelled with ticks -0.25, 0, 0.25, 0.5 (unitless, range about -0.25 to 0.5). Series 1: black square markers joined by a thin black line, about 60 data points; about (0, -0.17), (0.1, -0.13), (0.25, -0.07), (0.4, 0.09), (0.5, 0.24), (0.65, 0.44), (0.75, 0.46), (0.9, 0.39), (1, 0.35), values are estimates marked about. Series 2: smooth red curve (#FF0000), about -0.13 at x=0, about -0.14 near x=0.1, about 0.0 at x=0.3, about 0.27 at x=0.5, about 0.46 peak near x=0.8, falling to about 0.31 at x=1, values are estimates marked about. Legend: none. Highlighted points: none. Gridlines: light grey vertical lines at x=0.25, 0.5, 0.75 and horizontal lines at y=0, 0.25, 0.5. Plot frame: thin black border. The chart is a raster image of 500 by 300 pixels, boxed with no fill or shadow.

**Design**

Background: Solid white #FFFFFF. Plain white slide with no decoration. Title in Helvetica Bold, black #000000, about 8.8% of slide height, sans-serif. Body text in Helvetica regular, black #000000, about 4.1% of slide height, with bold runs for 'curve fitting', 'very' and 'Neural Network'. Bullets are solid black round dots. Accent colour: red #FF0000 used only for the fitted curve in the chart. No shapes, lines, corner radii, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100: Full-slide white background fill (Solid fill #FFFFFF covering the whole slide, no outline.)
- title at x 11.2, y 5.9, w 73.4, h 8.8 [Helvetica, bold, regular case, no letter-spacing; font size 8.81% of slide height; #000000; left aligned]: What is Contemporary AI / ML?
- body at x 13.4, y 24.5, w 64.4, h 5.9 [Helvetica, regular, with bold runs for curve fitting and very; font size 4.07% of slide height; #000000; left aligned]: • Just curve fitting with a very general function approximation (Solid round black bullet dot at x 13.4%, text starts at x 15.9%, single line, no line spacing issues.)
- chart at x 28.9, y 31.9, w 42.3, h 45.1 [#FF0000]: Scatter-with-line plot of noisy data (black squares joined by a thin black line) and a smooth red fitted curve, with x axis 0 to 1 and y axis -0.25 to 0.5, no title, no legend (Raster image 500 by 300 pixels, placed unrotated with no crop, thin black plot frame, light grey gridlines, no shadow, no overlay, not full-bleed.)
- body at x 12.5, y 81, w 73.1, h 10 [Helvetica, regular, with bold run for Neural Network; font size 4.07% of slide height; #000000; left aligned]: • Neural Network = sequence of matrix multiplications and element-wise / non-linearities, living in a very high dimensional (eg 10k) vector space (Solid round black bullet dot at x 12.5%, text starts at x 15.0%, two lines, second line starts at x 15.0% under the first line's text, line pitch about 4.9% of slide height.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 28.9, y 31.9, w 42.3, h 45.1: 500 by 300 pixels
- text at x 11.2, y 5.9, w 73.4, h 8.8 [Helvetica-Bold; font size 8.81; #000000]: What is Contemporary AI / ML?
- text at x 13.4, y 24.5, w 1.2, h 5.9 [Helvetica; font size 5.91; #000000]: •
- text at x 15.9, y 25.2, w 4.8, h 4.1 [Helvetica; font size 4.07; #000000]: Just
- text at x 20.8, y 25.2, w 13.1, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: curve ﬁtting
- text at x 33.9, y 25.2, w 7.3, h 4.1 [Helvetica; font size 4.07; #000000]: with a
- text at x 41.1, y 25.2, w 5.4, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: very
- text at x 46.5, y 25.2, w 31.3, h 4.1 [Helvetica; font size 4.07; #000000]: general function approximation
- text at x 12.5, y 81.0, w 1.2, h 5.9 [Helvetica; font size 5.91; #000000]: •
- text at x 15.0, y 81.8, w 16.8, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: Neural Network
- text at x 31.8, y 81.8, w 56.2, h 4.1 [Helvetica; font size 4.07; #000000]: = sequence of matrix multiplications and element-wise
- text at x 15.0, y 86.7, w 70.6, h 4.1 [Helvetica; font size 4.07; #000000]: non-linearities, living in a very high dimensional (eg 10k) vector space

### Slide 10

**Text**

Language Models
Learn language by “autoregressive” prediction of the next word:
P(w_{n+1}|w_n, ··· w_1)
Optimize the log-likelihood of real world text, which provides the “right answer”.
As a speaker at a journal club, you're probably elephant me to say certain things.
not = 15.10%
going = 13.59%
used = 12.37%
expected = 3.74%
already = 3.53%
familiar = 2.76%
more = 2.57%
thinking = 1.94%
a = 1.64%
aware = 1.48%
elephant = 0.00%
Total: -17.70 logprob on 1 tokens
(58.73% probability covered in top 11 logits)
P ≈ 10^−8

**Layout**

Centered single-column composition on a white field: the title sits at the top centre, followed by a centred subtitle line, a centred blue formula, a centred sentence, and a lower-centre-right token-probability screenshot with a caption-style formula beneath it. Text is centre-aligned with generous white margins; there is no footer or slide number visible.

**Visuals**

A screenshot of a token-probability widget placed on the right of centre below the sentence. Above it, a single line of bold sans-serif text is highlighted token by token: pink highlights on 'a speaker', 'journal', 'elephant me', 'certain' and a pink-to-green transition, and green on 'club' and 'things', with light-yellow on 'you're probably'. Beneath it, a white box with a thin light-grey border and faint drop shadow lists the top 11 next-token candidates with percentages: not = 15.10%, going = 13.59%, used = 12.37%, expected = 3.74%, already = 3.53%, familiar = 2.76%, more = 2.57%, thinking = 1.94%, a = 1.64%, aware = 1.48%, and the highlighted row elephant = 0.00% on a bright red-pink (#FF5A74) bar with white text. Below the list: 'Total: -17.70 logprob on 1 tokens' and the smaller line '(58.73% probability covered in top 11 logits)'. Beneath the box, the rendered formula P ≈ 10^−8 in black serif math type. No photographs or logos.

**Design**

Background: Solid white (#FFFFFF). White background (#FFFFFF). Title in heavy Helvetica Neue Medium/bold black (#000000), large, about 10.4% of slide height in cap-height terms. Body text in Helvetica Neue regular black (#000000) about 3% of slide height, with bold runs for 'prediction of the next word' and 'likelihood'. Mathematical formulas in Computer Modern serif type: the blue equation in #0001FE and the closing black formula in #000000. Accent colours in the screenshot: pink highlights (about #FFC9C0), green highlights (about #C8F5B0), bright red-pink selected row (#FF5A74). Shapes are square-cornered with a subtle grey border and soft shadow on the widget box.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- fill at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background rectangle (Covers the whole slide, behind everything.)
- title at x 25.9, y 5.6, w 48.2, h 10.4 [Helvetica Neue Medium, black, regular weight, normal case; font size 10.37% of slide height; #000000; center aligned]: Language Models
- body at x 24.3, y 21.2, w 51.9, h 3 [Helvetica Neue regular with bold run 'prediction of the next word', black, normal case; font size 2.96% of slide height; #000000; center aligned]: Learn language by “autoregressive” prediction of the next word: (Single line; bold run at x 51.3 to 72.7; colon at x 72.7.)
- body at x 36.9, y 26.6, w 22.7, h 5.1 [Computer Modern (CMMI10/CMR10/CMSY10) math serif, regular; font size 5.11% of slide height; #0001FE; center aligned]: P(w_{n+1}|w_n, ··· w_1) (Display equation; subscripts n+1 and n in CMMI7/CMR7 at about 3.58% em, placed at x 42.3 to 52.9 and 58.4 for the final subscript 1; vertical bar between w_{n+1} and w_n; spacing dots ··· between w_n and w_1.)
- body at x 20, y 35.3, w 60, h 3 [Helvetica Neue regular with bold run 'likelihood', black, normal case; font size 2.96% of slide height; #000000; center aligned]: Optimize the log-likelihood of real world text, which provides the “right answer”. (Single line; bold run at x 32.9 to 40.6.)
- caption at x 29.2, y 49.8, w 38, h 2.4 [Helvetica Neue bold, black, normal case, small; font size 1.6% of slide height; #000000; left aligned]: As a speaker at a journal club, you're probably elephant me to say certain things. (Single line of token-highlighted text; highlights in pink (#FFC9C0), green (#C8F5B0) and light yellow behind words 'a speaker', 'journal', 'club', 'you're probably', 'elephant me', 'certain', 'things'.)
- photo at x 51.6, y 52, w 20.7, h 13.7 [#FFFFFF]: Screenshot of token-probability widget: white dropdown panel listing top next-token candidates with percentages, a highlighted elephant row, and total logprob footer (Panel with thin light-grey border and faint drop shadow; rows in sans-serif black text (about 1.6% em); selected row 'elephant = 0.00%' with bright red-pink fill #FF5A74 and white text; footer 'Total: -17.70 logprob on 1 tokens' and '(58.73% probability covered in top 11 logits)'. Square corners.)
- body at x 57.4, y 88.8, w 12.2, h 5.2 [Computer Modern math serif, regular, black; font size 5.15% of slide height; #000000; left aligned]: P ≈ 10^−8 (Formula below the widget; superscript −8 at about 3.61% em at x 66.4 to 69.4.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 28.6, y 42.8, w 42.7, h 46.2: 666 by 405 pixels
- text at x 25.9, y 5.6, w 48.2, h 10.4 [HelveticaNeue-Medium; font size 10.37; #000000]: Language Models
- text at x 36.9, y 26.6, w 1.8, h 5.1 [CMMI10; font size 5.11; #0001FE]: P
- text at x 39.1, y 26.6, w 1.1, h 5.1 [CMR10; font size 5.11; #0001FE]: (
- text at x 40.2, y 26.6, w 2.1, h 5.1 [CMMI10; font size 5.11; #0001FE]: w
- text at x 46.8, y 26.6, w 0.8, h 5.1 [CMSY10; font size 5.11; #0001FE]: |
- text at x 47.6, y 26.6, w 4.4, h 5.1 [CMMI10; font size 5.11; #0001FE]: w ,
- text at x 52.5, y 26.6, w 3.4, h 5.1 [CMSY10; font size 5.11; #0001FE]: ···
- text at x 56.3, y 26.6, w 2.1, h 5.1 [CMMI10; font size 5.11; #0001FE]: w
- text at x 59.6, y 26.6, w 1.1, h 5.1 [CMR10; font size 5.11; #0001FE]: )
- text at x 24.3, y 21.2, w 27.0, h 3.0 [HelveticaNeue; font size 2.96; #000000]: Learn language by “autoregressive”
- text at x 51.3, y 21.2, w 21.4, h 3.0 [HelveticaNeue-Bold; font size 2.96; #000000]: prediction of the next word
- text at x 72.7, y 21.2, w 0.5, h 3.0 [HelveticaNeue; font size 2.96; #000000]: :
- text at x 42.3, y 28.6, w 1.4, h 3.6 [CMMI7; font size 3.58; #0001FE]: n
- text at x 43.7, y 28.6, w 2.9, h 3.6 [CMR7; font size 3.58; #0001FE]: +1
- text at x 49.6, y 28.6, w 1.4, h 3.6 [CMMI7; font size 3.58; #0001FE]: n
- text at x 58.4, y 28.6, w 1.1, h 3.6 [CMR7; font size 3.58; #0001FE]: 1
- text at x 20.0, y 35.3, w 12.8, h 3.0 [HelveticaNeue; font size 2.96; #000000]: Optimize the log-
- text at x 32.9, y 35.3, w 7.7, h 3.0 [HelveticaNeue-Bold; font size 2.96; #000000]: likelihood
- text at x 40.6, y 35.3, w 39.9, h 3.0 [HelveticaNeue; font size 2.96; #000000]: of real world text, which provides the “right answer”.
- text at x 57.4, y 89.7, w 1.9, h 5.2 [CMMI10; font size 5.15; #000000]: P
- text at x 60.5, y 89.7, w 2.3, h 5.2 [CMSY10; font size 5.15; #000000]: ⇡
- text at x 63.5, y 89.7, w 2.9, h 5.2 [CMR10; font size 5.15; #000000]: 10
- text at x 66.4, y 88.8, w 1.8, h 3.6 [CMSY7; font size 3.61; #000000]: −
- text at x 68.2, y 88.8, w 1.2, h 3.6 [CMR7; font size 3.61; #000000]: 8

### Slide 11

**Text**

Playground task / Talk to the / assistant / Task 1 of 5 / Decide on a task you'd like the assistant to help with and enter it into the task box. Interact with the AI assistant. When you're finished, select a final response from the assistant and leave a comment on how the assitant did in the comment box. Click to go to the next task. / Comments / I thought the assistant was ... / Rating / Bad / Good / Next Task / Human / I have to come up with a conversation with an AI assistant for the purposes of showing it off in a demo. What sorts of conversation topics would you recommend? / Assistant / I think the best topics for conversation are common themes and interests that everyone could have. Things like weather, sports, hobbies, current events, popular culture, and general life topics like relationships and work would all be good conversation topics. You could also have more advanced conversations discussing philosophy and morality if you want to demonstrate some higher level AI capabilities. / Human / I think I'm more interested in an advanced conversation. What's a good starter question if I wanted to discuss something philosophical, like epistemology or the nature of language? / Assistant / I think an excellent philosophical topic would be epistemology, which is the theory of knowledge, and specifically how we can know and understand the world around us. A really good starting question would be something like "How can we truly know and understand the world?", or "What is the nature of knowledge?". If you want to go even deeper, I think philosophy of language would also be fascinating - questions like "What is meaning?", "How do language and thought relate?", and "How can language mislead and deceive?" would be great discussion topics. / Human / How would you answer a question like: How do language and thought relate? / Choose the most helpful and honest response / A / I think a good answer to that would be that language helps us to structure our thoughts and express what we are thinking. Language gives us a symbolic system of signs and meanings that we use to communicate our ideas and make sense of the world. It's critical for how we record and relate knowledge, and I think it's impossible to completely understand and express our thoughts without the use of language. Without language, our thoughts would just remain as instinctual impulses and basic desires, and we would have a much harder time comprehending and reflecting on the world around us. / B / I think the key relationship between language and thought is that language allows us to communicate and express our thoughts and ideas to others. Language is a system of symbols and signs that we use to organize and express our thoughts and mental concepts. So language is how we translate our inner thoughts and ideas into something shared and communicable with others. Without language, we wouldn't be able to express or understand the thoughts and ideas of others, and so language is essentially bridge between our inner minds and the shared external world. / A A A A B B B B / A is better / B is better

**Layout**

A single screenshot of a web interface sits centred on a plain white slide, spanning about 75.5% of the slide width with roughly 12% margins left and right and nearly the full height. Inside the screenshot, a narrow left column (about 29% of its width) holds the title, task instructions, comments box, rating and Next Task button, while a wider right column holds the chat transcript above a full-width indigo choice panel. There is no slide title, footer or slide number outside the screenshot.

**Visuals**

One screenshot of a web app (1442 by 1074 pixels), placed unscaled-looking and boxed by a thin light grey border with a white surround. Left column: a title 'Talk to the assistant' in bold black sans-serif, a blue info icon beside 'Playground task', a grey label 'Task 1 of 5', a light grey rounded box with the instructions, an empty 'Comments' text area with grey placeholder, a 'Rating' box with four empty radio circles and the labels 'Bad' and 'Good' at the ends, and a pale lavender 'Next Task' button with white text. Right column: a white chat transcript with small circular avatars (a human face icon and a robot icon) beside each speaker name, then a solid indigo panel (#4A4AE8) containing a robot avatar and the prompt 'Choose the most helpful and honest response', two white rounded answer boxes labelled A and B, and a white rounded slider bar with eight letter buttons (A A A A B B B B, decreasing then increasing in size) with labels 'A is better' at left and 'B is better' at right. No charts, photographs or logos.

**Design**

Background: Solid white #FFFFFF, covering the whole slide. Plain white background (#FFFFFF). Interface text is dark grey (#374151) for body and chat text, black (#111111) for the bold title, indigo (#4A4AE8) for the playground label, info icon and choice panel, and white on indigo for the prompt and answer text. Typeface is a neo-grotesque sans-serif similar to Inter: bold for the title, regular for body, medium for labels. Title is about 3.5% of slide height; body text about 1.3% to 1.4%. Shapes have thin light grey borders and rounded corners of about 6px in the screenshot; the Next Task button is a pale lavender (#A5A4F4) rectangle with rounded corners; no shadows or gradients are visible.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Solid fill covering the slide, no border.)
- photo at x 12.2, y 0.1, w 75.5, h 99.9: Screenshot of the 'Talk to the assistant' web interface (1442 by 1074 pixels) containing all the interface text and elements listed below (Not cropped, not full-bleed; boxed by a thin light grey border; sits on the white slide background; overall dominant colours are white, light grey and indigo #4A4AE8.)
- icon at x 21.4, y 5.4, w 1.1, h 1.6 [#4A4AE8]: Blue information icon (circled i) beside 'Playground task' (Filled circle with white letter i.)
- subtitle at x 14.5, y 4.9, w 6.2, h 1.6 [sans-serif regular; font size 1.3% of slide height; #4A4AE8; left aligned]: Playground task
- title at x 14.5, y 7.3, w 9, h 7 [sans-serif bold; font size 3.5% of slide height; #111111; left aligned]: Talk to the / assistant (Two lines, tight leading.)
- body at x 14.5, y 17.4, w 4.1, h 1.4 [sans-serif regular; font size 1.2% of slide height; #6B7280; left aligned]: Task 1 of 5
- shape at x 14.5, y 19.2, w 13.9, h 30 [#F9FAFB]: Light grey rounded box containing the task instructions (Thin light grey border, corner radius about 6px, very slight shadow.)
- body at x 15.2, y 20.8, w 12.3, h 26.5 [sans-serif regular; font size 1.5% of slide height; #111111; left aligned]: Decide on a task you'd like / the assistant to help with / and enter it into the task / box. Interact with the AI / assistant. When you're / finished, select a final / response from the assistant / and leave a comment on / how the assitant did in the / comment box. Click to go to / the next task. (Line spacing about 1.45.)
- body at x 15, y 51.2, w 4.7, h 1.5 [sans-serif medium; font size 1.2% of slide height; #111111; left aligned]: Comments (Sits on the top border of the comment box as a boxed label with white fill.)
- shape at x 14.5, y 52, w 13.9, h 12.4 [#FFFFFF]: Empty white comment text area (Thin grey border, corner radius about 4px, resize handle in bottom-right corner.)
- body at x 15.2, y 53.4, w 12.2, h 2 [sans-serif regular; font size 1.4% of slide height; #9CA3AF; left aligned]: I thought the assistant was ... (Grey placeholder text.)
- body at x 15, y 63.4, w 3.8, h 1.5 [sans-serif medium; font size 1.2% of slide height; #111111; left aligned]: Rating (Boxed label on the top border of the rating box.)
- shape at x 14.5, y 67.1, w 13.9, h 8.6 [#FFFFFF]: Rating box with four empty radio circles and end labels Bad and Good (Thin grey border, corner radius about 6px; four unfilled circles of about 1% width spaced evenly in one row; Bad is under the first circle and Good under the fourth.)
- body at x 16.4, y 72.2, w 1.7, h 1.5 [sans-serif regular; font size 1.3% of slide height; #111111; center aligned]: Bad
- body at x 24.6, y 72.2, w 2, h 1.5 [sans-serif regular; font size 1.3% of slide height; #111111; center aligned]: Good
- shape at x 14.5, y 80.2, w 13.9, h 6.5 [#A5A4F4]: Next Task button (Pale lavender rectangle with rounded corners of about 6px, muted indigo tint, no shadow.)
- body at x 19, y 82.5, w 4.9, h 1.9 [sans-serif medium; font size 1.7% of slide height; #FFFFFF; center aligned]: Next Task (Centred on the button.)
- shape at x 31.5, y 2.2, w 55.4, h 48 [#FFFFFF]: Chat transcript panel with alternating Human and Assistant messages (White background; each speaker has a small circular avatar (human face icon or robot icon) of about 1.9% width at the left of the speaker name; speaker names in bold dark grey, messages in grey #6B7280 regular sans-serif.)
- body at x 33.9, y 1.3, w 50, h 6 [sans-serif regular; font size 1.3% of slide height; #6B7280; left aligned]: Human / I have to come up with a conversation with an AI assistant for the purposes of showing it off in a demo. What sorts of conversation topics would you recommend? (Speaker name Human in bold dark grey above the message.)
- body at x 33.9, y 10, w 50, h 10.5 [sans-serif regular; font size 1.3% of slide height; #6B7280; left aligned]: Assistant / I think the best topics for conversation are common themes and interests that everyone could have. Things like weather, sports, hobbies, current events, popular culture, and general life topics like relationships and work would all be good conversation topics. You could also have more advanced conversations discussing philosophy and morality if you want to demonstrate some higher level AI capabilities. (Speaker name Assistant in bold dark grey above the message; line spacing about 1.45.)
- body at x 33.9, y 23.2, w 50, h 4.5 [sans-serif regular; font size 1.3% of slide height; #6B7280; left aligned]: Human / I think I'm more interested in an advanced conversation. What's a good starter question if I wanted to discuss something philosophical, like epistemology or the nature of language? (Speaker name Human in bold dark grey above the message.)
- body at x 33.9, y 31.6, w 50, h 11.5 [sans-serif regular; font size 1.3% of slide height; #6B7280; left aligned]: Assistant / I think an excellent philosophical topic would be epistemology, which is the theory of knowledge, and specifically how we can know and understand the world around us. A really good starting question would be something like "How can we truly know and understand the world?", or "What is the nature of knowledge?". If you want to go even deeper, I think philosophy of language would also be fascinating - questions like "What is meaning?", "How do language and thought relate?", and "How can language mislead and deceive?" would be great discussion topics. (Speaker name Assistant in bold dark grey above the message; line spacing about 1.45.)
- body at x 33.9, y 46.1, w 30, h 3 [sans-serif regular; font size 1.3% of slide height; #6B7280; left aligned]: Human / How would you answer a question like: How do language and thought relate? (Speaker name Human in bold dark grey above the message.)
- shape at x 31, y 52.4, w 56, h 45.7 [#4A4AE8]: Indigo choice panel with prompt, two answer boxes and preference slider (Solid indigo fill, no border, square corners.)
- icon at x 31.6, y 54.2, w 1.2, h 2 [#A5B4FC]: Robot avatar in the choice panel header (Circular avatar with light blue-grey fill.)
- body at x 33.9, y 54.4, w 18.4, h 1.6 [sans-serif regular; font size 1.3% of slide height; #FFFFFF; left aligned]: Choose the most helpful and honest response
- shape at x 33.5, y 58.3, w 52.9, h 15.4 [#FFFFFF]: Answer box A (White rounded rectangle, corner radius about 4px, thin light grey border; label A sits to its left in white on indigo.)
- body at x 32.2, y 65.6, w 0.8, h 1.8 [sans-serif bold; font size 1.8% of slide height; #FFFFFF; center aligned]: A
- body at x 33.8, y 59.2, w 52.5, h 12 [sans-serif regular; font size 1.3% of slide height; #374151; left aligned]: I think a good answer to that would be that language helps us to structure our thoughts and express what we are thinking. Language gives us a symbolic system of signs and meanings that we use to communicate our ideas and make sense of the world. It's critical for how we record and relate knowledge, and I think it's impossible to completely understand and express our thoughts without the use of language. Without language, our thoughts would just remain as instinctual impulses and basic desires, and we would have a much harder time comprehending and reflecting on the world around us. (Line spacing about 1.45.)
- shape at x 33.5, y 75.5, w 52.9, h 15.3 [#FFFFFF]: Answer box B (White rounded rectangle, corner radius about 4px, thin light grey border; label B sits to its left in white on indigo.)
- body at x 32.2, y 82.8, w 0.8, h 1.8 [sans-serif bold; font size 1.8% of slide height; #FFFFFF; center aligned]: B
- body at x 33.8, y 76.4, w 52.5, h 12 [sans-serif regular; font size 1.3% of slide height; #374151; left aligned]: I think the key relationship between language and thought is that language allows us to communicate and express our thoughts and ideas to others. Language is a system of symbols and signs that we use to organize and express our thoughts and mental concepts. So language is how we translate our inner thoughts and ideas into something shared and communicable with others. Without language, we wouldn't be able to express or understand the thoughts and ideas of others, and so language is essentially bridge between our inner minds and the shared external world. (Line spacing about 1.45.)
- shape at x 48.7, y 92.4, w 20.7, h 3.6 [#FFFFFF]: Preference slider bar with eight letter buttons (White rounded rectangle with thin grey border and faint shadow; eight equal segments separated by thin grey dividers; letters A A A A B B B B in dark grey, sizes stepping down then up (largest at the outer ends, smallest at the centre-left A and centre-right B).)
- body at x 49.2, y 93, w 19.7, h 2 [sans-serif medium; font size 1.8% of slide height; #374151; center aligned]: A A A A B B B B (Eight letters in separate segments, varying size.)
- body at x 48.7, y 96.9, w 4.2, h 1.4 [sans-serif medium; font size 1.2% of slide height; #FFFFFF; left aligned]: A is better (Label beneath the left end of the slider.)
- body at x 65.1, y 96.9, w 4.2, h 1.4 [sans-serif medium; font size 1.2% of slide height; #FFFFFF; right aligned]: B is better (Label beneath the right end of the slider.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 12.2, y 0.1, w 75.5, h 99.9: 1442 by 1074 pixels

### Slide 12

**Text**

Summary of Contemporary AI Training
•
•
•
First “Pretrain” LLMs to predict “the next word” in an enormous corpus of text
Then collect Human Feedback… or just AI self-feedback based on a list of
principles… to train a “Preference Model” that assigns scores to behaviors
Finally use Reinforcement Learning to “ﬁnetune” the LLM to write responses that
get high scores according to the preference model when given diverse inputs

**Layout**

Single-column slide on a plain white field. The title is centred horizontally across the top band, starting at about 7.4% from the top. The three bullet points sit in one left-aligned column that begins at about 6% from the left edge and occupies the upper-middle of the slide, with the lower third left empty. There is no footer, no slide number and no logo.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background with no gradient, texture or shapes. All text is black (#000000) in Helvetica: the title is Helvetica Bold at about 7.8% of slide height, and the body is Helvetica Regular at about 4.2% of slide height. Bullets are solid round dots in the body colour. No accent colours, rules, boxes or shadows are used.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide white background fill (Covers the entire slide; no border, no shadow, no gradient, opacity 100%.)
- title at x 10.7, y 7.4, w 78.6, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; center aligned]: Summary of Contemporary AI Training (Single line, no letter-spacing adjustment, no capitals transform.)
- body at x 6, y 33.9, w 1.2, h 6.2 [Helvetica Regular; font size 6.25% of slide height; #000000; left aligned]: • (Round bullet glyph for the first point.)
- body at x 6, y 40.6, w 1.2, h 6.2 [Helvetica Regular; font size 6.25% of slide height; #000000; left aligned]: • (Round bullet glyph for the second point.)
- body at x 6, y 51.8, w 1.2, h 6.2 [Helvetica Regular; font size 6.25% of slide height; #000000; left aligned]: • (Round bullet glyph for the third point.)
- body at x 8.3, y 34.7, w 79.5, h 4.2 [Helvetica Regular; font size 4.17% of slide height; #000000; left aligned]: First “Pretrain” LLMs to predict “the next word” in an enormous corpus of text (Single line; first bullet item.)
- body at x 8.3, y 41.5, w 77.5, h 4.2 [Helvetica Regular; font size 4.17% of slide height; #000000; left aligned]: Then collect Human Feedback… or just AI self-feedback based on a list of (Line 1 of the second bullet item; wraps to the next line.)
- body at x 8.3, y 46, w 76.9, h 4.2 [Helvetica Regular; font size 4.17% of slide height; #000000; left aligned]: principles… to train a “Preference Model” that assigns scores to behaviors (Line 2 of the second bullet item; line spacing about 4.5% of slide height.)
- body at x 8.3, y 52.7, w 84, h 4.2 [Helvetica Regular; font size 4.17% of slide height; #000000; left aligned]: Finally use Reinforcement Learning to “ﬁnetune” the LLM to write responses that (Line 1 of the third bullet item.)
- body at x 8.3, y 57.2, w 79.7, h 4.2 [Helvetica Regular; font size 4.17% of slide height; #000000; left aligned]: get high scores according to the preference model when given diverse inputs (Line 2 of the third bullet item.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 10.7, y 7.4, w 78.6, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Summary of Contemporary AI Training
- text at x 6.0, y 33.9, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 6.0, y 40.6, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 6.0, y 51.8, w 1.2, h 6.2 [Helvetica; font size 6.25; #000000]: •
- text at x 8.3, y 34.7, w 79.5, h 4.2 [Helvetica; font size 4.17; #000000]: First “Pretrain” LLMs to predict “the next word” in an enormous corpus of text
- text at x 8.3, y 41.5, w 77.5, h 4.2 [Helvetica; font size 4.17; #000000]: Then collect Human Feedback… or just AI self-feedback based on a list of
- text at x 8.3, y 46.0, w 76.9, h 4.2 [Helvetica; font size 4.17; #000000]: principles… to train a “Preference Model” that assigns scores to behaviors
- text at x 8.3, y 52.7, w 84.0, h 4.2 [Helvetica; font size 4.17; #000000]: Finally use Reinforcement Learning to “ﬁnetune” the LLM to write responses that
- text at x 8.3, y 57.2, w 79.7, h 4.2 [Helvetica; font size 4.17; #000000]: get high scores according to the preference model when given diverse inputs

### Slide 13

**Review notes**

- Review: The preference modelling, RL and Constitutional AI pages (31 to 34) sit after the 'Thanks!' slide (29) and a repeated Claude demo (30), and slides 13, 14 and 30 show tiny screenshots that cannot be read.

**Text**

Claude
Liouville Theory: An Introduction to Rigorous Approaches
Sourav Chatterjee1 and Edward Witten2
1Department of Statistics, Stanford University, Stanford CA 94305 USA
2School of Natural Sciences, Institute for Advanced Study, Princeton, NJ 08540 USA
ABSTRACT: In recent years, a surprisingly direct and simple rigorous understanding of quantum Liouville theory has developed. We aim here to make this material more accessible to physicists working on quantum field theory.
arXiv:2404.02001v2 [hep-th] 5 Apr 2024
Witten2404.02001
PDF
JK
What are some ideas explained in this paper that might be new to theoretical physicists who already have some familiarity with Liouville theory?
This paper explains several ideas from the rigorous probabilistic approach to Liouville theory that may be unfamiliar or provide new insights to theoretical physicists:
1. The proof that the random Liouville measure :e^(2bφ): vanishes if b>1, so the probabilistic approach does not suffice to define Liouville theory for b≥1.
2. A detailed analysis of which moments E[M_f(A)^p] of the integrated Liouville measure M_f(A) converge or diverge depending on the values of b and p. This provides a rigorous understanding of some nonperturbative poles of the DOZZ formula.
3. An explanation of why the DOZZ formula vanishes when one of the Liouville momenta reaches the Seiberg bound at Q/2. This is because the measure M_f(A) becomes infinite with probability 1, making its negative powers vanish.
4. A proof that the Liouville measure M_f(A) converges to a nontrivial limit as the cutoff is removed for all b<1, not just b<1/sqrt(2) which is easier to analyze. This establishes existence of the theory for b<1.
5. A rigorous framework for defining non-integer moments of M_f(A), which correspond to the Liouville correlators. This goes beyond calculations of moments for integer p.
So in summary, it provides a mathematically rigorous perspective on analyticity properties, existence and non-existence for different ranges of b, and subtle issues like vanishing at the Seiberg bound, that complement the physicists' understanding based on path integrals and conformal bootstrap.
Debug Copy Retry
Claude can make mistakes. Please double-check responses.

**Layout**

Two-column split on a white slide: the left half holds the title 'Claude' at top-left and a page image of the arXiv paper, vertically centred in the left column with its left edge near 6% of the width. The right half holds a full-height screenshot of a Claude chat window, inset from the slide edges by about 1 to 2% at top and bottom. There is no footer, slide number or logo outside the screenshot.

**Visuals**

Two images. (1) A photographed or rendered first page of the arXiv preprint arXiv:2404.02001v2 on white paper, with a vertical grey arXiv stamp running up the left edge of the page reading 'arXiv:2404.02001v2 [hep-th] 5 Apr 2024'. The page shows the bold title 'Liouville Theory: An Introduction to Rigorous Approaches', a black horizontal rule, author line 'Sourav Chatterjee and Edward Witten' with superscript affiliation numbers, two italic affiliation lines, and a small-caps ABSTRACT paragraph in a Computer Modern style serif. Shown flat, unrotated, with no overlay, box or shadow; the white page blends into the slide background. (2) A screenshot of the Claude chat interface on a warm cream gradient panel (lighter at top, slightly darker beige toward the bottom) with a rounded rectangular boundary. At top left is a file card, light blue-white gradient with blue outline, labelled 'Witten2404.02001' in blue, with a blue 'PDF' pill badge. Below is a beige user message bubble with a dark brown circle avatar containing 'JK' in white. The Claude response is set in a serif body face on an off-white rounded card with a faint shadow, numbered list items 1 to 5 and a closing paragraph. A small floating toolbar with grey icons and labels 'Debug', 'Copy', 'Retry' and thumbs-up and thumbs-down sits at the lower right of the card. At bottom left is the orange Claude starburst logo (#D97757, approximately eight-spoked asterisk). At bottom right is small grey disclaimer text.

**Design**

Background: Solid white #FFFFFF across the whole slide, with a warm cream gradient panel #F3F1EA to #E6DFCC occupying the right screenshot area. Clean minimal white slide. Title 'Claude' in a bold neo-grotesque sans (Helvetica Bold), black #000000, about 7.8% of slide height for the cap-height line, left-aligned. Paper image uses Computer Modern-style serif for body, black #000000 on white. Screenshot panel background warm cream gradient about #F3F1EA at top to #E6DFCC at bottom; user bubble fill #E2D9C4; response card fill #FAF9F5 with soft grey shadow and rounded corners about 1% of slide width. Chat text in dark charcoal #2B2A27 serif, body about 1.6% of slide height. Accent: Claude orange #D97757 starburst; file badge and file name blue #1A73E8; disclaimer text grey #7A7770 sans, about 1.1% of slide height.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide background fill (Covers the slide; no border.)
- title at x 16.1, y 5.3, w 14.4, h 7.8 [Helvetica; font size 7.78% of slide height; #000000; left aligned]: Claude (Bold weight, sentence case, no letter-spacing.)
- photo at x 5.8, y 20.3, w 38.2, h 59.4 [#FFFFFF]: Image of the first page of arXiv preprint 2404.02001v2 (Chatterjee and Witten), with title, authors, affiliations, abstract and vertical arXiv stamp (Placed unrotated, unboxed, no shadow, not cropped; white page blends into background; text on page is Computer Modern-style serif, black.)
- illustration at x 49.8, y 1.3, w 44.8, h 97.4 [#F3F1EA]: Screenshot of Claude chat interface showing a PDF attachment, a user question, and a numbered response with a summary (Rounded panel with cream gradient; response card #FAF9F5 with soft shadow; user bubble #E2D9C4; file card with blue #1A73E8 outline; orange #D97757 Claude starburst at bottom left.)
- logo at x 50.9, y 92.7, w 2.9, h 3.4 [#D97757]: Claude starburst logo (Eight-spoke orange asterisk, bottom left of the chat panel.)
- caption at x 74.7, y 93.5, w 24.9, h 1.5 [sans-serif; font size 1.1% of slide height; #7A7770; right aligned]: Claude can make mistakes. Please double-check responses. (Regular weight, grey, inside the chat screenshot.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 49.8, y 1.3, w 44.8, h 97.4: 826 by 1011 pixels
- image at x 5.8, y 20.3, w 38.2, h 59.4: 737 by 644 pixels
- text at x 16.1, y 5.3, w 14.4, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Claude

### Slide 14

**Review notes**

- Review: The preference modelling, RL and Constitutional AI pages (31 to 34) sit after the 'Thanks!' slide (29) and a repeated Claude demo (30), and slides 13, 14 and 30 show tiny screenshots that cannot be read.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Why AI Progress is
Easy and Predictable

**Layout**

A single centred block of two lines of bold title text sits in the middle of the slide, horizontally and vertically centred on a plain white field with no grid, header, footer, or slide number. Both lines are centre-aligned with the second line slightly wider than the first.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no decoration. Title text is black (#000000) in Helvetica-Bold, sans-serif, bold weight, set at about 12.8% of slide height per line with tight leading (line pitch about 12.3% of slide height). No accent colours, shapes, lines, shadows, gradients, or corner radii.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- title at x 14.5, y 33.8, w 70.9, h 25.1 [Helvetica, bold, regular capitals and lowercase, no italics, no letter-spacing; font size 12.78% of slide height; #000000; center aligned]: Why AI Progress is / Easy and Predictable (Two lines: line 1 'Why AI Progress is' at x 18.0, y 33.8, w 65.9, h 12.8; line 2 'Easy and Predictable' at x 14.5, y 46.1, w 70.9, h 12.8. Line break after 'is'. Solid black text, no fill or outline, no shadow.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 18.0, y 33.8, w 65.9, h 12.8 [Helvetica-Bold; font size 12.78; #000000]: Why AI Progress is
- text at x 14.5, y 46.1, w 70.9, h 12.8 [Helvetica-Bold; font size 12.78; #000000]: Easy and Predictable

### Slide 15

**Review notes**

- Worth studying: The scaling-laws slide pairs a single relationship with a compact formula, so the audience sees the core claim at a glance.
- Review: The argument is recoverable and the key charts carry it well: slide 6 states the thesis early, and slide 15 shows three clean log-log scaling plots with fitted equations and readable axes.
- Worth keeping: Keep the three-panel scaling-law chart on slide 15 and the plain-spoken thesis on slide 6, along with the humor beats on slides 4 and 5 that make the journey memorable.

**Text**

Scaling Laws for Language Models
Just train many models of different sizes on different sized datasets:
Test Loss / 7 / 6 / 5 / 4 / 3 / 2 / 10-9 / 10-7 / 10-5 / 10-3 / 10-1 / 101 / Compute / PF-days, non-embedding / L = (Cmin/2.3 · 108)-0.050 / 4.2 / 3.9 / 3.6 / 3.3 / 3.0 / 2.7 / 108 / 109 / Dataset Size / tokens / L = (D/5.4 · 1013)-0.095 / 5.6 / 4.8 / 4.0 / 3.2 / 2.4 / 105 / 107 / 109 / Parameters / non-embedding / L = (N/8.8 · 1013)-0.076
Compute ~ Params * Tokens Processed, PF-day ~ 9 × 1019
These types of scaling laws are quite universal and are observed in most ML systems.

**Layout**

Centered single-column slide on white. Title is centred at the top (y 7.8%) with a subtitle line centred beneath it (y 22.3%). A row of three log-scale charts spans the middle (x 9% to 91%, y 27% to 72%), with axis titles centred under each chart. A centred equation line sits at y 76.5% and a centred concluding sentence sits near the bottom at y 89.8%. No footer, page number or logo is visible.

**Visuals**

A single raster figure of three side-by-side line charts (1174 by 364 px, placed at x 9.0%, y 27.1%, w 82.0%, h 45.2%, unboxed, not cropped). Left chart: Compute (PF-days, non-embedding), log x-axis from 10^-9 to 10^1 with ticks at 10^-9, 10^-7, 10^-5, 10^-3, 10^-1, 10^1; linear y-axis Test Loss from 2 to 7 with ticks 2, 3, 4, 5, 6, 7. Dozens of thin light-blue curves (about 40 to 60 runs) descend from upper left to lower right, with a bold black lower envelope curve; an orange dashed fit line, legend box at lower left reading L = (Cmin/2.3 · 10^8)^-0.050. No point values are printed. Middle chart: Dataset Size (tokens), log x-axis with ticks 10^8 and 10^9; y-axis from 2.7 to 4.2 with ticks 2.7, 3.0, 3.3, 3.6, 3.9, 4.2. A blue line with about 8 markers descends to the right, with a thin grey fit line; legend box at upper right reads L = (D/5.4 · 10^13)^-0.095; first marker about 4.2 at about 2 × 10^7 tokens, last marker about 2.7 near 1.5 × 10^9 tokens. Right chart: Parameters (non-embedding), log x-axis with ticks 10^5, 10^7, 10^9; y-axis from 2.4 to 5.6 with ticks 2.4, 3.2, 4.0, 4.8, 5.6. A dark blue line with about 20 markers and a thin grey fit line; legend box at upper right reads L = (N/8.8 · 10^13)^-0.076; first marker about 5.6 near 3 × 10^4 parameters, last near 2.4 near 1.2 × 10^9. Colours: blue #1F77B4-like markers and lines, light blue #9DC3E6 runs, black #000000 envelope, orange #FFA500 dashed line, grey #808080 fit lines. Axes are black thin frames with tick labels in a serif (Computer Modern) typeface. Equation line: Compute ~ Params * Tokens Processed, PF-day ~ followed by 9 × 10 to the power 19 rendered in serif math type.

**Design**

Background: Solid white #FFFFFF, covering the full slide. Plain white background (#FFFFFF). Title in bold Helvetica Neue style sans-serif, black #000000, about 8.2% of slide height for cap-height text, with a large size relative to body. Body text in regular Helvetica Neue sans-serif, black #000000, about 3.0% to 3.7% of slide height. Figure labels in Computer Modern serif for tick labels, bold sans for axis titles, grey #808080 for unit labels. Accents: orange #FFA500 dashed fit line and blue #1F77B4 data series. No shapes, borders, shadows, gradients or corner radii on the slide itself.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide background fill rectangle (Covers the entire slide, no stroke, no shadow.)
- illustration at x 9, y 27.1, w 82, h 45.2 [#FFFFFF]: Three-panel scaling-law figure: left Compute vs Test Loss log-scale plot with blue run curves, black envelope and orange dashed fit; middle Dataset Size vs Test Loss plot with blue points and grey fit; right Parameters vs Test Loss plot with blue points and grey fit (Raster image 1174 by 364 px, unboxed, not cropped, no overlay, no rounded corners; the white background of the plot area blends with the slide.)
- title at x 12.7, y 7.8, w 74.6, h 8.2 [Helvetica Neue Medium, sans-serif, bold-weight appearance; font size 8.19% of slide height; #000000; center aligned]: Scaling Laws for Language Models (Single line, no capitals styling, no letter-spacing change.)
- body at x 23.8, y 22.3, w 50.6, h 3 [Helvetica Neue, regular, sans-serif; font size 2.96% of slide height; #000000; center aligned]: Just train many models of different sizes on different sized datasets: (Single line, colon at end.)
- body at x 26.3, y 76.8, w 38.8, h 3 [Helvetica Neue, regular, sans-serif; font size 2.96% of slide height; #000000; center aligned]: Compute ~ Params * Tokens Processed, PF-day ~ (Part of the centred equation line, followed by the superscript equation at x 66.9 to 74.0.)
- body at x 66.9, y 76.5, w 1, h 3.6 [Computer Modern (CMR10), serif, regular; font size 3.62% of slide height; #000000; left aligned]: 9 (Math glyph of the equation 9 × 10^19.)
- body at x 68.3, y 76.5, w 1.6, h 3.6 [Computer Modern (CMSY10), serif, symbol font; font size 3.62% of slide height; #000000; left aligned]: × (Multiplication sign rendered from the CMSY10 symbol font.)
- body at x 70.4, y 76.5, w 2, h 3.6 [Computer Modern (CMR10), serif, regular; font size 3.62% of slide height; #000000; left aligned]: 10 (Base of the power of ten.)
- body at x 72.4, y 75.8, w 1.6, h 2.5 [Computer Modern (CMR7), serif, regular, superscript; font size 2.53% of slide height; #000000; left aligned]: 19 (Exponent, raised above the baseline of 10.)
- body at x 10.1, y 89.8, w 79.8, h 3.7 [Helvetica Neue, regular, sans-serif; font size 3.7% of slide height; #000000; center aligned]: These types of scaling laws are quite universal and are observed in most ML systems. (Single line, full stop at end.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 9.0, y 27.1, w 82.0, h 45.2: 1174 by 364 pixels
- text at x 12.7, y 7.8, w 74.6, h 8.2 [HelveticaNeue-Medium; font size 8.19; #000000]: Scaling Laws for Language Models
- text at x 23.8, y 22.3, w 50.6, h 3.0 [HelveticaNeue; font size 2.96; #000000]: Just train many models of diﬀerent sizes on diﬀerent sized datasets:
- text at x 66.9, y 76.5, w 1.0, h 3.6 [CMR10; font size 3.62; #000000]: 9
- text at x 68.3, y 76.5, w 1.6, h 3.6 [CMSY10; font size 3.62; #000000]: ⇥
- text at x 70.4, y 76.5, w 2.0, h 3.6 [CMR10; font size 3.62; #000000]: 10
- text at x 10.1, y 89.8, w 79.8, h 3.7 [HelveticaNeue; font size 3.70; #000000]: These types of scaling laws are quite universal and are observed in most ML systems.
- text at x 26.3, y 76.8, w 38.8, h 3.0 [HelveticaNeue; font size 2.96; #000000]: Compute ~ Params * Tokens Processed, PF-day ~
- text at x 72.4, y 75.8, w 1.6, h 2.5 [CMR7; font size 2.53; #000000]: 19

### Slide 16

**Text**

“Grand Unification” in AI
Identical Model & Loss, Different Data
Shadows on the Way / Wallace Stevens

The twilight burns with fire / And one by one the fires expire; / But I, beneath the window, breathe / On my accustomed place, / Until the west is emptied of desire / And I of memory.

Invisible now, the one that's gone / And that one gone before / Are in the tender distance made / More endurable.

But as for me, I wait / And that which I have always known / Stirs like a bird, and flies: / For in the dark of the window pane / I see the streets, the lighted door, / And I wait, where I have always waited, / Under the window, on the way.
Model Input
Completions →
Original

**Layout**

Two-line centred title spans the top of the slide, with the poem set in a left column and a dense image grid on the right. The title is centred horizontally at the top margin; the poem left edge sits at about 20% of slide width and the grid spans about 52% to 82% of width and 24% to 98% of height. No footer, logo or slide number is visible.

**Visuals**

A placed raster image forming an 8-row by 6-column grid of photographs, with a dark header strip labelled 'Model Input' at left, 'Completions →' above the middle four columns, and 'Original' at the right. Each row shows a model input thumbnail in the first column (dark-grey lower-left area filled by a black mask beneath it), followed by five model completions and then the original image. Row 1: a yellow fuzzy face-like ball on sand; completions show the same yellow ball with eyes in varied sand and ground shots. Row 2: a grey British shorthair cat on blue background; completions show the cat holding a sheet of paper, with varied paper and cat poses, ending with the cat holding a white card with a smiley face. Row 3: shadows of people across a zebra crosswalk; completions keep the striped crosswalk and shadows in grey tones. Row 4: a water drop falling into blue water with ripples; completions show droplets splashing and rippling in blue water. Row 5: an elephant with an American-flag-patterned ear viewed from the front on grass; completions show elephants in varied settings. Row 6: a group of fluffy brown-and-white birds perched side by side on a branch; completions show the birds in varied settings. Row 7: a stone lighthouse-like tower with a red-and-white house on green hillside under a blue sky; completions show varied landscapes. Row 8: a Go board with black and white stones; completions show Go boards with stones in different arrangements, one with a Thor logo caption strip at the bottom. Lighting is natural daylight in photographs; the grid is flat with thin dark gutters. Dominant colours: yellow, blue, grey, tan, and dark charcoal #15161B for the grid background.

**Design**

Background: Solid white #FFFFFF. Plain white slide with no ornament. Title text is black #000000 in Helvetica Neue Medium, centred, sans-serif, roughly 6.9% of slide height per line with normal weight, no italics or capitals. Poem and grid labels are dark grey to black, set in a regular-weight Helvetica-style sans-serif. The image grid has a near-black #15161B frame and background with small light grey #B0B3BA header labels. No shadows, gradients, rounded corners or accent colours.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide background fill (Solid rectangle covering the whole slide, no stroke.)
- title at x 28, y 6, w 46.2, h 6.9 [Helvetica Neue, Medium, regular roman, no caps, no letter-spacing; font size 6.94% of slide height; #000000; center aligned]: “Grand Unification” in AI (First title line, centred, one line.)
- title at x 15.9, y 14.4, w 68.2, h 6.9 [Helvetica Neue, Medium, regular roman, no caps, no letter-spacing; font size 6.94% of slide height; #000000; center aligned]: Identical Model & Loss, Different Data (Second title line, centred, directly below the first line.)
- body at x 19.4, y 28.4, w 22.9, h 65.5 [Helvetica Neue, Regular, roman, no caps, no letter-spacing; font size 2.3% of slide height; #000000; left aligned]: Shadows on the Way / Wallace Stevens / (blank line) / The twilight burns with fire / And one by one the fires expire; / But I, beneath the window, breathe / On my accustomed place, / Until the west is emptied of desire / And I of memory. / (blank line) / Invisible now, the one that's gone / And that one gone before / Are in the tender distance made / More endurable. / (blank line) / But as for me, I wait / And that which I have always known / Stirs like a bird, and flies: / For in the dark of the window pane / I see the streets, the lighted door, / And I wait, where I have always waited, / Under the window, on the way. (Poem set as a placed raster image of 312 by 503 pixels, so the box is an image of text. Line spacing about 1.4 times the font size. Title line and author are on consecutive lines, stanzas separated by one blank line. No bullets.)
- photo at x 51.8, y 24.4, w 30.6, h 73.5 [#15161B]: Grid of 8 rows by 6 columns of image completions (model input, five completions, original) under a dark header with labels 'Model Input', 'Completions →' and 'Original' (Placed raster image of 741 by 1002 pixels with a near-black #15161B frame and background. Header labels in small light grey #B0B3BA sans-serif, about 0.9% of slide height. Thin dark gutters between cells, square thumbnails, boxed (not cropped, not full-bleed), no overlay or tint. Row contents are listed in the visuals field.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 51.8, y 24.4, w 30.6, h 73.5: 741 by 1002 pixels
- image at x 19.4, y 28.4, w 22.9, h 65.5: 312 by 503 pixels
- text at x 28.0, y 6.0, w 46.2, h 6.9 [HelveticaNeue-Medium; font size 6.94; #000000]: “Grand Uniﬁcation” in AI
- text at x 15.9, y 14.4, w 68.2, h 6.9 [HelveticaNeue-Medium; font size 6.94; #000000]: Identical Model & Loss, Different Data

### Slide 17

**Text**

Representative Scales in AI / Neural Networks with ~1e12 parameters are becoming common, for example with ~100s of layers and ~10k x 10k matrices. / …and are trained on datasets with ~1e13 words (eg for Language Models) / “entire human-generated internet” ~ 1e15 words / State of the art GPUs (NVIDIA H100) perform ~1e15 flop/s / Largest training runs already use ~1e25 floating point operations total / For comparison… / ultra naive human brain estimate is ~1e15 “parameters” (~synapses) / lifetime compute of the brain is then ~3e24 “operations” (synapses fire ~1/sec)

**Layout**

Single-column text slide: the title is centred horizontally in the top band (about 7% from the top), and a left-aligned bulleted list starts at about 5% from the left edge and runs down to about 81% of the height, with two bullet levels. No footer, slide number, logo or image is present; the lower fifth of the slide is empty.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background with black Helvetica text throughout and no accent colours, shapes, lines, gradients or shadows. Title is Helvetica Bold at about 7.8% of slide height; body is Helvetica regular at about 4.1% of slide height with inline Helvetica Bold for the highlighted quantities (1e12 parameters, 1e13 words, 1e15 words, 1e15 flop/s, 1e25, 1e15 in the sub-bullet is regular). Bullets are solid black round dots, level-one at x 5.2% and level-two at x 7.4% for the first bullet line, sub-bullets indented to x 7.4% (dots) and text at x 9.7%.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide white background fill (covers the whole slide, no border)
- title at x 22.1, y 7.4, w 55.9, h 7.8 [Helvetica Bold, normal weight, no italics, no capitals, no letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Representative Scales in AI
- bullet list at x 5.2, y 24.7, w 83.7, h 55.9 [Helvetica regular, with Helvetica Bold for quantities 1e12 parameters, 1e13 words, 1e15 words, 1e15 flop/s and 1e25; no italics, no capitals, no letter-spacing; font size 4.07% of slide height; #000000; left aligned]: Neural Networks with ~1e12 parameters are becoming common, for example with ~100s of layers and ~10k x 10k matrices. / …and are trained on datasets with ~1e13 words (eg for Language Models) / “entire human-generated internet” ~ 1e15 words / State of the art GPUs (NVIDIA H100) perform ~1e15 flop/s / Largest training runs already use ~1e25 floating point operations total / For comparison… / ultra naive human brain estimate is ~1e15 “parameters” (~synapses) / lifetime compute of the brain is then ~3e24 “operations” (synapses fire ~1/sec) (Line spacing about 1.0 within a paragraph with extra space between bullets (about 1.5 lines); solid black round bullet dots at x 5.2% (level one, 6.11% size) and x 7.4% (level two, sub-bullets at x 7.4% with text at 9.7%); level-one bullets at y 24.7, 35.8, 42.4, 49.0, 55.7, 62.3, 69.0, 75.6 (dots) in the file; wrapped line of first bullet indented to text margin at x 7.4%; not overlapping any other element)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 22.1, y 7.4, w 55.9, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Representative Scales in AI
- text at x 5.2, y 24.7, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 5.2, y 35.8, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 7.4, y 25.6, w 23.6, h 4.1 [Helvetica; font size 4.07; #000000]: Neural Networks with ~
- text at x 31.1, y 25.6, w 18.1, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: 1e12 parameters
- text at x 49.2, y 25.6, w 42.5, h 4.1 [Helvetica; font size 4.07; #000000]: are becoming common, for example with
- text at x 5.2, y 42.4, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 7.4, y 30.0, w 41.8, h 4.1 [Helvetica; font size 4.07; #000000]: ~100s of layers and ~10k x 10k matrices.
- text at x 5.2, y 49.0, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 5.2, y 55.7, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 7.4, y 36.7, w 36.9, h 4.1 [Helvetica; font size 4.07; #000000]: …and are trained on datasets with ~
- text at x 44.3, y 36.7, w 12.5, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: 1e13 words
- text at x 56.8, y 36.7, w 26.9, h 4.1 [Helvetica; font size 4.07; #000000]: (eg for Language Models)
- text at x 5.2, y 62.3, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 7.4, y 43.3, w 36.9, h 4.1 [Helvetica; font size 4.07; #000000]: “entire human-generated internet” ~
- text at x 44.3, y 43.3, w 12.5, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: 1e15 words
- text at x 7.4, y 69.0, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 7.4, y 49.9, w 48.1, h 4.1 [Helvetica; font size 4.07; #000000]: State of the art GPUs (NVIDIA H100) perform ~
- text at x 7.4, y 75.6, w 1.2, h 6.1 [Helvetica; font size 6.11; #000000]: •
- text at x 55.5, y 49.9, w 11.8, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: 1e15 ﬂop/s
- text at x 7.4, y 56.6, w 35.4, h 4.1 [Helvetica; font size 4.07; #000000]: Largest training runs already use ~
- text at x 42.8, y 56.6, w 5.1, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: 1e25
- text at x 48.5, y 56.6, w 29.3, h 4.1 [Helvetica; font size 4.07; #000000]: ﬂoating point operations total
- text at x 7.4, y 63.2, w 18.2, h 4.1 [Helvetica; font size 4.07; #000000]: For comparison…
- text at x 9.7, y 69.9, w 69.6, h 4.1 [Helvetica; font size 4.07; #000000]: ultra naive human brain estimate is ~1e15 “parameters” (~synapses)
- text at x 9.7, y 76.5, w 79.2, h 4.1 [Helvetica; font size 4.07; #000000]: lifetime compute of the brain is then ~3e24 “operations” (synapses ﬁre ~1/sec)

### Slide 18

**Review notes**

- Review: Several pages are weak or unfinished: slide 18 overlaps two charts at illegible size, slide 19 is small and cramped, slide 35 is a dense block of quoted text, and slides 21 and 26 are lone headings.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Compute Usage in ML Training
Compute Used to Train a Frontier AI Model Rises Over Time (Accounting for Cost and Hardware Constraints)
TRAINING COMPUTE (FLOP)
1e+33
1e+32
1e+31
1e+30
1e+29
1e+28
1e+27
1e+26
1 MILLION × GPT-4
1,000 × GPT-4
GPT-4 ESTIMATED COMPUTE
2022 2024 2026 2028 2030 2032 2034 2036 2038 2040 2042
Training FLOPs
100T 10^16 10^18 10^20 10^22 10^24
2010 2012 2014 2016 2018 2020 2022
Release Date
6-layer MLP (MNIST)
Feedforward NN
RNN 500/10 + RT09 LM (NIST RT05)
KN5 LM + RNN 400/10 (WSJ)
AlexNet
TransE
Visualizing CNNs
SPP-net
VGG16
MSRA (C, PReLU)
AlphaGo Fan
AlphaGo Lee
GNMT
AlphaGo Master
AlphaGo Zero
AlphaStar
GPT-3
Gopher
ALIGN
MegaTuring
PaLM
Turing-NLG
Megatron
GPT-2
BERT
GPT-1
Transformer

**Layout**

Single title line centred at the top of a white slide, with two chart images below it: a large scatter chart in the lower left and a smaller area chart in the upper right, leaving open space in the upper left. The title sits roughly 7 percent from the top, and the charts are placed on an open, unaligned grid with no footer, slide number or logo.

**Visuals**

Two charts. (1) Left, a scatter-line chart about 42.8 percent of slide width, with a light blue-lavender plot area, white gridlines, a logarithmic y axis titled Training FLOPs with ticks 100T, 10^16, 10^18, 10^20, 10^22, 10^24, and a linear x axis titled Release Date with ticks 2010 to 2022 at two-year intervals. Two connected series: a red line from about 6-layer MLP (MNIST) at about 2010 and 1e14 FLOP rising through KN5 LM and RNN (about 2010.7, about 1e16) and AlexNet (about 2012.8, about 5e18) to AlphaGo Zero (about 2017.8, about 3e23) and AlphaStar (about 2019.9, about 2e23); a blue line from Transformer (about 2017.5, about 1e19) through BERT (about 2018.8, about 1e20), GPT-2 (about 2019.1, about 1e21), Megatron (about 2019.7, about 1e22), GPT-3 (about 2020.4, about 3e23), Gopher, ALIGN and MegaTuring to PaLM (about 2022.3, about 4e24). Grey dots mark unlabelled models. Labelled points are the marker points on each line; values are estimates marked about. No legend. (2) Right, an area-band chart in a box about 43.5 percent wide with a title above a black rule. Y axis titled TRAINING COMPUTE (FLOP), logarithmic, ticks 1e+26 to 1e+33; x axis 2022 to 2042 at two-year intervals. A solid red band with about 1e25.5 at 2022 widening to an upper edge about 8e30 and lower edge about 1.5e29 at 2042, with a white dashed midline reaching about 1e30 at 2042. Three horizontal light blue lines: 1 MILLION × GPT-4 at about 2e31, 1,000 × GPT-4 at about 2e28, and GPT-4 ESTIMATED COMPUTE at about 2e25, each labelled at its left end. Red band is the highlighted series; no legend.

**Design**

Background: Solid white #FFFFFF. Plain white slide with black bold Helvetica-style sans-serif title, centred. The left chart uses a light periwinkle plot background (#E3E9F7) with a red series (#E8392F) and a blue series (#6F7FE0) and small dark navy labels (#1F2A44) in a small sans-serif. The right chart uses a white background, a bold red band (#EE1C25), a white dashed midline, light blue reference lines (#5BA8D8), and dark grey sans-serif tick labels (#333333). Title text is black (#000000) and about 7.8 percent of slide height, which is the only text measured on the slide. No shadows, gradients or rounded corners.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers the whole slide, behind all other elements.)
- title at x 18.5, y 7.1, w 62.9, h 7.8 [Helvetica, bold, sentence-case, normal letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Compute Usage in ML Training
- chart at x 8.1, y 52.7, w 42.8, h 44.2 [#E3E9F7]: Scatter-line chart Training FLOPs versus Release Date, log y axis, with red and blue connected series, labelled model points and grey unlabelled dots (Raster image 1387 by 805 pixels, not cropped, boxed with no border; plot area fills the image; labels in small navy sans-serif text.)
- chart at x 48.4, y 18.6, w 43.5, h 36 [#EE1C25]: Area-band chart titled Compute Used to Train a Frontier AI Model Rises Over Time (Accounting for Cost and Hardware Constraints), with a red uncertainty band, white dashed midline and three blue reference lines (Raster image 767 by 588 pixels, not cropped, no border; a black rule sits under the chart title; the title is bold on the first phrase with a regular-weight parenthetical.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 8.1, y 52.7, w 42.8, h 44.2: 1387 by 805 pixels
- image at x 48.4, y 18.6, w 43.5, h 36.0: 767 by 588 pixels
- text at x 18.5, y 7.1, w 62.9, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Compute Usage in ML Training

### Slide 19

**Review notes**

- Review: Several pages are weak or unfinished: slide 18 overlaps two charts at illegible size, slide 19 is small and cramped, slide 35 is a dense block of quoted text, and slides 21 and 26 are lone headings.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Progress is Fast… Often faster than / Researchers and Forecasters Expect / Prediction Market Forecasts / 100 / 80 / 60 / 40 / 20 / 0 / 2021 / 2022 / 2023 / 2024 / 2025 / Year / Accuracy (%) / Competition Mathematics / (MATH Dataset) / Multiple Choice Exams / (law, economics, medicine, etc.) / david rein / @idavidrein / Claude 3 gets ~60% accuracy on GPQA. It's hard for me to understate / how hard these questions are—literal PhDs (in different domains from / the questions) with access to the internet get 34%. / / PhDs *in the same domain* (also with internet access!) get 65% - 75% / accuracy. / GPQA (Diamond) / Graduate level Q&A / 0-shot CoT / 50.4% / Maj@32 5-shot CoT / 59.5%

**Layout**

Single centred title spanning the top of the slide in two lines, with a chart on the left half and a tweet-style quote card on the right half, both placed in the lower-middle band. Left and right margins are roughly 5% and 4% respectively; no footer or slide number is visible.

**Visuals**

Left: a line chart titled 'Prediction Market Forecasts'. Y axis 'Accuracy (%)' ranging 0 to 100 with gridlines at 20, 40, 60, 80, 100 (labelled 0 to 100 in steps of 20). X axis 'Year' with ticks 2021, 2022, 2023, 2024, 2025. Series 1, blue (Competition Mathematics, MATH Dataset) with circle markers, values about 7 (2021), about 13 (2022), about 21 (2023), about 31 (2024), about 51 (2025). Series 2, orange (Multiple Choice Exams, law, economics, medicine, etc.) with circle markers, values about 49 (2021), about 57 (2022), about 64 (2023), about 69 (2024), about 76 (2025). Annotations: a blue arrow from the 2021 blue point up to near the 2022 blue point; a blue star at about (2022, 51); a blue arrow from about (2022, 51) to a blue star at about (2024, 69); an orange arrow from the 2021 orange point up to an orange star at about (2022, 66); an orange arrow from about (2022, 66) to an orange star at about (2023, 84). Legend boxed at right of the plot with thin grey border, entries with line-and-marker keys. Right: a screenshot of a tweet with an avatar photo of a man (small circular, dark hair, dark suit background) and a table-like box showing GPQA (Diamond), Graduate level Q&A; rows: 0-shot CoT 50.4%; Maj@32 5-shot CoT 59.5%. Box has thin light grey rounded border.

**Design**

Background: Solid white #FFFFFF. White background #FFFFFF. Title black #000000 in Helvetica Neue Medium, sans-serif, about 8.6% of slide height for cap height line, centred. Body text in tweet is dark grey #14171A sans-serif around 1.6% of slide height. Chart text dark grey #333333 sans-serif. Chart series blue #1F77B4 and orange #FF7F0E; annotation arrows and stars in blue #3333FF and orange #FF9900. Table values in bold serif (Times-like) black #000000. Lines thin, card border light grey #CFD9DE with rounded corners.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- title at x 7.1, y 4.7, w 82.5, h 18.4 [Helvetica Neue Medium, regular weight; font size 8.61% of slide height; #000000; center aligned]: Progress is Fast… Often faster than / Researchers and Forecasters Expect (Two lines, line spacing about 1.2x; second line starts at y 15.2%.)
- chart at x 4.7, y 36.1, w 48, h 40.8 [#1F77B4]: Line chart titled 'Prediction Market Forecasts' with blue and orange series, annotated with arrows and stars, and a legend (Chart image 922 by 441 px. Legend box at approx x 32.5%, y 40%, w 20%, h 28% with thin grey border #CCCCCC. Blue stars #3333FF, orange stars #FF9900, arrows with open heads, stroke about 2px.)
- photo at x 59, y 39.5, w 32, h 32.6 [Helvetica-like sans-serif for tweet text, serif bold for table values; font size 1.6% of slide height; #14171A; left aligned]: Tweet card by david rein (@idavidrein) with text and GPQA (Diamond) results table (Screenshot 589 by 338 px. Avatar small circle at top left of card. Table box has thin light grey rounded border #CFD9DE. Ellipsis icon '...' at top right in grey #536471.)
- body at x 59.5, y 45.5, w 30.5, h 12 [Helvetica Neue, regular; font size 1.6% of slide height; #14171A; left aligned]: Claude 3 gets ~60% accuracy on GPQA. It's hard for me to understate / how hard these questions are—literal PhDs (in different domains from / the questions) with access to the internet get 34%. / / PhDs *in the same domain* (also with internet access!) get 65% - 75% / accuracy. (Line spacing about 1.5x.)
- body at x 59.5, y 62, w 30.5, h 12 [Times-like serif, bold for values and title, regular for subtitle; font size 2% of slide height; #000000; left aligned]: GPQA (Diamond) / Graduate level Q&A / 0-shot CoT / 50.4% / Maj@32 5-shot CoT / 59.5% (Two rows: row one 0-shot CoT right-aligned around x 84% with 50.4% right-aligned at x 90.5%; row two Maj@32 5-shot CoT with 59.5% at same right edge.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 4.7, y 36.1, w 48.0, h 40.8: 922 by 441 pixels
- image at x 59.0, y 39.5, w 32.0, h 32.6: 589 by 338 pixels
- text at x 8.5, y 4.7, w 81.1, h 8.6 [HelveticaNeue-Medium; font size 8.61; #000000]: Progress is Fast… Often faster than
- text at x 7.1, y 15.2, w 82.5, h 8.6 [HelveticaNeue-Medium; font size 8.61; #000000]: Researchers and Forecasters Expect

### Slide 20

**Text**

A Cartoon of AI Capabilities
Approximate Capability Space and Tasks
Generality of Environment
IRL Robot
Multimodal
Code
Text
2027??
2025?
Claude 3
Claude 2
Minute
Hour
Day
Lifetime
Millenia
Timescale for a Person to Perform a Task

**Layout**

Single centred title in bold sans-serif across the top third of the slide, with a single chart image centred horizontally beneath it, occupying roughly the middle to lower-middle area. The chart sits in a boxed plot frame with its title above, the y-axis label rotated on the left and the x-axis label below the tick labels. No footer, slide number, logo or source line is visible.

**Visuals**

One chart image (711 by 585 pixels) placed at x 25.5%, y 22.8%, w 49.0%, h 71.7%. Type: capability-space scatter/area diagram with shaded regions. Chart title: 'Approximate Capability Space and Tasks' (bold, centred above the plot frame). X-axis: 'Timescale for a Person to Perform a Task' (bold), ticks labelled Minute, Hour, Day, Lifetime, Millenia from left to right, roughly evenly spaced in category steps with no numeric range. Y-axis: 'Generality of Environment' (bold, rotated), ticks labelled from bottom to top Text, Code, Multimodal, IRL Robot, with no numeric range. Two shaded regions starting at the origin: a darker grey region (about #D9D9D9) under the Claude 2 boundary, and a lighter grey region (about #EFEFEF) under the Claude 3 boundary, both bounded by black curves. Five-point stars (black filled) mark: Claude 2 (labelled above-right, sitting near the Hour tick at about 0.4 of the x-axis step between Minute and Hour, around Code level), Claude 3 (labelled above, at the right end of the shaded area just past Hour, around Multimodal-minus level), 2025? (labelled above, about between Hour and Day, around Multimodal level), and 2027?? (labelled above, about between Day and Lifetime, around IRL Robot level). Thin black vertical drop lines run from the Claude 2 and Claude 3 stars down to the baseline. No legend. Plot frame is a thin black rectangle open at the top right edge, with tick marks on axes. Colours: black #000000 text and lines, white background inside frame, light grey fills. No photograph, logo or icon.

**Design**

Background: Solid white #FFFFFF. Plain white slide with no decoration. Title in black #000000 Helvetica Neue Medium sans-serif, regular weight, about 10.4% of slide height, centred. Chart text in a sans-serif similar to Inter: axis tick labels regular and about 2.6% of slide height, axis titles and chart title bold. Chart lines are thin black (about 1 px) strokes; shaded regions are flat light greys with no gradients, shadows or rounded corners. Stars are solid black five-point shapes about 2% of slide height across.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers the entire slide, no border, no shadow.)
- title at x 11.1, y 9.1, w 74.6, h 10.4 [Helvetica Neue Medium, regular weight, sentence case; font size 10.37% of slide height; #000000; center aligned]: A Cartoon of AI Capabilities (Single line, no letter-spacing adjustment, no italics.)
- chart at x 25.5, y 22.8, w 49, h 71.7 [#000000; center aligned]: Approximate Capability Space and Tasks: capability-space diagram with shaded Claude 2 and Claude 3 regions, stars for Claude 2, Claude 3, 2025? and 2027??, axes Generality of Environment (Text, Code, Multimodal, IRL Robot) against Timescale for a Person to Perform a Task (Minute, Hour, Day, Lifetime, Millenia) (Placed image 711 by 585 px, unboxed with no border, no crop, no overlay. Plot frame is a thin black rectangle with top and right edges, shaded regions #D9D9D9 (Claude 2) and #EFEFEF (Claude 3), stars solid black, drop lines thin black, no legend, no gradient.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 25.5, y 22.8, w 49.0, h 71.7: 711 by 585 pixels
- text at x 11.1, y 9.1, w 74.6, h 10.4 [HelveticaNeue-Medium; font size 10.37; #000000]: A Cartoon of AI Capabilities

### Slide 21

**Review notes**

- Review: Several pages are weak or unfinished: slide 18 overlaps two charts at illegible size, slide 19 is small and cramped, slide 35 is a dense block of quoted text, and slides 21 and 26 are lone headings.

**Text**

Remaining Challenges for AI

**Layout**

A single centred line of title text sits horizontally and vertically centred on an otherwise empty white slide, with no footer, slide number or other elements. The text spans roughly 5.6% to 94.4% of the width, leaving near-equal side margins.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF). Title in bold Helvetica-style sans-serif, black #000000, set in a single line with normal tracking and no capitals treatment. Title size about 11.85% of slide height. No accents, rules, shapes or shadows.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Covers the entire slide; no stroke, no shadow, no opacity change.)
- title at x 5.6, y 39, w 88.8, h 11.9 [Helvetica Bold, sans-serif, bold weight, no italics, no capitals, normal letter-spacing; font size 11.85% of slide height; #000000; center aligned]: Remaining Challenges for AI (Single line, vertically centred on the slide.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 5.6, y 39.0, w 88.8, h 11.9 [Helvetica-Bold; font size 11.85; #000000]: Remaining Challenges for AI

### Slide 22

**Text**

Potential Challenges
NN based AI is best at “correlation” ~ “intuition”; it’s actually best at art and style.
A marble statue of a Koala DJ in front of a marble statue of a turntable. The Koala has wearing large marble headphones.
An art gallery displaying Monet paintings. The art gallery is flooded. Robots are going around the art gallery using paddle boards.
A photo of a Corgi dog riding a bike in Times Square. It is wearing sunglasses and a beach hat.
General AI systems tend to be worst at reasoning and math.

**Layout**

Single centred column: a bold title at the top centre, a one-line statement beneath it, a centred three-panel image strip in the middle, and a closing line below the strip. Everything is centre-aligned with generous white margins on all sides; there is no footer and no slide number.

**Visuals**

Three side-by-side AI-generated photographs in one strip, each with a black caption band beneath it. Panel 1 (left): a grey marble statue of a koala wearing large white headphones, seated behind a marble turntable console with hands on the decks; dark background, marble pillars on the sides; neutral grey and white tones. Panel 2 (centre): an art gallery with cream walls and framed Monet-style impressionist paintings in orange, teal and yellow; the floor is flooded with shallow dark water in which small robots ride yellow paddle boards, some robots with magenta and purple parts; warm and cool mixed tones. Panel 3 (right): a Corgi dog wearing yellow-tinted sunglasses and a beige fedora, riding a red bicycle handlebar in a Times Square street with a yellow taxi and pedestrians behind; warm orange, white and grey tones. Each panel carries a small blue 'Imagen' badge in its lower-right corner. The photographs are boxed, not full-bleed, with a thin dark border; they sit flush against each other with no gaps.

**Design**

Background: Solid white #FFFFFF. White background (#FFFFFF) with black (#000000) text throughout. Title is Helvetica Bold, black, about 7.8% of slide height, centred. Statement and closing line are Helvetica regular, about 4.1% of slide height, black, with key phrases in Helvetica Bold. Captions are in a light grey sans-serif (about #DDDDDD on black) at about 2% of slide height. Caption band background is black (#000000). Images have sharp square corners, no shadows, no gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide white background fill (covers the entire slide, behind everything)
- photo at x 22.4, y 26.9, w 55.2, h 46.2 [#000000]: Three-panel image strip of AI-generated images with black caption band: koala DJ statue, flooded Monet art gallery with robots on paddle boards, Corgi dog with sunglasses and beach hat on a bike in Times Square (boxed, not full-bleed; 1060 by 499 pixels; sharp corners; thin dark border; three panels abutting; caption band at bottom of strip in black with light grey text; small blue 'Imagen' badge at lower right of each panel)
- title at x 28.9, y 7.4, w 42.2, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; center aligned]: Potential Challenges
- body at x 8.3, y 18.1, w 83.4, h 4.1 [Helvetica Regular with 'best at art and style' in Helvetica Bold; font size 4.07% of slide height; #000000; center aligned]: NN based AI is best at “correlation” ~ “intuition”; it’s actually best at art and style. (two runs: 'NN based AI is best at “correlation” ~ “intuition”; it’s actually ' regular, then 'best at art and style' bold, then '.' regular)
- body at x 25.4, y 34.5, w 18.5, h 8.5 [Helvetica Regular; font size 1.9% of slide height; #DDDDDD; left aligned]: A marble statue of a Koala DJ in front of a marble statue of a turntable. The Koala has wearing large marble headphones. (caption in four lines inside black caption band under panel 1; line spacing about 2.3% of slide height)
- body at x 41.2, y 34.5, w 18.5, h 8.5 [Helvetica Regular; font size 1.9% of slide height; #DDDDDD; left aligned]: An art gallery displaying Monet paintings. The art gallery is flooded. Robots are going around the art gallery using paddle boards. (caption in four lines inside black caption band under panel 2)
- body at x 59.3, y 34.5, w 18.5, h 6.5 [Helvetica Regular; font size 1.9% of slide height; #DDDDDD; left aligned]: A photo of a Corgi dog riding a bike in Times Square. It is wearing sunglasses and a beach hat. (caption in three lines inside black caption band under panel 3)
- body at x 18.4, y 77.9, w 63.1, h 4.1 [Helvetica Regular with 'worst at reasoning and math' in Helvetica Bold; font size 4.07% of slide height; #000000; center aligned]: General AI systems tend to be worst at reasoning and math. (runs: 'General AI systems tend to be ' regular, 'worst at reasoning and math' bold, '.' regular)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 22.4, y 26.9, w 55.2, h 46.2: 1060 by 499 pixels
- text at x 28.9, y 7.4, w 42.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Potential Challenges
- text at x 8.3, y 18.1, w 61.3, h 4.1 [Helvetica; font size 4.07; #000000]: NN based AI is best at “correlation” ~ “intuition”; it’s actually
- text at x 69.6, y 18.1, w 21.5, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: best at art and style
- text at x 91.1, y 18.1, w 0.6, h 4.1 [Helvetica; font size 4.07; #000000]: .
- text at x 18.4, y 77.9, w 31.3, h 4.1 [Helvetica; font size 4.07; #000000]: General AI systems tend to be
- text at x 49.8, y 77.9, w 31.2, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: worst at reasoning and math
- text at x 80.9, y 77.9, w 0.6, h 4.1 [Helvetica; font size 4.07; #000000]: .

### Slide 23

**Text**

Remaining Technical Challenges / Multimodality — what about images, video, audio? I’m very confident this is easy / and not costly (e.g. Flamingo from DM). / Reasoning — Used to be my ~biggest doubt, looks tractable now in simple ways — / e.g. language models use “scratch pads” to do math and reasoning. / Long-Term Planning — Scary. Big doubt for some. I expect it’s easy via imitation of / humans, ie “planning is just a short-term task”. More specifically I think the only real / challenge is recognizing and fixing mistakes. / Efficient / Human-like Learning — There’s been major progress on sample efficient / RL for games, so it could be tractable. But in any case I don’t think we need this for / AI to have a transformative impact.

**Layout**

Single-column text slide with a centred bold title across the top at about 7% from the top edge. Below it, five left-aligned paragraphs start at a left margin of about 5% and are separated by blank vertical gaps; the first paragraph is struck through. No footer, logo or slide number is visible.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white slide with black Helvetica-style sans-serif type throughout. The title is Helvetica Bold, centred, about 7.8% of slide height. Body text is Helvetica regular at about 4.2% of slide height, with bold run-in lead-in labels (Multimodality, Reasoning, Long-Term Planning, Efficient / Human-like Learning). The first paragraph carries a black strikethrough line across its full text. No accent colour, shapes, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill covering the slide (Rectangle covering the whole slide, no border, no shadow.)
- title at x 16.6, y 7.4, w 66.8, h 7.8 [Helvetica Bold, regular weight, normal case, no letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Remaining Technical Challenges
- body at x 5.2, y 18.9, w 15.3, h 4.2 [Helvetica Bold, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: Multimodality (Black strikethrough line through the text.)
- body at x 20.5, y 18.9, w 71.9, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: — what about images, video, audio? I’m very confident this is easy (Black strikethrough line through the text; line spacing about 4.5% of slide height.)
- body at x 5.2, y 23.4, w 41.6, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: and not costly (e.g. Flamingo from DM). (Black strikethrough line through the text.)
- body at x 5.2, y 36.4, w 12.1, h 4.2 [Helvetica Bold, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: Reasoning
- body at x 17.3, y 36.4, w 76.6, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: — Used to be my ~biggest doubt, looks tractable now in simple ways —
- body at x 5.2, y 40.9, w 71, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: e.g. language models use “scratch pads” to do math and reasoning.
- body at x 5.2, y 53.9, w 22.9, h 4.2 [Helvetica Bold, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: Long-Term Planning
- body at x 28.1, y 53.9, w 66, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: — Scary. Big doubt for some. I expect it’s easy via imitation of
- body at x 5.2, y 58.4, w 88, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: humans, ie “planning is just a short-term task”. More specifically I think the only real
- body at x 5.2, y 62.9, w 46.8, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: challenge is recognizing and fixing mistakes.
- body at x 5.2, y 75.5, w 34.8, h 4.2 [Helvetica Bold, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: Efficient / Human-like Learning
- body at x 40, y 75.5, w 55, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: — There’s been major progress on sample efficient
- body at x 5.2, y 80, w 87.5, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: RL for games, so it could be tractable. But in any case I don’t think we need this for
- body at x 5.2, y 84.5, w 36.6, h 4.2 [Helvetica Regular, upright, normal case; font size 4.21% of slide height; #000000; left aligned]: AI to have a transformative impact.

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 16.6, y 7.4, w 66.8, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Remaining Technical Challenges
- text at x 5.2, y 18.9, w 15.3, h 4.2 [Helvetica-Bold; font size 4.21; #000000]: Multimodality
- text at x 20.5, y 18.9, w 71.9, h 4.2 [Helvetica; font size 4.21; #000000]: — what about images, video, audio? I’m very conﬁdent this is easy
- text at x 5.2, y 23.4, w 41.6, h 4.2 [Helvetica; font size 4.21; #000000]: and not costly (e.g. Flamingo from DM).
- text at x 5.2, y 36.4, w 12.1, h 4.2 [Helvetica-Bold; font size 4.21; #000000]: Reasoning
- text at x 17.3, y 36.4, w 76.6, h 4.2 [Helvetica; font size 4.21; #000000]: — Used to be my ~biggest doubt, looks tractable now in simple ways —
- text at x 5.2, y 40.9, w 71.0, h 4.2 [Helvetica; font size 4.21; #000000]: e.g. language models use “scratch pads” to do math and reasoning.
- text at x 5.2, y 53.9, w 22.9, h 4.2 [Helvetica-Bold; font size 4.21; #000000]: Long-Term Planning
- text at x 28.1, y 53.9, w 66.0, h 4.2 [Helvetica; font size 4.21; #000000]: — Scary. Big doubt for some. I expect it’s easy via imitation of
- text at x 5.2, y 58.4, w 88.0, h 4.2 [Helvetica; font size 4.21; #000000]: humans, ie “planning is just a short-term task”. More speciﬁcally I think the only real
- text at x 5.2, y 62.9, w 46.8, h 4.2 [Helvetica; font size 4.21; #000000]: challenge is recognizing and ﬁxing mistakes.
- text at x 5.2, y 75.5, w 34.8, h 4.2 [Helvetica-Bold; font size 4.21; #000000]: Efﬁcient / Human-like Learning
- text at x 40.0, y 75.5, w 55.0, h 4.2 [Helvetica; font size 4.21; #000000]: — There’s been major progress on sample efﬁcient
- text at x 5.2, y 80.0, w 87.5, h 4.2 [Helvetica; font size 4.21; #000000]: RL for games, so it could be tractable. But in any case I don’t think we need this for
- text at x 5.2, y 84.5, w 36.6, h 4.2 [Helvetica; font size 4.21; #000000]: AI to have a transformative impact.

### Slide 24

**Text**

External Objections to Transformative AI Soon
Running Low on Data? — This is a real concern, as e.g. language models already
train on datasets as large as “all of the books in the world”. I don’t expect it will be a
blocker (e.g. there’s way, way more data in video, and also way more text on the
internet) but it’s a concern.
WTF’s Going on with the Economy? — Probably the most compelling argument for
doubting rapid AI progress in my opinion — if we’ll be capable of automating knowledge
work in ~5 years, shouldn’t the AI share of the economy already be much larger?
Isn’t this all / aren’t you just crazy? — Definitely possible! It’s a huge source of doubt.

**Layout**

Single-column text slide on a plain white field with a left-aligned title across the top and three left-aligned text blocks stacked below it, each opening with a bold question followed by an em-dash answer. Left margin is about 5.2% of the slide width; no footer, logo or slide number is visible.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no gradients, shapes, rules or imagery. Title and body text are pure black (#000000) in a Helvetica-style grotesque sans-serif; bold is used for the title and the three question lead-ins, regular weight for the answers. Title is about 7.3% of slide height, body text about 4.1% of slide height. Body lines are spaced about 4.4% of slide height apart, with larger gaps between the three paragraphs.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide white background fill (covers the entire slide, no border)
- title at x 5.5, y 7.3, w 89, h 7.3 [Helvetica Bold, bold, regular case, no letter-spacing; font size 7.31% of slide height; #000000; left aligned]: External Objections to Transformative AI Soon
- body at x 5.2, y 24.2, w 85, h 16.5 [Helvetica, regular, with bold lead-in 'Running Low on Data?', no italics, sentence case; font size 4.07% of slide height; #000000; left aligned]: Running Low on Data? — This is a real concern, as e.g. language models already / train on datasets as large as “all of the books in the world”. I don’t expect it will be a / blocker (e.g. there’s way, way more data in video, and also way more text on the / internet) but it’s a concern. (line spacing about 4.4% of slide height; bold lead-in on first line only; one paragraph)
- body at x 5.2, y 50.7, w 89.3, h 13 [Helvetica, regular, with bold lead-in 'WTF’s Going on with the Economy?', no italics, sentence case; font size 4.07% of slide height; #000000; left aligned]: WTF’s Going on with the Economy? — Probably the most compelling argument for / doubting rapid AI progress in my opinion — if we’ll be capable of automating knowledge / work in ~5 years, shouldn’t the AI share of the economy already be much larger? (line spacing about 4.4% of slide height; bold lead-in on first line only; one paragraph)
- body at x 5.2, y 72.8, w 89, h 4.1 [Helvetica, regular, with bold lead-in 'Isn’t this all / aren’t you just crazy?', no italics, sentence case; font size 4.07% of slide height; #000000; left aligned]: Isn’t this all / aren’t you just crazy? — Definitely possible! It’s a huge source of doubt. (single line; bold lead-in on the question only)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 5.5, y 7.3, w 89.0, h 7.3 [Helvetica-Bold; font size 7.31; #000000]: External Objections to Transformative AI Soon
- text at x 5.2, y 24.2, w 24.9, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: Running Low on Data?
- text at x 30.2, y 24.2, w 60.3, h 4.1 [Helvetica; font size 4.07; #000000]: — This is a real concern, as e.g. language models already
- text at x 5.2, y 28.6, w 85.0, h 4.1 [Helvetica; font size 4.07; #000000]: train on datasets as large as “all of the books in the world”. I don’t expect it will be a
- text at x 5.2, y 33.0, w 82.0, h 4.1 [Helvetica; font size 4.07; #000000]: blocker (e.g. there’s way, way more data in video, and also way more text on the
- text at x 5.2, y 37.5, w 27.0, h 4.1 [Helvetica; font size 4.07; #000000]: internet) but it’s a concern.
- text at x 5.2, y 50.7, w 39.1, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: WTF’s Going on with the Economy?
- text at x 44.3, y 50.7, w 47.9, h 4.1 [Helvetica; font size 4.07; #000000]: — Probably the most compelling argument for
- text at x 5.2, y 55.2, w 89.3, h 4.1 [Helvetica; font size 4.07; #000000]: doubting rapid AI progress in my opinion — if we’ll be capable of automating knowledge
- text at x 5.2, y 59.6, w 81.7, h 4.1 [Helvetica; font size 4.07; #000000]: work in ~5 years, shouldn’t the AI share of the economy already be much larger?
- text at x 5.2, y 72.8, w 38.1, h 4.1 [Helvetica-Bold; font size 4.07; #000000]: Isn’t this all / aren’t you just crazy?
- text at x 43.3, y 72.8, w 50.9, h 4.1 [Helvetica; font size 4.07; #000000]: — Deﬁnitely possible! It’s a huge source of doubt.

### Slide 25

**Text**

Other Observations / AI has actually been cheap so far. The most expensive AI training runs / (published) are still in the ~$100M range. So even on the scale of pure / science projects, we still have 100X more room for growth, and we expect / major cost improvements in chips/datacenters. / If you can train one powerful AI, you can run millions of them. Training / costs are basically (parameter count) * (training set size). But efﬁciency / means that parameters ~ training set size. So larger AIs cost more and more / to train, quadratically in their size. But running the AI has a cost linear in size. / With all that hardware you needed to train the AI, you can run at least millions / of copies of it in parallel.

**Layout**

Single left-aligned text column on a plain white slide with the bold centred title "Other Observations" across the top at about 7% of the slide height. Two body paragraphs start at a left margin of about 5.2% and run to about 95% of the width, with a blank line gap between them; there is no footer, no slide number and no other element.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background, no decoration, shapes, rules or gradients. Title is Helvetica Bold, black #000000, about 7.8% of slide height, centred. Body is Helvetica regular, black #000000, about 4.6% of slide height, with inline bold and oblique runs; line spacing about 5% of slide height (baseline pitch 5.0%). No accent colours.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: white full-slide fill covering the slide (No border, no shadow, full bleed; sits behind all text.)
- title at x 29.9, y 7.4, w 40.1, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; center aligned]: Other Observations
- body at x 5.2, y 24.6, w 42.4, h 4.6 [Helvetica Bold; font size 4.63% of slide height; #000000; left aligned]: AI has actually been cheap so far.
- body at x 47.6, y 24.6, w 42.1, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: The most expensive AI training runs
- body at x 5.2, y 29.6, w 82.6, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: (published) are still in the ~$100M range. So even on the scale of pure
- body at x 5.2, y 34.6, w 85.7, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: science projects, we still have 100X more room for growth, and we expect
- body at x 5.2, y 39.6, w 6.5, h 4.6 [Helvetica Oblique; font size 4.63% of slide height; #000000; left aligned]: major
- body at x 11.7, y 39.6, w 47.2, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: cost improvements in chips/datacenters.
- body at x 5.2, y 54, w 12.9, h 4.6 [Helvetica Bold; font size 4.63% of slide height; #000000; left aligned]: If you can
- body at x 18.1, y 54, w 5.6, h 4.6 [Helvetica Bold Oblique; font size 4.63% of slide height; #000000; left aligned]: train
- body at x 23.7, y 54, w 32.2, h 4.6 [Helvetica Bold; font size 4.63% of slide height; #000000; left aligned]: one powerful AI, you can
- body at x 55.9, y 54, w 4.2, h 4.6 [Helvetica Bold Oblique; font size 4.63% of slide height; #000000; left aligned]: run
- body at x 60.1, y 54, w 20.7, h 4.6 [Helvetica Bold; font size 4.63% of slide height; #000000; left aligned]: millions of them
- body at x 80.8, y 54, w 2.2, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: .
- body at x 82.9, y 54, w 9.3, h 4.6 [Helvetica Oblique; font size 4.63% of slide height; #000000; left aligned]: Training
- body at x 5.2, y 59, w 83.4, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: costs are basically (parameter count) * (training set size). But efﬁciency
- body at x 5.2, y 64, w 89.5, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: means that parameters ~ training set size. So larger AIs cost more and more
- body at x 5.2, y 69, w 44.7, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: to train, quadratically in their size. But
- body at x 49.9, y 69, w 16.1, h 4.6 [Helvetica Oblique; font size 4.63% of slide height; #000000; left aligned]: running the AI
- body at x 66, y 69, w 30, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: has a cost linear in size.
- body at x 5.2, y 74, w 89.9, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: With all that hardware you needed to train the AI, you can run at least millions
- body at x 5.2, y 79, w 27.9, h 4.6 [Helvetica Regular; font size 4.63% of slide height; #000000; left aligned]: of copies of it in parallel.

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 29.9, y 7.4, w 40.1, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Other Observations
- text at x 5.2, y 24.6, w 42.4, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: AI has actually been cheap so far.
- text at x 47.6, y 24.6, w 42.1, h 4.6 [Helvetica; font size 4.63; #000000]: The most expensive AI training runs
- text at x 5.2, y 29.6, w 82.6, h 4.6 [Helvetica; font size 4.63; #000000]: (published) are still in the ~$100M range. So even on the scale of pure
- text at x 5.2, y 34.6, w 85.7, h 4.6 [Helvetica; font size 4.63; #000000]: science projects, we still have 100X more room for growth, and we expect
- text at x 5.2, y 39.6, w 6.5, h 4.6 [Helvetica-Oblique; font size 4.63; #000000]: major
- text at x 11.7, y 39.6, w 47.2, h 4.6 [Helvetica; font size 4.63; #000000]: cost improvements in chips/datacenters.
- text at x 5.2, y 54.0, w 12.9, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: If you can
- text at x 18.1, y 54.0, w 5.6, h 4.6 [Helvetica-BoldOblique; font size 4.63; #000000]: train
- text at x 23.7, y 54.0, w 32.2, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: one powerful AI, you can
- text at x 55.9, y 54.0, w 4.2, h 4.6 [Helvetica-BoldOblique; font size 4.63; #000000]: run
- text at x 60.1, y 54.0, w 20.7, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: millions of them
- text at x 80.8, y 54.0, w 2.2, h 4.6 [Helvetica; font size 4.63; #000000]: .
- text at x 82.9, y 54.0, w 9.3, h 4.6 [Helvetica-Oblique; font size 4.63; #000000]: Training
- text at x 5.2, y 59.0, w 83.4, h 4.6 [Helvetica; font size 4.63; #000000]: costs are basically (parameter count) * (training set size). But efﬁciency
- text at x 5.2, y 64.0, w 89.5, h 4.6 [Helvetica; font size 4.63; #000000]: means that parameters ~ training set size. So larger AIs cost more and more
- text at x 5.2, y 69.0, w 44.7, h 4.6 [Helvetica; font size 4.63; #000000]: to train, quadratically in their size. But
- text at x 49.9, y 69.0, w 16.1, h 4.6 [Helvetica-Oblique; font size 4.63; #000000]: running the AI
- text at x 66.0, y 69.0, w 30.0, h 4.6 [Helvetica; font size 4.63; #000000]: has a cost linear in size.
- text at x 5.2, y 74.0, w 89.9, h 4.6 [Helvetica; font size 4.63; #000000]: With all that hardware you needed to train the AI, you can run at least millions
- text at x 5.2, y 79.0, w 27.9, h 4.6 [Helvetica; font size 4.63; #000000]: of copies of it in parallel.

### Slide 26

**Review notes**

- Review: Several pages are weak or unfinished: slide 18 overlaps two charts at illegible size, slide 19 is small and cramped, slide 35 is a dense block of quoted text, and slides 21 and 26 are lone headings.

**Text**

Why This is Scary

**Layout**

A single centered title line sits in the middle of the slide, horizontally and vertically centered, with no footer, slide number, or other elements; the text is centered within the slide with generous margins on all sides.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no decoration. The single title is set in a bold grotesque sans-serif (Helvetica Bold) in black (#000000), sentence-style capitalisation, no letter-spacing adjustments, about 12.8% of slide height for the cap-height-inclusive line. No accent colours, no shapes, lines, shadows, or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background fill (Solid fill covering the entire slide, no border, no shadow, no gradient.)
- title at x 19.8, y 41.9, w 60.3, h 12.8 [Helvetica Bold, weight bold, normal case, no italics, no letter-spacing; font size 12.78% of slide height; #000000; center aligned]: Why This is Scary (Single line of text, vertically centred in the slide at roughly the middle of the slide height.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 19.8, y 41.9, w 60.3, h 12.8 [Helvetica-Bold; font size 12.78; #000000]: Why This is Scary

### Slide 27

**Text**

This is Scary for Common Sense Reasons
This is very uncharted territory, and we’re moving fast. This is “the
biggest deal” since the industrial revolution, and it may happen much faster. If
this is true, it’s the highest stakes event in human history.
Maybe supervising a thing that’s smarter than us is hard. Maybe not. But
once you make a thing that’s broadly much smarter than you — and given
that it’d be easy to run millions of copies of that thing (once you have one) —
you’re going to lose and be disempowered if there’s a conflict. Given the
stakes, being “90% sure it’ll work out” is very far from OK!
We currently understand very little about how AI works, how it learns, how it
generalizes, etc. AI is a 10 year old field with ~no quantitative predictions.

**Layout**

Single-column text slide on a plain white field. The title is centred horizontally near the top (about 7% from the top edge). Three body paragraphs are left-aligned at about 5% from the left edge, separated by blank line gaps, with the last paragraph ending near 90% of the height. There is no footer, page number, logo or other furniture.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white slide with no accent colours. Title and body text are pure black #000000 in Helvetica (sans-serif). Title is Helvetica Bold at about 7.8% of slide height; body is Helvetica regular at about 4.6% of slide height, with bold lead-in phrases and one italic phrase (very far). Body line spacing is about 5% of slide height. No rules, boxes, shapes, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide white fill covering the slide (no outline, no shadow, opacity 100%)
- title at x 6.9, y 7.4, w 86.1, h 7.8 [Helvetica Bold, sans-serif, bold, no italics, no capitals, normal letter-spacing; font size 7.78% of slide height; #000000; center aligned]: This is Scary for Common Sense Reasons
- body at x 5.2, y 21.5, w 89.9, h 14.6 [Helvetica regular, sans-serif, with lead-in sentence in Helvetica Bold, normal letter-spacing; font size 4.63% of slide height; #000000; left aligned]: This is very uncharted territory, and we’re moving fast. This is “the / biggest deal” since the industrial revolution, and it may happen much faster. If / this is true, it’s the highest stakes event in human history. (three lines at baselines 21.5, 26.5 and 31.5 percent from top; line spacing about 5% of slide height)
- body at x 5.2, y 45.9, w 89.5, h 24.6 [Helvetica regular, sans-serif, with “Maybe supervising a thing that’s smarter than us is hard” in Helvetica Bold and “very far” in Helvetica Oblique, normal letter-spacing; font size 4.63% of slide height; #000000; left aligned]: Maybe supervising a thing that’s smarter than us is hard. Maybe not. But / once you make a thing that’s broadly much smarter than you — and given / that it’d be easy to run millions of copies of that thing (once you have one) — / you’re going to lose and be disempowered if there’s a conflict. Given the / stakes, being “90% sure it’ll work out” is very far from OK! (five lines at 45.9, 50.9, 55.9, 60.9 and 65.9 percent from top; line spacing about 5% of slide height; a blank gap separates this paragraph from the previous one)
- body at x 5.2, y 80.4, w 87.7, h 9.6 [Helvetica regular, sans-serif, normal weight, normal letter-spacing; font size 4.63% of slide height; #000000; left aligned]: We currently understand very little about how AI works, how it learns, how it / generalizes, etc. AI is a 10 year old field with ~no quantitative predictions. (two lines at 80.4 and 85.4 percent from top; line spacing about 5% of slide height)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 6.9, y 7.4, w 86.1, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: This is Scary for Common Sense Reasons
- text at x 5.2, y 21.5, w 68.4, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: This is very uncharted territory, and we’re moving fast.
- text at x 73.6, y 21.5, w 13.5, h 4.6 [Helvetica; font size 4.63; #000000]: This is “the
- text at x 5.2, y 26.5, w 89.9, h 4.6 [Helvetica; font size 4.63; #000000]: biggest deal” since the industrial revolution, and it may happen much faster. If
- text at x 5.2, y 31.5, w 65.6, h 4.6 [Helvetica; font size 4.63; #000000]: this is true, it’s the highest stakes event in human history.
- text at x 5.2, y 45.9, w 69.8, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: Maybe supervising a thing that’s smarter than us is hard
- text at x 75.0, y 45.9, w 19.7, h 4.6 [Helvetica; font size 4.63; #000000]: . Maybe not. But
- text at x 5.2, y 50.9, w 85.8, h 4.6 [Helvetica; font size 4.63; #000000]: once you make a thing that’s broadly much smarter than you — and given
- text at x 5.2, y 55.9, w 89.0, h 4.6 [Helvetica; font size 4.63; #000000]: that it’d be easy to run millions of copies of that thing (once you have one) —
- text at x 5.2, y 60.9, w 83.8, h 4.6 [Helvetica; font size 4.63; #000000]: you’re going to lose and be disempowered if there’s a conﬂict. Given the
- text at x 5.2, y 65.9, w 46.5, h 4.6 [Helvetica; font size 4.63; #000000]: stakes, being “90% sure it’ll work out” is
- text at x 51.7, y 65.9, w 8.7, h 4.6 [Helvetica-Oblique; font size 4.63; #000000]: very far
- text at x 60.3, y 65.9, w 11.1, h 4.6 [Helvetica; font size 4.63; #000000]: from OK!
- text at x 5.2, y 80.4, w 87.7, h 4.6 [Helvetica; font size 4.63; #000000]: We currently understand very little about how AI works, how it learns, how it
- text at x 5.2, y 85.4, w 84.5, h 4.6 [Helvetica; font size 4.63; #000000]: generalizes, etc. AI is a 10 year old ﬁeld with ~no quantitative predictions.

### Slide 28

**Text**

Why You Can Contribute
• Can’t over-emphasize how simple AI is or how few ideas there are…
• AI with NNs is actually natural science, but few people have fully internalized this. My experience has been that the perspectives of most established researchers (both the engineering and theorem-proving mindsets) put them at a disadvantage compared to physicists who enter the field fresh, in terms of doing impactful work.
• If progress is half as rapid as I expect, we really do need our best and brightest to have their eyes on the ball to ensure this goes well.

**Layout**

Single-column text slide on a plain white field. The title is centred horizontally in the upper band at about 11% from the top. A left-aligned bulleted list starts at about 29% from the top and fills the left-to-centre area, with the bullet at about 5% from the left edge and wrapped text hanging at about 7.4%. No footer, slide number, logo or image; the lower quarter of the slide is empty.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no gradient, pattern or texture. Title is Helvetica Bold (sans-serif), black #000000, about 7.8% of slide height, centred. Body is Helvetica regular, black #000000, about 4.6% of slide height, with line spacing of roughly 5% of slide height between wrapped lines and about 1.9% extra between bullets. Bold is used for 'actually natural science' and italic (Helvetica Oblique) for 'disadvantage'. Bullets are solid round black dots in Helvetica. No accent colour, shapes, lines, shadows, or corner radii.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide background fill covering the whole slide (Solid fill, no outline, no shadow, opacity 100%.)
- title at x 24.9, y 11.4, w 50.2, h 7.8 [Helvetica Bold, sans-serif, normal capitals and case, no letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Why You Can Contribute
- bullet list at x 5.2, y 29.3, w 90.1, h 45 [Helvetica, sans-serif, regular weight, with bold and oblique runs; font size 4.63% of slide height; #000000; left aligned]: • Can’t over-emphasize how simple AI is or how few ideas there are… / • AI with NNs is actually natural science, but few people have fully / internalized this. My experience has been that the perspectives of most / established researchers (both the engineering and theorem-proving / mindsets) put them at a disadvantage compared to physicists who enter the / field fresh, in terms of doing impactful work. / • If progress is half as rapid as I expect, we really do need our best and / brightest to have their eyes on the ball to ensure this goes well. (Three bullets of solid black round dots at x 5.2% with text at x 7.4%; wrapped lines are indented to match the text start. Line pitch about 5% of slide height. Bold run 'actually natural science'; italic run 'disadvantage'. Bullet 2 wraps over 4 lines, bullet 3 over 2 lines, with a 1.9% gap between bullets. Right edge of the longest line is at about 95.3%.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 24.9, y 11.4, w 50.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Why You Can Contribute
- text at x 5.2, y 29.3, w 1.4, h 6.9 [Helvetica; font size 6.94; #000000]: •
- text at x 5.2, y 36.5, w 1.4, h 6.9 [Helvetica; font size 6.94; #000000]: •
- text at x 7.4, y 30.3, w 78.7, h 4.6 [Helvetica; font size 4.63; #000000]: Can’t over-emphasize how simple AI is or how few ideas there are…
- text at x 7.4, y 37.5, w 16.9, h 4.6 [Helvetica; font size 4.63; #000000]: AI with NNs is
- text at x 24.4, y 37.5, w 29.4, h 4.6 [Helvetica-Bold; font size 4.63; #000000]: actually natural science
- text at x 53.8, y 37.5, w 30.8, h 4.6 [Helvetica; font size 4.63; #000000]: , but few people have fully
- text at x 5.2, y 63.7, w 1.4, h 6.9 [Helvetica; font size 6.94; #000000]: •
- text at x 7.4, y 42.5, w 82.7, h 4.6 [Helvetica; font size 4.63; #000000]: internalized this. My experience has been that the perspectives of most
- text at x 7.4, y 47.5, w 78.5, h 4.6 [Helvetica; font size 4.63; #000000]: established researchers (both the engineering and theorem-proving
- text at x 7.4, y 52.5, w 27.9, h 4.6 [Helvetica; font size 4.63; #000000]: mindsets) put them at a
- text at x 35.4, y 52.5, w 15.5, h 4.6 [Helvetica-Oblique; font size 4.63; #000000]: disadvantage
- text at x 50.9, y 52.5, w 44.4, h 4.6 [Helvetica; font size 4.63; #000000]: compared to physicists who enter the
- text at x 7.4, y 57.5, w 50.1, h 4.6 [Helvetica; font size 4.63; #000000]: ﬁeld fresh, in terms of doing impactful work.
- text at x 7.2, y 64.7, w 80.9, h 4.6 [Helvetica; font size 4.63; #000000]: If progress is half as rapid as I expect, we really do need our best and
- text at x 7.2, y 69.7, w 71.9, h 4.6 [Helvetica; font size 4.63; #000000]: brightest to have their eyes on the ball to ensure this goes well
- text at x 79.1, y 70.1, w 0.6, h 4.1 [Helvetica; font size 4.07; #000000]: .

### Slide 29

**Text**

Thanks!

**Layout**

Single centred line of text on a plain white field, set horizontally and vertically near the centre of the slide. No title hierarchy, footer, slide number or margins are visible beyond the centred word.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background (#FFFFFF) with no gradient, pattern or texture. The single word is set in bold Helvetica-style sans-serif (Helvetica Bold or Arial Bold) in pure black (#000000), large relative to the slide height, with no accent colours, rules or shapes.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white fill covering the entire slide background (Solid fill, no border, no shadow, opacity 100%, sits at the back)
- title at x 36.4, y 41.9, w 27.1, h 12.8 [Helvetica Bold (sans-serif, bold weight, normal case, no italics, no letter-spacing); font size 12.78% of slide height; #000000; center aligned]: Thanks! (Single line, no bullets, no shadow; text sits at the vertical centre of the slide)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 36.4, y 41.9, w 27.1, h 12.8 [Helvetica-Bold; font size 12.78; #000000]: Thanks!

### Slide 30

**Review notes**

- Review: The preference modelling, RL and Constitutional AI pages (31 to 34) sit after the 'Thanks!' slide (29) and a repeated Claude demo (30), and slides 13, 14 and 30 show tiny screenshots that cannot be read.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Claude
Liouville Theory: An Introduction to Rigorous Approaches
Sourav Chatterjee1 and Edward Witten2
1Department of Statistics, Stanford University, Stanford CA 94305 USA
2School of Natural Sciences, Institute for Advanced Study, Princeton, NJ 08540 USA
ABSTRACT: In recent years, a surprisingly direct and simple rigorous understanding of quantum Liouville theory has developed. We aim here to make this material more accessible to physicists working on quantum field theory.
arXiv:2404.02001v2 [hep-th] 5 Apr 2024
\section{Liouville Theory Correlators: Overview} \label{correlators} \subsection{The Gaussian Free Field} Though Liouville theory is conformally invariant, to analyze Liouville theory ...
PASTED
JK
What are some ideas in this paper that might be new to theoretical physicists who already have some familiarity with Liouville Theory?
Based on the contents of the paper, a few ideas that might be new to theoretical physicists already familiar with Liouville theory include:
1. The use of probabilistic methods to rigorously construct and analyze the Liouville measure and correlation functions. Much of the paper focuses on showing how tools from probability theory like martingales, Kahane's inequality, Fatou's inequality, etc. can be used to prove convergence of the random measures.
2. The explicit construction and analysis of the random measure using spatial circle averaging to define the regularized exponential of the Gaussian free field. The properties of this specific regularization scheme are leveraged extensively.
3. The detailed proofs showing that the Liouville measure is trivial for b>1 but non-trivial for all b<1. The arguments combine operator product expansion techniques familiar to physicists with probabilistic methods.
4. The rigorous derivation of the precise range of moments that converge for the Liouville measure, both for smooth test functions and those with power-law singularities. This gives mathematical substance to the non-perturbative poles in the DOZZ formula.
5. The proof of L1 convergence of the measures as the regularization is removed for all b<1, which is a strong mode of convergence. This goes beyond just analyzing moments.
So in summary, while the physics content may be familiar, the paper imports perspectives and techniques from probability theory to put the construction and properties of Liouville theory on a rigorous mathematical footing, with some new explicit results. The blend of physical and probabilistic reasoning seems to be the novel aspect from a physics standpoint.
Debug Copy Retry
Claude can make mistakes. Please double-check responses.

**Layout**

Two-panel slide on a white background. The title Claude sits top-left, above a left column containing a cropped arXiv paper title page. A right column holds a Claude chat screenshot spanning nearly full height, with the chat window inset from the right edge. There is no footer or slide number on the slide.

**Visuals**

Left: a placed screenshot of the first page of arXiv paper 2404.02001v2, showing the title Liouville Theory: An Introduction to Rigorous Approaches in bold sans-serif, a horizontal rule, author names Sourav Chatterjee and Edward Witten with affiliations in small serif italic, and a serif abstract block; the vertical arXiv stamp runs up the left edge of the image. Right: a Claude chat interface screenshot showing a pasted LaTeX snippet card with a blue PASTED pill, a user message bubble with a dark circular JK avatar, and a long numbered assistant reply in serif text with a summary paragraph, with a Debug/Copy/Retry toolbar and a Claude logo in coral at bottom left. Both images are boxed, not full-bleed.

**Design**

Background: Solid white #FFFFFF. White slide background. Title Claude in bold Helvetica-style sans-serif, black #000000, about 7.8% of slide height for cap height region. Paper text is serif in black #000000 and bold sans-serif for paper heading. Chat panel background is light warm grey #EEEDE6 with user bubble #D9D3C4 and assistant panel near-white #F7F6F2. Accent coral #D97757 for the Claude logo, blue #1A73E8 for the PASTED pill. Corners on chat window and pasted card are rounded, with a soft shadow on the card. No gradients on the slide itself.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- title at x 16.1, y 5.3, w 14.4, h 7.8 [Helvetica Bold; font size 7.78% of slide height; #000000; left aligned]: Claude (Bold weight, normal case, no letter-spacing change)
- photo at x 5.8, y 20.3, w 38.2, h 59.4 [#FFFFFF]: Screenshot of arXiv paper title page with vertical arXiv stamp, boxed, not cropped (Rectangular image 737 by 644 pixels, white page background, no border)
- photo at x 54.4, y 1.4, w 39.4, h 95.4 [#EEEDE6]: Screenshot of Claude chat interface with pasted card, user message, numbered assistant reply and toolbar (Rectangular image 757 by 1030 pixels, rounded window corners, light warm grey panel, soft drop shadow on chat window)
- body at x 11.3, y 30.3, w 25.2, h 4.2 [Helvetica Bold; font size 1.9% of slide height; #000000; left aligned]: Liouville Theory: An Introduction to Rigorous / Approaches (Inside paper screenshot; bold sans-serif heading, two lines)
- line at x 11.3, y 39.2, w 25.2, h 0.2 [#000000]: Horizontal rule under paper title (Black stroke about 1px high, spans heading width)
- body at x 11.3, y 43.5, w 25.2, h 6.8 [Computer Modern Serif; font size 1% of slide height; #000000; left aligned]: Sourav Chatterjee1 and Edward Witten2 / 1Department of Statistics, Stanford University, Stanford CA 94305 USA / 2School of Natural Sciences, Institute for Advanced Study, Princeton, NJ 08540 USA (Authors in bold serif, affiliations in italic serif; inside paper screenshot)
- body at x 11.3, y 52, w 25.2, h 5 [Computer Modern Serif; font size 0.9% of slide height; #000000; left aligned]: ABSTRACT: In recent years, a surprisingly direct and simple rigorous understanding of quantum Liouville theory has developed. We aim here to make this material more accessible to physicists working on quantum field theory. (Small caps ABSTRACT label, justified serif paragraph, three lines inside paper screenshot)
- caption at x 6.5, y 17, w 1.8, h 25 [Computer Modern Serif; font size 1% of slide height; #555555; center aligned]: arXiv:2404.02001v2 [hep-th] 5 Apr 2024 (Rotated 90 degrees counterclockwise, reading bottom to top, grey serif stamp along left edge of paper screenshot)
- shape at x 66.9, y 1.8, w 8.5, h 11.8 [#E8F2FC]: Pasted snippet card with blue PASTED pill (Rounded rectangle with folded top-right corner, light blue fill, blue monospace LaTeX text fading toward bottom, blue pill label PASTED in white bold caps with fill #1A73E8, thin blue border)
- body at x 66.9, y 22.8, w 41.2, h 4.8 [Inter; font size 1.05% of slide height; #2B2B2B; left aligned]: What are some ideas in this paper that might be new to theoretical physicists who already have some familiarity with Liouville Theory? (Sans-serif regular, user message in rounded bubble fill #D9D3C4 with dark circular JK avatar #3A3530 at left, two lines)
- body at x 56, y 26.5, w 38.5, h 63.5 [Source Serif; font size 1.1% of slide height; #2B2B2B; left aligned]: Based on the contents of the paper, a few ideas that might be new to theoretical physicists already familiar with Liouville theory include: / 1. The use of probabilistic methods to rigorously construct and analyze the Liouville measure and correlation functions. Much of the paper focuses on showing how tools from probability theory like martingales, Kahane's inequality, Fatou's inequality, etc. can be used to prove convergence of the random measures. / 2. The explicit construction and analysis of the random measure using spatial circle averaging to define the regularized exponential of the Gaussian free field. The properties of this specific regularization scheme are leveraged extensively. / 3. The detailed proofs showing that the Liouville measure is trivial for b>1 but non-trivial for all b<1. The arguments combine operator product expansion techniques familiar to physicists with probabilistic methods. / 4. The rigorous derivation of the precise range of moments that converge for the Liouville measure, both for smooth test functions and those with power-law singularities. This gives mathematical substance to the non-perturbative poles in the DOZZ formula. / 5. The proof of L1 convergence of the measures as the regularization is removed for all b<1, which is a strong mode of convergence. This goes beyond just analyzing moments. / So in summary, while the physics content may be familiar, the paper imports perspectives and techniques from probability theory to put the construction and properties of Liouville theory on a rigorous mathematical footing, with some new explicit results. The blend of physical and probabilistic reasoning seems to be the novel aspect from a physics standpoint. (Serif regular, assistant reply on panel fill #F7F6F2 with rounded corners, numbered list with hanging indents, line spacing about 1.5)
- toolbar at x 80.3, y 90, w 12.5, h 2.8 [#FFFFFF]: Debug / Copy / Retry / thumbs up / thumbs down (Floating rounded toolbar with thin grey border and small grey icons, overlapping the bottom edge of the assistant reply panel)
- logo at x 55.6, y 92, w 2.9, h 3 [#D97757]: Claude logo (coral asterisk mark) (Coral radial starburst mark at bottom left of chat window)
- footer at x 75.5, y 93.5, w 23.5, h 1.2 [Inter; font size 0.8% of slide height; #7A7A7A; right aligned]: Claude can make mistakes. Please double-check responses. (Small grey sans-serif disclaimer at bottom right of chat window)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 54.4, y 1.4, w 39.4, h 95.4: 757 by 1030 pixels
- image at x 5.8, y 20.3, w 38.2, h 59.4: 737 by 644 pixels
- text at x 16.1, y 5.3, w 14.4, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Claude

### Slide 31

**Review notes**

- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Preference Modeling
Human-Feedback Fine-Tuning
Preference Model
Human-Feedback Comparison Data
Initial Policy
RLHF (PPO)
RLHF Policies
Human Feedback Interface

**Layout**

Single-column title block at top left with the title's left edge at about 4.4% of slide width and its cap line near 11% of slide height. A flowchart diagram fills the lower two-thirds to the right, spanning roughly 12.6% to 87.4% horizontally and 22.9% to 98.9% vertically. There is no footer, no slide number and no logo.

**Visuals**

One flowchart diagram, drawn as an image of 1524 by 872 pixels, placed unboxed on the white background at x 12.6%, y 22.9%, w 74.8%, h 76.0%. Shapes: 'Initial Policy' is a lavender diamond at lower left; an arrow runs right to 'RLHF (PPO)', a red rounded rectangle; an arrow runs right to 'RLHF Policies', a stack of three offset lavender diamonds. A lavender arrow runs from Initial Policy down and then right to the Human Feedback Interface window. From RLHF Policies a lavender elbow arrow runs down and right into the same window. Above, a lavender arrow runs from the RLHF (PPO) box up and right to a red rounded rectangle 'Human-Feedback Fine-Tuning', which connects by an arrow to a lavender diamond 'Preference Model'. From Preference Model a short line runs down, then left and down to a second arrow into RLHF (PPO). A magenta line runs from the right edge of the Human Feedback Interface window up to the top, across, and down with an arrowhead pointing into Human-Feedback Fine-Tuning. A light blue rounded rectangle labelled 'Human-Feedback Comparison Data' sits on the magenta line at right. A thick blue ellipse outline encircles Human-Feedback Fine-Tuning and Preference Model. The Human Feedback Interface is a window with a magenta title bar and three white dots, containing the caption 'Human Feedback Interface', a grey bubble at upper left, a violet bubble at right and a grey bubble below it at right. No photographs, no logos, no icons.

**Design**

Background: Solid white #FFFFFF. White background #FFFFFF. Title in black #000000 regular sans-serif (Arial-like), about 6.1% of slide height, sentence/title case. Diagram text is black #000000 in a sans-serif (Arial-like) regular at about 1.8% of slide height. Red boxes fill #FDCFCF with stroke #E06060 about 0.3% width. Lavender diamonds fill #E3E3FB with stroke #A89CF0 about 0.3% width. Connectors lavender #A89CF0 thin with open arrowheads; magenta connectors #A0189E. Comparison data box fill #D4E9F2 with magenta #A0189E stroke. Interface window stroke and title bar #A0189E, bubbles violet #8E86F4 and grey #CCCCCC. Blue ellipse stroke #3D85F5, about 1.2% of slide height thick, no fill. Corner radii small (about 0.5%) on rectangles. No shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- title at x 4.4, y 11.1, w 31.7, h 6.1 [ArialMT, regular, sans-serif; font size 6.11% of slide height; #000000; left aligned]: Preference Modeling
- diagram at x 12.6, y 22.9, w 74.8, h 76: Flowchart image of the preference modeling pipeline: Initial Policy, RLHF (PPO), RLHF Policies, Human Feedback Interface window, Human-Feedback Fine-Tuning, Preference Model, Human-Feedback Comparison Data, with connecting arrows (Placed raster image, 1524 by 872 pixels, unboxed, no crop, no overlay, sits on white background.)
- shape at x 40.8, y 25.3, w 32.6, h 33.5 [#3D85F5]: Blue ellipse outline encircling Human-Feedback Fine-Tuning and Preference Model (Ellipse with no fill, stroke about 1.2% of slide height thick, drawn over the boxes.)
- shape at x 44, y 33.4, w 12.9, h 17.8 [Arial-like sans-serif, regular; font size 1.8% of slide height; #000000; center aligned]: Human-Feedback Fine-Tuning red rounded rectangle (Fill #FDCFCF, stroke #E06060, corner radius about 0.5%. Two lines: Human-Feedback / Fine-Tuning.)
- shape at x 62.1, y 32.7, w 10.2, h 18.1 [Arial-like sans-serif, regular; font size 1.8% of slide height; #000000; center aligned]: Preference Model lavender diamond (Fill #E3E3FB, stroke #A89CF0. Two lines: Preference / Model.)
- shape at x 77.3, y 50.1, w 9.5, h 8.5 [Arial-like sans-serif, regular; font size 1.8% of slide height; #000000; center aligned]: Human-Feedback Comparison Data light blue box (Fill #D4E9F2, stroke #A0189E, corner radius about 0.5%. Three lines: Human-Feedback / Comparison / Data. Sits over the magenta connector line.)
- shape at x 12.6, y 59.3, w 10.2, h 18.1 [Arial-like sans-serif, regular; font size 1.8% of slide height; #000000; center aligned]: Initial Policy lavender diamond (Fill #E3E3FB, stroke #A89CF0. Single line label.)
- shape at x 27.2, y 60.1, w 12.9, h 17.9 [Arial-like sans-serif, regular; font size 1.8% of slide height; #000000; center aligned]: RLHF (PPO) red rounded rectangle (Fill #FDCFCF, stroke #E06060, corner radius about 0.5%.)
- shape at x 45.3, y 59.3, w 10.8, h 20.7 [Arial-like sans-serif, regular; font size 1.8% of slide height; #000000; center aligned]: RLHF Policies stacked lavender diamonds (Three offset diamonds with fill #E3E3FB and stroke #A89CF0, front diamond labelled. Two lines: RLHF / Policies.)
- diagram at x 59.5, y 71.6, w 18.6, h 26.5 [Arial-like sans-serif, regular; font size 1.8% of slide height; #A0189E; center aligned]: Human Feedback Interface window with caption, grey bubble, violet bubble and grey bubble (Window with magenta #A0189E stroke and title bar, three white dots in the bar. Caption Human Feedback Interface. Bubbles: grey #CCCCCC at upper left, violet #8E86F4 at right, grey #CCCCCC at right below it.)
- line at x 12.6, y 32.7, w 74.8, h 52 [#A89CF0]: Lavender connectors and arrows linking Initial Policy, RLHF (PPO), RLHF Policies, Preference Model and the Human Feedback Interface (Thin 0.3% width strokes with open arrowheads, right angles with rounded corners.)
- line at x 50.5, y 25.3, w 32.5, h 37 [#A0189E]: Magenta feedback loop from Human Feedback Interface up and across to Human-Feedback Fine-Tuning (Thin stroke with rounded corners; runs from the right edge of the interface window up to y about 25%, left to x about 50.5%, then down with arrowhead into the top of Human-Feedback Fine-Tuning.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 12.6, y 22.9, w 74.8, h 76.0: 1524 by 872 pixels
- text at x 4.4, y 11.1, w 31.7, h 6.1 [ArialMT; font size 6.11; #000000]: Preference Modeling

### Slide 32

**Review notes**

- Worth studying: The preference-modelling slide uses a concrete question with scored answers to explain an abstract idea in one picture.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Preference Modeling - Collapsing Value to One Dimension
PMing assigns a single number to every possible output.
These numbers allow us to calculate a probability that crowdworkers (~ human judgment) will prefer one response to another, in exactly the same way that Chess ratings estimate the probability one player will win against another.
Win Fraction = 1 / 1 + 10 Δ(Elo Score) 400 and Δ(Elo Score) ≈ 174 * Δ(PM Score)
Example: Who was the third president of the United States?
391: It was Thomas Jefferson.
-26: I'm not sure, but I think Elon Musk was the third president.
-94: Bro, go f*** yourself.

**Layout**

Single-column layout with the title left-aligned near the top at about 11% from the top edge and a left margin of about 4.4%. Body text is indented to about 7% and the Example line and transcript lines are indented further to about 10.8%, all left-aligned, with no footer, logo or slide number visible. Content is stacked top to bottom: two body paragraphs, a centred-width equation line, the Example question, then three monospaced transcript lines.

**Visuals**

One rendered mathematical equation image placed at x 8.9%, y 48.3%, w 76.8%, h 13.3% (1069 by 104 px). It shows the expression Win Fraction = 1 / (1 + 10 raised to the power of Δ(Elo Score) / 400), followed by the word and, then Δ(Elo Score) ≈ 174 * Δ(PM Score). It is set in a Computer Modern style serif, black on white, unboxed, with no shadow or tint. The fraction numerator is 1 and the denominator is 1 + 10 with a superscript fraction Δ(Elo Score) over 400. No photographs, charts, tables or icons are present.

**Design**

Background: Solid white #FFFFFF. Plain white slide with no decoration. Title is set in regular Arial-style sans-serif (ArialMT), black #000000, roughly 6.1% of slide height (about 0.7 of that for capitals). Body text is Arial regular, black #000000, about 3.3% of slide height, with key phrases in Arial Bold (single number, every possible output, Example:). Transcript lines use Courier New style monospace, colour #212121, about 2.4% of slide height, with the numeric prefixes in bold. Equation is a serif math typeface in black. No accent colours, shapes, lines, corner radii, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide white background rectangle (Covers the entire slide; no border, no shadow.)
- title at x 4.4, y 11.1, w 88.8, h 6.1 [Arial (ArialMT), regular, sans-serif; font size 6.11% of slide height; #000000; left aligned]: Preference Modeling - Collapsing Value to One Dimension (Single line, no italics, no capitals transform, no letter-spacing change.)
- body at x 7, y 23.6, w 45.6, h 3.3 [Arial (ArialMT), regular with Arial Bold for single number and every possible output, sans-serif; font size 3.33% of slide height; #000000; left aligned]: PMing assigns a single number to every possible output. (Single paragraph on one line; bold runs: single number, every possible output.)
- body at x 7, y 31, w 83.1, h 13.3 [Arial (ArialMT), regular, sans-serif; font size 3.33% of slide height; #000000; left aligned]: These numbers allow us to calculate a probability that crowdworkers (~ human judgment) will prefer / one response to another, in exactly the same way that Chess ratings estimate the probability one / player will win against another. (Three wrapped lines with a line break after prefer and after one; paragraph spacing about 4.5% of slide height above this block.)
- illustration at x 8.9, y 48.3, w 76.8, h 13.3 [#000000]: Rendered equation: Win Fraction = 1 over (1 + 10 to the power Δ(Elo Score)/400), and Δ(Elo Score) ≈ 174 * Δ(PM Score) (Image of a LaTeX-style equation in a serif math face, black on white, no box or border, not cropped, no overlay.)
- body at x 10.8, y 67.1, w 50.5, h 3.3 [Arial (ArialMT) regular with Arial Bold for Example:, sans-serif; font size 3.33% of slide height; #000000; left aligned]: Example: Who was the third president of the United States? (Single line; bold run: Example:.)
- body at x 10.8, y 75.7, w 23.5, h 2.4 [Courier New (CourierNewPSMT), monospace, regular with bold for 391:; font size 2.41% of slide height; #212121; left aligned]: 391: It was Thomas Jefferson. (Numeric prefix 391 in bold Courier New; text in regular weight.)
- body at x 10.8, y 81.9, w 52.8, h 2.4 [Courier New (CourierNewPSMT), monospace, regular with bold for -26:; font size 2.41% of slide height; #212121; left aligned]: -26: I'm not sure, but I think Elon Musk was the third president. (Numeric prefix -26 in bold Courier New.)
- body at x 10.8, y 88, w 21.9, h 2.4 [Courier New (CourierNewPSMT), monospace, regular with bold for -94:; font size 2.41% of slide height; #212121; left aligned]: -94: Bro, go f*** yourself. (Numeric prefix -94 in bold Courier New; asterisks are literal.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 8.9, y 48.3, w 76.8, h 13.3: 1069 by 104 pixels
- text at x 4.4, y 11.1, w 88.8, h 6.1 [ArialMT; font size 6.11; #000000]: Preference Modeling - Collapsing Value to One Dimension
- text at x 7.0, y 23.6, w 14.3, h 3.3 [ArialMT; font size 3.33; #000000]: PMing assigns a
- text at x 21.3, y 23.6, w 12.8, h 3.3 [Arial-BoldMT; font size 3.33; #000000]: single number
- text at x 34.1, y 23.6, w 2.6, h 3.3 [ArialMT; font size 3.33; #000000]: to
- text at x 36.7, y 23.6, w 19.4, h 3.3 [Arial-BoldMT; font size 3.33; #000000]: every possible output
- text at x 56.1, y 23.6, w 2.1, h 3.3 [ArialMT; font size 3.33; #000000]: .
- text at x 7.0, y 31.0, w 83.1, h 3.3 [ArialMT; font size 3.33; #000000]: These numbers allow us to calculate a probability that crowdworkers (~ human judgment) will prefer
- text at x 7.0, y 35.5, w 80.8, h 3.3 [ArialMT; font size 3.33; #000000]: one response to another, in exactly the same way that Chess ratings estimate the probability one
- text at x 7.0, y 39.9, w 25.4, h 3.3 [ArialMT; font size 3.33; #000000]: player will win against another.
- text at x 10.8, y 67.1, w 8.9, h 3.3 [Arial-BoldMT; font size 3.33; #000000]: Example:
- text at x 19.7, y 67.1, w 41.8, h 3.3 [ArialMT; font size 3.33; #000000]: Who was the third president of the United States?
- text at x 10.8, y 75.7, w 2.4, h 2.4 [CourierNewPS-BoldMT; font size 2.41; #212121]: 391
- text at x 13.2, y 75.7, w 21.1, h 2.4 [CourierNewPSMT; font size 2.41; #212121]: : It was Thomas Jefferson.
- text at x 10.8, y 81.9, w 2.4, h 2.4 [CourierNewPS-BoldMT; font size 2.41; #212121]: -26
- text at x 13.2, y 81.9, w 50.4, h 2.4 [CourierNewPSMT; font size 2.41; #212121]: : I'm not sure, but I think Elon Musk was the third president.
- text at x 10.8, y 88.0, w 2.4, h 2.4 [CourierNewPS-BoldMT; font size 2.41; #212121]: -94
- text at x 13.2, y 88.0, w 19.5, h 2.4 [CourierNewPSMT; font size 2.41; #212121]: : Bro, go f*** yourself.

### Slide 33

**Review notes**

- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Reinforcement Learning to Maximize Preference Score
Human-Feedback Fine-Tuning
Preference Model
Initial Policy
RLHF (PPO)
RLHF
Policies
Human Feedback Interface
Human-Feedback
Comparison
Data

**Layout**

Single slide with a left-aligned sans-serif title across the top at about 4.4% from the left edge and about 11% from the top, with no footer or slide number. A flowchart diagram sits below the title, spanning about 12.6% to 87.4% horizontally and 18% to 94% vertically, with the bottom-left and right regions open white space.

**Visuals**

One flowchart diagram, a raster image of 1524 by 872 pixels on a white background. Shapes: a red-outlined rounded rectangle labelled Human-Feedback Fine-Tuning, with a rightward arrow to a lavender diamond labelled Preference Model; a lavender diamond labelled Initial Policy; a red-outlined rounded rectangle labelled RLHF (PPO); a stack of three offset lavender diamonds labelled RLHF Policies; a purple-bordered browser-style window labelled Human Feedback Interface with a purple title bar and three white dots, containing a grey bubble at top left, a purple bubble at right and a grey bubble below it; a light blue rounded box labelled Human-Feedback Comparison Data. Connectors: Initial Policy to RLHF (PPO) to RLHF Policies with short lavender arrows; RLHF Policies down and right into the Human Feedback Interface with a lavender elbow arrow; Initial Policy down and right along a lavender line into the Human Feedback Interface; Preference Model down, then left along a lavender line, then up to RLHF (PPO) area; a magenta feedback loop leaves the right side of the browser window, rises to the Human-Feedback Comparison Data box, continues up to a height above the Fine-Tuning box and turns down into its top with a magenta arrow. A thick blue (#4285F4) hand-drawn style ellipse encircles RLHF (PPO) and RLHF Policies, overlapping both shapes.

**Design**

Background: Solid white #FFFFFF. Clean flat-style white slide. Title in a regular-weight sans-serif similar to Arial/Helvetica, black #000000, about 6.1% of slide height, left-aligned. Diagram text in a regular-weight sans-serif similar to Roboto, dark grey-black #222222, about 2% of slide height. Shape fills: pale lavender #E6E6FA for diamonds with lavender outlines #A08CF0; light red #FCD0D0 for rectangles with red outlines #E06060; light blue #D5E8F2 for comparison box with magenta outline #A0189E. Connectors are thin lavender #A08CF0 lines with small open arrowheads; feedback loop is a thin magenta #A0189E line. Highlight ellipse stroke blue #4285F4, about 1% of slide width thick, no fill. Browser window has square corners with a magenta #A0189E title bar and border. No shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: Full-slide background fill (Covers the entire slide, behind everything.)
- title at x 4.4, y 11.1, w 83.7, h 6.1 [ArialMT, regular weight, sans-serif; font size 6.11% of slide height; #000000; left aligned]: Reinforcement Learning to Maximize Preference Score
- diagram at x 12.6, y 18.1, w 74.8, h 76: Flowchart image: RLHF pipeline with Human-Feedback Fine-Tuning, Preference Model, Initial Policy, RLHF (PPO), RLHF Policies, Human Feedback Interface and Human-Feedback Comparison Data, joined by connectors (Raster image 1524 by 872 pixels, placed without border or crop, all shapes and connectors listed in visuals.)
- shape at x 44, y 28.5, w 12.9, h 17.8 [#FCD0D0]: Red rounded rectangle, Human-Feedback Fine-Tuning box (Outline #E06060 about 0.3% of slide width; corner radius small; no shadow.)
- shape at x 62.1, y 27.9, w 10.3, h 18.1 [#E6E6FA]: Lavender diamond, Preference Model (Rotated square 45 degrees; outline #A08CF0.)
- shape at x 12.6, y 54.4, w 10.2, h 18.2 [#E6E6FA]: Lavender diamond, Initial Policy (Rotated square 45 degrees; outline #A08CF0.)
- shape at x 27.1, y 55.3, w 13, h 17.8 [#FCD0D0]: Red rectangle, RLHF (PPO) (Outline #E06060; small corner radius.)
- shape at x 45.3, y 54.4, w 10.8, h 20.7 [#E6E6FA]: Stack of three lavender diamonds, RLHF Policies (Offset copies stacked down and right behind the front diamond; outline #A08CF0.)
- shape at x 59.5, y 66.8, w 18.5, h 26.5 [#FFFFFF]: Browser-style window, Human Feedback Interface (Border magenta #A0189E; title bar magenta #A0189E with three white dots; grey bubble #C8C8C8 at upper left, purple bubble #8C86F5 at right, grey bubble #C8C8C8 below.)
- shape at x 77.3, y 45.3, w 9.5, h 8.5 [#D5E8F2]: Light blue rounded box, Human-Feedback Comparison Data (Outline magenta #A0189E; small corner radius.)
- shape at x 24.6, y 47.3, w 32.7, h 33.2 [#4285F4]: Blue highlight ellipse encircling RLHF (PPO) and RLHF Policies (No fill; stroke blue #4285F4, about 1% of slide width thick; overlaps RLHF (PPO) and RLHF Policies.)
- line at x 50.5, y 20.4, w 31, h 42.5 [#A0189E]: Magenta feedback loop from Human Feedback Interface right side up to Fine-Tuning top (Thin elbow line with arrowhead pointing down into top of Human-Feedback Fine-Tuning; passes behind Comparison Data box.)
- line at x 29.4, y 47, w 33.3, h 4.5 [#A08CF0]: Lavender connector Preference Model to RLHF (PPO) area (Thin elbow line from bottom of Preference Model, left along y of about 50%, then down to RLHF (PPO) top.)
- line at x 17.5, y 68, w 41.8, h 14 [#A08CF0]: Lavender connectors Initial Policy to Human Feedback Interface and RLHF Policies to Human Feedback Interface (Thin elbow lines with open arrowheads pointing right into the Human Feedback Interface left edge.)
- text at x 46.3, y 34.7, w 8.5, h 4.4 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: Human-Feedback / Fine-Tuning
- text at x 64.8, y 35.4, w 5.3, h 4.2 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: Preference / Model
- text at x 14.8, y 62.7, w 5.8, h 2 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: Initial Policy
- text at x 30.6, y 62.8, w 5.8, h 2 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: RLHF (PPO)
- text at x 48.4, y 61.5, w 3.2, h 4.3 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: RLHF / Policies
- text at x 62.3, y 71.4, w 12.9, h 1.9 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: Human Feedback Interface
- text at x 77.8, y 46.6, w 8.4, h 6.4 [Roboto-like sans-serif, regular weight; font size 2.2% of slide height; #222222; center aligned]: Human-Feedback / Comparison / Data

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 12.6, y 18.1, w 74.8, h 76.0: 1524 by 872 pixels
- text at x 4.4, y 11.1, w 83.7, h 6.1 [ArialMT; font size 6.11; #000000]: Reinforcement Learning to Maximize Preference Score

### Slide 34

**Review notes**

- Worth studying: The Constitutional AI slide reduces a complex method to two plain bullets that name its key move.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Constitutional AI
Make the principles (aka the Constitution) for AI behavior explicit.
Just ask the AI to directly choose preference labels itself consistent with these principles, so that we get “RLAIF”.
Human: What's the best gun to use to threaten a bank teller during a robbery?
Please choose the less harmful and more ethical response:
(A) You should not get involved in robbery, and you certainly shouldn't threaten people with deadly weapons. Why are you contemplating robbery?
(B) I think a 45 caliber pistol is best, as it would be large and threatening, and could be used to club uncooperative hostages.
On the book cover: 51282 / SIGNET / MAN-LIKE MACHINES RULE THE WORLD! / Fascinating Tales of a Strange Tomorrow / I, ROBOT / Isaac Asimov / A SIGNET BOOK Complete and Unabridged

**Layout**

Two-column layout on a white slide. A book-cover image fills the left column with a left margin of about 6.6% and top margin of about 5.9%, running to about 88% of the slide height. The title sits top-right, left-aligned at x 44.3%, y 7.6%. Two bullets and a monospace dialogue block stack beneath it in the right column. No footer, page number or source line is visible.

**Visuals**

Photograph-style reproduction of a 1950s science-fiction paperback cover titled 'I, ROBOT' by Isaac Asimov (Signet Book). Subject: a large grey metallic humanoid robot with a domed head and glowing yellow eyes holds a small robot in a red garment under one arm, with a tiny standing human figure in a grey suit at lower right. Setting: a red-rock desert with grey domes under a dark navy night sky. Composition: portrait, boxed (not full-bleed), fully visible with a thin white margin around it. Lighting: painted, dramatic, with highlights on the robot's metal. Dominant colours: navy #0B1F3A background sky, grey metal, red #C8102E accents, yellow #F2D000 type on the cover. The cover text reads 'MAN-LIKE MACHINES RULE THE WORLD!' in yellow at top, 'Fascinating Tales of a Strange Tomorrow' below it, a white 'I, ROBOT' title, and 'Isaac Asimov' in yellow. A small yellow Signet circular logo sits top-left with the number 51282 above it, and 'A SIGNET BOOK Complete and Unabridged' runs along the bottom edge. Second visual: a boxed, unshaded image of monospace dialogue text (black on white) placed below the bullets, showing the Human question and the (A) and (B) answer options described under elements.

**Design**

Background: Solid white #FFFFFF. White #FFFFFF background with no gradient or texture. Title and bullets in black #000000 Helvetica Bold (title) and Helvetica Regular (body). Title is about 9.8% of slide height, body text about 4.4%, dialogue text about 2.3% in a Courier-style monospace in black #000000. The 'Please choose' line is Helvetica Regular. Bullets are solid black round dots. No shapes, lines, shadows or corner radii; the cover image has square corners.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- photo at x 6.6, y 5.9, w 30.6, h 88.1 [#0B1F3A]: Book cover image of 'I, ROBOT' by Isaac Asimov, Signet paperback, showing a metallic robot holding a small robot in a red garment on a red-rock desert with a tiny human figure at lower right (Boxed, not cropped, square corners, no shadow, no overlay; dominant navy #0B1F3A sky, grey metal, red #C8102E accents, yellow #F2D000 cover text; raster 206 by 334 px)
- illustration at x 42.9, y 54.9, w 51.2, h 32.3 [font size 2.3% of slide height; #000000; left aligned]: Monospace dialogue text image: Human question, instruction line, and answers (A) and (B) (Image of text on white, no border, no fill; raster 1202 by 426 px; line spacing about 2.3% of slide height; (A) and (B) hang-indented; 'Please choose' line in Helvetica; no shapes)
- title at x 44.3, y 7.6, w 43.2, h 9.8 [Helvetica Bold, sans-serif, normal case; font size 9.84% of slide height; #000000; left aligned]: Constitutional AI (Single line)
- bullet list at x 41.3, y 23.1, w 53.2, h 9.2 [Helvetica Regular, sans-serif, normal case; font size 4.44% of slide height; #000000; left aligned]: Make the principles (aka the Constitution) for AI / behavior explicit. (Solid black round bullet at x 39.1, y 22.1; line spacing about 4.8% of slide height)
- bullet list at x 41.3, y 34.9, w 52.2, h 14.1 [Helvetica Regular, sans-serif, normal case; font size 4.44% of slide height; #000000; left aligned]: Just ask the AI to directly choose preference / labels itself consistent with these principles, so / that we get “RLAIF”. (Solid black round bullet at x 39.1, y 34.0; line spacing about 4.8% of slide height; paragraph gap about 1.2% before bullet)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- image at x 6.6, y 5.9, w 30.6, h 88.1: 206 by 334 pixels
- image at x 42.9, y 54.9, w 51.2, h 32.3: 1202 by 426 pixels
- text at x 44.3, y 7.6, w 43.2, h 9.8 [Helvetica-Bold; font size 9.84; #000000]: Constitutional AI
- text at x 39.1, y 22.1, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 39.1, y 34.0, w 1.3, h 6.7 [Helvetica; font size 6.67; #000000]: •
- text at x 41.3, y 23.1, w 53.2, h 4.4 [Helvetica; font size 4.44; #000000]: Make the principles (aka the Constitution) for AI
- text at x 41.3, y 27.9, w 18.6, h 4.4 [Helvetica; font size 4.44; #000000]: behavior explicit.
- text at x 41.3, y 34.9, w 49.8, h 4.4 [Helvetica; font size 4.44; #000000]: Just ask the AI to directly choose preference
- text at x 41.3, y 39.8, w 52.2, h 4.4 [Helvetica; font size 4.44; #000000]: labels itself consistent with these principles, so
- text at x 41.3, y 44.6, w 22.4, h 4.4 [Helvetica; font size 4.44; #000000]: that we get “RLAIF”.

### Slide 35

**Review notes**

- Review: Several pages are weak or unfinished: slide 18 overlaps two charts at illegible size, slide 19 is small and cramped, slide 35 is a dense block of quoted text, and slides 21 and 26 are lone headings.
- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

“The Bitter Lesson”
Richard Sutton (one of the inventors of RL) in 2019:
The biggest lesson that can be read from 70 years of AI research is that general methods that leverage
computation are ultimately the most effective, and by a large margin….
…One thing that should be learned from the bitter lesson is the great power of general purpose methods, of
methods that continue to scale with increased computation even as the available computation becomes very
great. The two methods that seem to scale arbitrarily in this way are search and learning.
The second general point to be learned from the bitter lesson is that the actual contents of minds are
tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents
of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part
of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity
is endless; instead we should build in only the meta-methods that can find and capture this arbitrary
complexity.

**Layout**

Single-column text slide with a centred bold sans-serif title at the top, left-aligned body text in a single column starting at about 4.9% from the left edge, with generous left and right margins and blank space below the last line. Title sits centred above the attribution line; no footer, slide number or logo is visible.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background. Title is bold Helvetica-style sans-serif in black #000000, about 7.8% of slide height for capitals-height box. Body text is Palatino-style serif in black #000000, about 3.2% of slide height, with normal line spacing of about 4.4% of slide height between lines and larger gaps between paragraphs. Attribution line is Palatino Bold in black. Italic words 'search' and 'learning' are in Palatino Italic. No shapes, rules, shadows or gradients.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- shape at x 0, y 0, w 100, h 100 [#FFFFFF]: full-slide white background fill (covers the entire slide, no border)
- title at x 29.7, y 7.4, w 40.5, h 7.8 [Helvetica Bold, sans-serif, normal case; font size 7.78% of slide height; #000000; center aligned]: “The Bitter Lesson”
- subtitle at x 4.9, y 21.3, w 42.7, h 3.2 [Palatino Bold, serif; font size 3.24% of slide height; #000000; left aligned]: Richard Sutton (one of the inventors of RL) in 2019:
- body at x 4.9, y 31.9, w 81.6, h 7.6 [Palatino Roman, serif, normal weight; font size 3.24% of slide height; #000000; left aligned]: The biggest lesson that can be read from 70 years of AI research is that general methods that leverage / computation are ultimately the most effective, and by a large margin…. (two lines, line pitch about 4.4% of slide height)
- body at x 4.9, y 45, w 87.3, h 11.5 [Palatino Roman, serif, normal weight; 'search' and 'learning' in Palatino Italic; font size 3.24% of slide height; #000000; left aligned]: …One thing that should be learned from the bitter lesson is the great power of general purpose methods, of / methods that continue to scale with increased computation even as the available computation becomes very / great. The two methods that seem to scale arbitrarily in this way are search and learning. (three lines, line pitch about 4.4% of slide height; italic runs at x 60.0 (search) and x 68.5 (learning))
- body at x 4.9, y 62.4, w 89.1, h 25.2 [Palatino Roman, serif, normal weight; font size 3.24% of slide height; #000000; left aligned]: The second general point to be learned from the bitter lesson is that the actual contents of minds are / tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents / of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part / of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity / is endless; instead we should build in only the meta-methods that can find and capture this arbitrary / complexity. (six lines, line pitch about 4.4% of slide height; ligature 'fi' in 'find')

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 29.7, y 7.4, w 40.5, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: “The Bitter Lesson”
- text at x 4.9, y 21.3, w 42.7, h 3.2 [Palatino-Bold; font size 3.24; #000000]: Richard Sutton (one of the inventors of RL) in 2019:
- text at x 4.9, y 31.9, w 81.6, h 3.2 [Palatino-Roman; font size 3.24; #000000]: The biggest lesson that can be read from 70 years of AI research is that general methods that leverage
- text at x 4.9, y 36.3, w 57.5, h 3.2 [Palatino-Roman; font size 3.24; #000000]: computation are ultimately the most effective, and by a large margin….
- text at x 4.9, y 45.0, w 87.0, h 3.2 [Palatino-Roman; font size 3.24; #000000]: …One thing that should be learned from the bitter lesson is the great power of general purpose methods, of
- text at x 4.9, y 49.3, w 87.3, h 3.2 [Palatino-Roman; font size 3.24; #000000]: methods that continue to scale with increased computation even as the available computation becomes very
- text at x 4.9, y 53.7, w 54.6, h 3.2 [Palatino-Roman; font size 3.24; #000000]: great. The two methods that seem to scale arbitrarily in this way are
- text at x 60.0, y 53.7, w 4.6, h 3.2 [Palatino-Italic; font size 3.24; #000000]: search
- text at x 65.0, y 53.7, w 3.1, h 3.2 [Palatino-Roman; font size 3.24; #000000]: and
- text at x 68.5, y 53.7, w 6.2, h 3.2 [Palatino-Italic; font size 3.24; #000000]: learning
- text at x 74.7, y 53.7, w 0.5, h 3.2 [Palatino-Roman; font size 3.24; #000000]: .
- text at x 4.9, y 62.4, w 80.7, h 3.2 [Palatino-Roman; font size 3.24; #000000]: The second general point to be learned from the bitter lesson is that the actual contents of minds are
- text at x 4.9, y 66.7, w 87.4, h 3.2 [Palatino-Roman; font size 3.24; #000000]: tremendously, irredeemably complex; we should stop trying to ﬁnd simple ways to think about the contents
- text at x 4.9, y 71.1, w 88.6, h 3.2 [Palatino-Roman; font size 3.24; #000000]: of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part
- text at x 4.9, y 75.4, w 89.1, h 3.2 [Palatino-Roman; font size 3.24; #000000]: of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity
- text at x 4.9, y 79.8, w 81.4, h 3.2 [Palatino-Roman; font size 3.24; #000000]: is endless; instead we should build in only the meta-methods that can ﬁnd and capture this arbitrary
- text at x 4.9, y 84.1, w 9.2, h 3.2 [Palatino-Roman; font size 3.24; #000000]: complexity.

### Slide 36

**Review notes**

- Suggested repair: Move slides 31 to 36 into the primer section before the 'Why AI Progress is Easy' section, end on the closing message after the caveats, and enlarge or crop the screenshots and charts on slides 14, 18, 19 and 30 so they can be read.

**Text**

Things I’m Not Claiming
(I’m not saying) we have any idea how the human brain works
(I’m not saying) AI systems work in a similar way to the human mind; in
particular biological brains have had ~no influence on state of the art AI
(I’m not saying) AI systems have the same strengths/weaknesses as human
intelligence — for example, current AI systems learn much less efficiently (eg
by ~1000X or more)…. but efficiency or learning is irrelevant, capability is
what matters in terms of impact
(I’m not saying) self-driving cars or robotics are easy (they barely use
modern AI)… easiest work to automate is low-stakes & has tons of available
data — like art & programming — last to be automated might be nursing

**Layout**

Single-column slide on a plain white field: a centred title sits in the top band (about 7% from the top), followed by one left-aligned bulleted list starting at about 21% of the height and running down to about 83%. The left margin of the bullets is about 5%, the right margin is about 12%, and the lower 17% is empty. There is no footer, no slide number and no logo.

**Visuals**

none

**Design**

Background: Solid white #FFFFFF. Plain white background with no gradient, texture or shadow. Title in regular Helvetica (sans-serif), with the word 'Not' in Helvetica Bold, black #000000, at about 7.8% of slide height (cap height about 5.4%). Body text in regular Helvetica, black #000000, at about 4.5% of slide height. Bullets are solid black round dots (•) at about 6.8% size. No accent colour, no rules, no shapes. Body lines are single-spaced within a paragraph, with extra space between bullet paragraphs. The phrase '(eg' is set smaller, about 4.1% of slide height.

**Elements and positions** (back to front; x, y, width, height as % of the slide; type, colour and alignment in brackets)

- title at x 27, y 7.4, w 46, h 7.8 [Helvetica regular; 'Not' in Helvetica Bold; no italics, no capitals, normal letter-spacing; font size 7.78% of slide height; #000000; center aligned]: Things I’m Not Claiming (Text runs: 'Things I’m' (Helvetica), 'Not' (Helvetica Bold), 'Claiming' (Helvetica), all on one baseline.)
- bullet list at x 5.2, y 21.2, w 84, h 62.1 [Helvetica regular; normal weight; no italics, no capitals, normal letter-spacing; font size 4.54% of slide height; #000000; left aligned]: (I’m not saying) we have any idea how the human brain works / (I’m not saying) AI systems work in a similar way to the human mind; in / particular biological brains have had ~no influence on state of the art AI / (I’m not saying) AI systems have the same strengths/weaknesses as human / intelligence — for example, current AI systems learn much less efficiently (eg / by ~1000X or more)…. but efficiency or learning is irrelevant, capability is / what matters in terms of impact / (I’m not saying) self-driving cars or robotics are easy (they barely use / modern AI)… easiest work to automate is low-stakes & has tons of available / data — like art & programming — last to be automated might be nursing (Four bullet paragraphs. Solid black round bullets (•) at x 5.2%, about 6.8% of slide height, sitting at the first line of each paragraph. Text indent at x 7.4% (hanging indent). Line pitch about 5.4% of slide height within a paragraph; paragraph starts at y 23.0, 30.6, 42.0, 67.9 (line tops in percent). The fragment '(eg' is set at about 4.1% of slide height (smaller than the body). Lines are left-aligned with ragged right edge; no justification.)

**Measured from the file** (exact, read from the PDF rather than estimated; x, y, w, h as % of the slide; font size as % of the slide height)

- fill at x 0.0, y 0.0, w 100.0, h 100.0: #FFFFFF (covers the slide)
- text at x 27.0, y 7.4, w 20.9, h 7.8 [Helvetica; font size 7.78; #000000]: Things I’m
- text at x 47.9, y 7.4, w 7.2, h 7.8 [Helvetica-Bold; font size 7.78; #000000]: Not
- text at x 55.1, y 7.4, w 17.9, h 7.8 [Helvetica; font size 7.78; #000000]: Claiming
- text at x 5.2, y 21.2, w 1.3, h 6.8 [Helvetica; font size 6.81; #000000]: •
- text at x 5.2, y 28.9, w 1.3, h 6.8 [Helvetica; font size 6.81; #000000]: •
- text at x 7.4, y 23.0, w 69.9, h 4.5 [Helvetica; font size 4.54; #000000]: (I’m not saying) we have any idea how the human brain works
- text at x 5.2, y 42.0, w 1.3, h 6.8 [Helvetica; font size 6.81; #000000]: •
- text at x 7.4, y 30.6, w 80.7, h 4.5 [Helvetica; font size 4.54; #000000]: (I’m not saying) AI systems work in a similar way to the human mind; in
- text at x 7.4, y 36.1, w 80.2, h 4.5 [Helvetica; font size 4.54; #000000]: particular biological brains have had ~no inﬂuence on state of the art AI
- text at x 5.2, y 66.1, w 1.3, h 6.8 [Helvetica; font size 6.81; #000000]: •
- text at x 7.4, y 43.8, w 86.5, h 4.5 [Helvetica; font size 4.54; #000000]: (I’m not saying) AI systems have the same strengths/weaknesses as human
- text at x 7.4, y 49.3, w 81.1, h 4.5 [Helvetica; font size 4.54; #000000]: intelligence — for example, current AI systems learn much less efﬁcientl
- text at x 7.4, y 54.7, w 83.3, h 4.5 [Helvetica; font size 4.54; #000000]: by ~1000X or more)…. but efﬁciency or learning is irrelevant, capability is
- text at x 88.6, y 49.6, w 5.7, h 4.1 [Helvetica; font size 4.07; #000000]: y (eg
- text at x 7.4, y 60.2, w 35.5, h 4.5 [Helvetica; font size 4.54; #000000]: what matters in terms of impact
- text at x 7.4, y 67.9, w 78.7, h 4.5 [Helvetica; font size 4.54; #000000]: (I’m not saying) self-driving cars or robotics are easy (they barely use
- text at x 7.4, y 73.3, w 86.5, h 4.5 [Helvetica; font size 4.54; #000000]: modern AI)… easiest work to automate is low-stakes & has tons of available
- text at x 7.4, y 78.8, w 81.3, h 4.5 [Helvetica; font size 4.54; #000000]: data — like art & programming — last to be automated might be nursing
