# Anthropic conference talk, 2023: deck review (machine-readable)

- Company: Anthropic
- Type: Conference talk
- Year: 2023
- Industry: Artificial intelligence
- Slides: 13
- Slide size: 16:9 (720 x 405 points)
- Review score: 6 out of 10
- Entry: https://slidedeck.co/library/anthropic-conference-2023/
- Original: https://aiconference.com/wp-content/uploads/2023/10/Ben-Mann-Anthropic-Slides-1-1.pdf
- Recording: https://www.youtube.com/watch?v=GWXfpFchkMQ
- 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 deck has a coherent visual identity (warm earth palette, hand-drawn illustrations, one friendly typeface) and a clear route: mission, agenda, then three sections marked by a highlighted agenda slide (8, 10). But the content slides are three-word bullet fragments such as 'Define principles' and 'Crowdsourcing' (7, 9), so without the speaker the audience cannot learn how Constitutional AI or oversight actually works. Slide 12, the only information-bearing graphic (ASL levels), is the weakest legible page: tiny captions in a flat four-box row with an arrow, and it appears after the deployment slide with no framing and no link to the agenda.

**Worth keeping:** Keep the earth-tone palette, the consistent illustration style, and the highlighted-agenda device (slides 5, 8, 10) that tells the audience where they are.

**Most useful repair:** Replace fragment bullets on slides 7, 9 and 11 with one concrete sentence or example each, enlarge the ASL captions on slide 12 and tie that slide to the deployment section.

**What would change the score:** Cut the duplicate agenda on slide 4 versus 5, and fix the pixelated illustration on slide 10 and the loose title and bullet alignment, with the title indented off the bullet marker on slides 7 and 9.

_Confidence medium. 13 of 13 pages inspected on contact sheets, 4 at readable size. Reviewed Reviewed 2026-11-08. Model-assisted review (claude-sonnet-5-5)._

## Summary

The deck is a conference talk by Ben Mann, cofounder of Anthropic, titled Aligning AI with Human Values, with lessons from building Claude. It opens with a title slide and a mission statement: developing AI that is helpful, honest, and harmless. A short agenda follows, repeated as a recurring section marker that lists alignment, scalable oversight, and responsible deployment. The alignment section defines the three mission terms with brief sub-points and then introduces Constitutional AI as a sequence of defining principles, revising responses accordingly, and applying reinforcement learning. The scalable oversight section lists automated testing, expert red teaming, and crowdsourcing as methods. The deployment section covers strong demand at launch, data privacy, and a layered response to new attacks. The deck closes with a single line, Let's build safe AGI. Most slides are short bullet lists with headline-style titles, so the agenda slides serve as a navigational spine. Slides 4 and 5 repeat the same agenda text, and slide 12 is blank.

## Questions about this deck

**What does the Anthropic deck say about its mission?**

Slide 3 states the mission as developing AI that is helpful, honest, and harmless. Slide 6 lists sub-points under each term, such as concise, efficient, and appropriate for helpful.

**How many slides are in the Anthropic deck and how is it structured?**

The Anthropic deck has 13 slides. Agenda slides 4, 5, 8, and 10 act as section markers for alignment, scalable oversight, and responsible deployment, and slide 13 closes with the line Let's build safe AGI.

**What methods does the Anthropic deck list for scalable oversight?**

Slide 9 lists automated testing, expert red teaming, and crowdsourcing as scalable oversight methods.

**What does the Anthropic deck say about deployment?**

Slide 11 covers strong demand at launch, data privacy, and a layered response to new attacks.

## 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**

Ben Mann, cofounder
ANTHROPIC

**Layout**

Single centred column with all three elements stacked on the vertical axis of the slide: the hands illustration at top, the ANTHROPIC wordmark beneath it, and the name line beneath that. The stack sits in the middle of the slide with no title, footer, or slide number.

**Visuals**

Two non-text graphics. (1) Illustration of about a dozen raised hands drawn in black, hand-drawn, thin uneven outline strokes with no fill, arranged in a tight irregular cluster of rows, fingers pointing up, wider at the top and narrower toward the bottom edge; the group is centred horizontally, unboxed and uncropped, on the flat background. (2) Logo wordmark 'ANTHROP\C' in heavy, bold, geometric sans-serif capitals in #212121, with the letter O in a round form and the C followed by a backslash-like glyph; unboxed, no overlay, no tint.

**Design**

Background: Flat solid fill #E6E4E0 covering the full slide. Flat light warm-grey background with no gradient or texture. Wordmark and name line are both #212121 near-black. The hands illustration is black. No accent colour. Name line is set in Lexend Medium, a geometric sans-serif, regular-to-medium weight, sentence case, about 8.9% of slide height. Wordmark is a heavy bold geometric sans in capitals. No shapes, shadows, corner radii 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 [#E6E4E0]: Full-slide background fill (Solid rectangle covering the entire slide, no border, no shadow.)
- illustration at x 41, y 21.6, w 18.1, h 32.1 [#000000]: Cluster of about a dozen black hand-drawn raised hands, fingers up, forming a rounded crowd shape (Black thin hand-drawn outline strokes on transparent background, no fill, centred horizontally, unboxed, not cropped.)
- logo at x 25.7, y 52.5, w 48.7, h 9.7 [#212121]: ANTHROP\C wordmark (Heavy bold geometric sans-serif capitals in #212121, unboxed, centred, placed directly under the hands illustration with a small gap.)
- body at x 23.8, y 63.7, w 52.5, h 8.9 [Lexend Medium, sans-serif, medium weight, normal style, no capitals, no letter-spacing; font size 8.89% of slide height; #212121; center aligned]: Ben Mann, cofounder (Single line, sits directly below the wordmark, centred.)

**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: #E6E4E0 (covers the slide)
- image at x 41.0, y 21.6, w 18.1, h 32.1: 256 by 256 pixels
- image at x 25.7, y 52.5, w 48.7, h 9.7: 1024 by 115 pixels
- text at x 23.8, y 63.7, w 52.5, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Ben Mann, cofounder

### Slide 2

**Text**

Aligning AI with Human Values
Lessons from Building Claude

**Layout**

Single centred column on a full-bleed terracotta field. Title block sits in the vertical middle-upper area with the illustration centred above it; text is horizontally centred with no footer, no slide number and no visible margins beyond the centred column. The title is centred at about 49.5% of slide height and the subtitle follows below it.

**Visuals**

One line illustration centred at the top third: a hand-drawn black outline of a head in profile made of several concentric arcs (about 5 nested curves) that open downward and end in wavy lines suggesting hair and face contour, overlaid with thin off-white/pale pink (#F2E3DC) radiating spokes from a central node, with seven small pale pink circular nodes at the spoke ends (top, upper-left, upper-right, right, left, lower-left, lower-center). The illustration is unboxed, sits on the background with no shadow, and occupies roughly x 42 to 57 percent of slide width and y 14 to 39 percent of slide height. No photograph, chart or logo.

**Design**

Background: Solid terracotta #CB785C, full-bleed across the entire slide.

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

- illustration at x 42, y 14, w 15, h 26 [#F2E3DC]: Line illustration of a head in profile formed by nested black hand-drawn arcs with pale pink radiating spokes and seven pale pink nodes at spoke ends (Black strokes approx 0.4 percent of slide height thick with slightly irregular hand-drawn quality; spokes approx 0.3 percent thick in off-white #F2E3DC; nodes are filled pale pink circles about 1.5 percent of slide width; image placed as 512 by 288 pixel bitmap at x 32.1, y 10.5, w 35.9, h 35.9 with the illustration occupying its centre; no crop, no border, no shadow)
- title at x 15.4, y 49.5, w 69.3, h 8 [Lexend, Medium weight, sans-serif, no italics, no capitals, normal letter-spacing; font size 8% of slide height; #212121; center aligned]: Aligning AI with Human Values (Single line; Lexend-Medium at font size 8.00 percent of slide height; near-black text)
- subtitle at x 25.7, y 61.3, w 48.6, h 5.9 [Lexend, Medium weight, sans-serif, no italics, no capitals, normal letter-spacing; font size 5.93% of slide height; #F8F8F8; center aligned]: Lessons from Building Claude (Single line; Lexend-Medium at font size 5.93 percent of slide height; off-white text)

**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: #CB785C (covers the slide)
- image at x 32.1, y 10.5, w 35.9, h 35.9: 512 by 288 pixels
- text at x 15.4, y 49.5, w 69.3, h 8.0 [Lexend-Medium; font size 8.00; #212121]: Aligning AI with Human Values
- text at x 25.7, y 61.3, w 48.6, h 5.9 [Lexend-Medium; font size 5.93; #F8F8F8]: Lessons from Building Claude

### Slide 3

**Review notes**

- Worth studying: A single-sentence mission statement given its own slide makes the core claim the audience should remember before the talk develops it.

**Text**

Mission
Develop AI that is helpful, honest, and harmless

**Layout**

A single centred text column: the title 'Mission' sits in the vertical middle of the slide (top edge about 48% down) with the subtitle line directly beneath it (top edge about 62% down), both centred on the horizontal midline. A line illustration is placed in the upper-right corner, occupying about the top 59% of the height and right third of the width, with no visible margin or frame. No footer, page number or logo appears.

**Visuals**

One illustration in the upper-right corner: a hand-drawn line drawing of a network. Five salmon-coloured filled circles (nodes) are connected by salmon straight lines, forming a branching network: a central node with lines running to upper-left, top, lower-left and lower-right nodes, plus a line running right to a node at the upper-right, which connects down-left to the lower-right node. A black, thick, hand-drawn outline of a hand (index finger extended upward, other fingers curled) points with its fingertip at the lower-right node. The subject is unphotographed; it is a flat illustration with no shading, no gradient and no text. Its bounding box is roughly square-ish within the top-right corner, not cropped, not boxed, with no overlay or tint. Dominant colours are salmon/terracotta (about #D9826B) and black (about #111111) on the light grey background.

**Design**

Background: Solid flat fill, light warm grey #E6E4E0, covering the whole slide. Flat, minimal look on a single warm light-grey background (#E6E4E0). The title 'Mission' is set in Lexend Medium at about 8.9% of slide height, colour #212121, in a bold-leaning geometric sans-serif. The subtitle is set in Lexend Regular at about 5.9% of slide height, colour #212121, in a geometric sans-serif. No accent colour in text; the only accent colours come from the illustration (salmon nodes and black hand outline). No shapes, lines, borders, corner radii, shadows or gradients are present apart from the illustration's own strokes.

**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 [#E6E4E0]: Full-slide background fill rectangle (Solid fill #E6E4E0, no border, no shadow, covers the entire slide edge to edge.)
- illustration at x 66.8, y 0, w 33.2, h 59.1 [#D9826B]: Line drawing of a network of five salmon-coloured nodes connected by lines, with a black hand-drawn hand outline with an extended index finger touching the lower-right node (Placed image, not cropped, no boxing, no overlay; nodes are filled circles about #D9826B with slightly irregular hand-drawn edges; connecting lines about #D9826B; hand outline black (about #111111) with thick hand-drawn stroke; transparent background so the #E6E4E0 fill shows through.)
- title at x 40.9, y 48.2, w 18.3, h 8.9 [Lexend Medium; font size 8.89% of slide height; #212121; center aligned]: Mission (Sentence-case, no italics, no capitals transform, normal letter-spacing, single line.)
- body at x 11.5, y 61.9, w 77, h 5.9 [Lexend Regular; font size 5.93% of slide height; #212121; center aligned]: Develop AI that is helpful, honest, and harmless (Sentence-case, no italics, no capitals transform, normal letter-spacing, single line, no 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: #E6E4E0 (covers the slide)
- image at x 66.8, y 0.0, w 33.2, h 59.1: 656 by 656 pixels
- text at x 40.9, y 48.2, w 18.3, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Mission
- text at x 11.5, y 61.9, w 77.0, h 5.9 [Lexend-Regular; font size 5.93; #212121]: Develop AI that is helpful, honest, and harmless

### Slide 4

**Review notes**

- What would change the score: Cut the duplicate agenda on slide 4 versus 5, and fix the pixelated illustration on slide 10 and the loose title and bullet alignment, with the title indented off the bullet marker on slides 7 and 9.

**Text**

Agenda
●
Alignment
●
Scalable oversight
●
Responsible deployment

**Layout**

Single-column left-aligned layout with the title and a three-item bullet list stacked in the left two-thirds of the slide, with a line illustration in the upper-right corner. The title sits at the top-left of the text block, the bullets are aligned to a common left edge, and the lower third of the slide is empty. No footer or slide number is visible.

**Visuals**

One line illustration in the upper-right corner: a network diagram drawn in terracotta-coral (about #E08A74) with five filled circular nodes connected by straight stroked lines, joined to a black hand-drawn outline of a hand with the index finger extended touching one node. Nodes sit at upper-left, top-centre, centre, right, lower-left and lower-centre, with connecting lines forming a web. The hand is a black hand-drawn outline stroke of roughly constant weight, no fill, and the index finger overlaps the lower-centre node. The image is a raster placed unboxed on the background, roughly 328 by 328 pixels, with no crop, no overlay and no tint.

**Design**

Background flat warm light grey #E6E4E0. Title and bullet text in near-black #212121 in Lexend Medium, a geometric sans-serif with medium weight, not italic and not capitalised. Title is about 8.9% of slide height and bullets are the same size, so hierarchy comes from spacing rather than size. Bullets are solid filled black circles in ArialMT. Illustration uses terracotta #E08A74 with black #111111 hand outline. No shapes, lines, shadows, gradients or corner radii on the slide.

**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 [#E6E4E0]: Full-slide background fill (Covers the entire slide, flat, no border.)
- illustration at x 66.8, y 0, w 33.2, h 59.1 [#E08A74]: Line illustration of a terracotta network diagram with five filled nodes and connecting lines, with a black hand-drawn hand whose index finger touches a node (Raster image 328 by 328 pixels, unboxed, no crop, no overlay, no shadow; nodes are filled coral circles with straight coral connecting lines; hand outline is black #111111 with variable stroke weight; sits over the background only.)
- title at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend Medium, sans-serif, medium weight, not italic, not capitalised, no letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Agenda
- body at x 15.7, y 38.5, w 3, h 8.9 [ArialMT, sans-serif, regular; font size 8.89% of slide height; #212121; left aligned]: ● (Solid filled round bullet marker.)
- bullet list at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend Medium, sans-serif, medium weight, not italic; font size 8.89% of slide height; #212121; left aligned]: Alignment (First bullet item; bullet marker at x 15.7.)
- body at x 15.7, y 49.2, w 3, h 8.9 [ArialMT, sans-serif, regular; font size 8.89% of slide height; #212121; left aligned]: ● (Solid filled round bullet marker.)
- bullet list at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend Medium, sans-serif, medium weight, not italic; font size 8.89% of slide height; #212121; left aligned]: Scalable oversight (Second bullet item; bullet marker at x 15.7.)
- body at x 15.7, y 59.8, w 3, h 8.9 [ArialMT, sans-serif, regular; font size 8.89% of slide height; #212121; left aligned]: ● (Solid filled round bullet marker.)
- bullet list at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend Medium, sans-serif, medium weight, not italic; font size 8.89% of slide height; #212121; left aligned]: Responsible deployment (Third bullet item; bullet marker at x 15.7.)

**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: #E6E4E0 (covers the slide)
- image at x 66.8, y 0.0, w 33.2, h 59.1: 328 by 328 pixels
- text at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Agenda
- text at x 15.7, y 38.5, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Alignment
- text at x 15.7, y 49.2, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Scalable oversight
- text at x 15.7, y 59.8, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Responsible deployment

### Slide 5

**Review notes**

- Worth keeping: Keep the earth-tone palette, the consistent illustration style, and the highlighted-agenda device (slides 5, 8, 10) that tells the audience where they are.

**Text**

Agenda
●  Alignment
●  Scalable oversight
●  Responsible deployment

**Layout**

Single left-aligned text column with the title Agenda above a bulleted list of three items, all starting at the same left margin of about 16% of slide width. The title sits near the top third and the list sits in the middle band, leaving the lower third empty. No footer or slide number is visible. A hand-drawn network illustration occupies the top-right corner.

**Visuals**

One illustration in the top-right corner: a network diagram in terracotta-pink (#D98C73 approx.) with about seven filled circles connected by straight lines, drawn with a black hand-drawn outline of a hand with a raised index finger touching one of the nodes. The hub node sits near the centre of the illustration and branches to satellite nodes at upper-left, top-centre, right, lower-left and lower-centre. The image is a boxed raster placed without crop, with a transparent background so the slide colour shows through. A terracotta highlight bar (#CC9B7A) sits behind the word Alignment, covering the first bullet row. No charts, tables or photographs.

**Design**

Background is a flat warm light grey (#E6E4E0). Title and bullet text are near-black (#212121) in Lexend Medium, a geometric sans-serif, rendered in a medium-bold weight. Title size is about 8.9% of slide height, and bullet text is also about 8.9% of slide height. Bullets are solid black round dots set in ArialMT. The only accent is the terracotta highlight bar #CC9B7A, a plain rectangle with square corners and no border or shadow. No gradients 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 15.7, y 37.1, w 30.8, h 10.7: Highlight rectangle behind first bullet row (Filled rectangle, colour #CC9B7A, square corners, no stroke, no shadow; sits behind the Alignment text.)
- illustration at x 66.8, y 0, w 33.2, h 59.1 [#D98C73]: Hand-drawn network diagram: terracotta nodes connected by lines, with a black hand outline touching one node (Raster image 164 by 164 px, transparent background, no crop, no border; nodes are filled circles in terracotta, connecting lines are thin terracotta strokes, hand outline is black hand-drawn stroke of about 1% of slide width.)
- title at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend Medium, sans-serif, medium weight; font size 8.89% of slide height; #212121; left aligned]: Agenda (Regular case, no italics, no letter-spacing adjustment.)
- body at x 15.7, y 38.5, w 3, h 8.9 [ArialMT; font size 8.89% of slide height; #212121; left aligned]: ● (Solid round bullet dot.)
- body at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend Medium, sans-serif, medium weight; font size 8.89% of slide height; #212121; left aligned]: Alignment (Sits on terracotta highlight bar #CC9B7A.)
- body at x 15.7, y 49.2, w 3, h 8.9 [ArialMT; font size 8.89% of slide height; #212121; left aligned]: ● (Solid round bullet dot.)
- body at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend Medium, sans-serif, medium weight; font size 8.89% of slide height; #212121; left aligned]: Scalable oversight (No highlight.)
- body at x 15.7, y 59.8, w 3, h 8.9 [ArialMT; font size 8.89% of slide height; #212121; left aligned]: ● (Solid round bullet dot.)
- body at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend Medium, sans-serif, medium weight; font size 8.89% of slide height; #212121; left aligned]: Responsible deployment (No highlight.)

**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: #E6E4E0 (covers the slide)
- fill at x 15.7, y 37.1, w 30.8, h 10.7: #CC9B7A
- image at x 66.8, y 0.0, w 33.2, h 59.1: 164 by 164 pixels
- text at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Agenda
- text at x 15.7, y 38.5, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Alignment
- text at x 15.7, y 49.2, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Scalable oversight
- text at x 15.7, y 59.8, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Responsible deployment

### Slide 6

**Text**

Helpful
Concise, efficient, appropriate
Honest
Accurate, expresses uncertainty
Harmless
Avoids harm, acts carefully

**Layout**

Single text block of nested bullets left-aligned in the middle band of the slide, starting at about 12.9% from the left edge and vertically centred; no title, footer or slide number, and the bullet glyphs sit in a hanging indent with the sub-points indented one level further.

**Visuals**

none

**Design**

Background: Solid colour #CB785C (terracotta). Flat terracotta background (#CB785C) with no gradient or texture. Text is near-black (#212121) in Lexend Bold for top-level items and Lexend Medium for sub-items, both sans-serif with geometric shapes and normal letter-spacing. Top-level text is about 7.4% of slide height (about 7.4 on a 100-unit scale, close to a 60pt body on a 7.5in slide), with sub-items at the same size. Level-one bullets are solid filled circles (●) in ArialMT; level-two bullets are open circles (○) in ArialMT. Line spacing is about 8.9% of slide height per line. No shapes, lines, shadows, corner radii or overlays.

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

- body at x 12.9, y 23.9, w 62.2, h 52 [Lexend Bold (top-level) and Lexend Medium (sub-items), sans-serif, regular letter-spacing; font size 7.41% of slide height; #212121; left aligned]: ● Helpful / ○ Concise, efficient, appropriate / ● Honest / ○ Accurate, expresses uncertainty / ● Harmless / ○ Avoids harm, acts carefully (Six lines as a bulleted list: four-line stack of bold top-level bullets (● in ArialMT, same size) each followed by a Medium-weight sub-bullet (○ in ArialMT) indented to x 17.9% for the marker and x 22.9% for the text; line pitch about 8.9% of slide height; no bullet colour difference from text; no fill or stroke; sits directly on the background with no overlap.)

**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: #CB785C (covers the slide)
- text at x 12.9, y 23.9, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 23.9, w 15.2, h 7.4 [Lexend-Bold; font size 7.41; #212121]: Helpful
- text at x 17.9, y 32.8, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ○
- text at x 22.9, y 32.8, w 62.2, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Concise, effi cient, appropriate
- text at x 12.9, y 41.7, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 41.7, w 14.7, h 7.4 [Lexend-Bold; font size 7.41; #212121]: Honest
- text at x 17.9, y 50.6, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ○
- text at x 22.9, y 50.6, w 65.8, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Accurate, expresses uncertainty
- text at x 12.9, y 59.5, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 59.5, w 19.7, h 7.4 [Lexend-Bold; font size 7.41; #212121]: Harmless
- text at x 17.9, y 68.4, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ○
- text at x 22.9, y 68.4, w 55.8, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Avoids harm, acts carefully

### Slide 7

**Review notes**

- Worth studying: Three short steps set out the Constitutional AI process as a compact sequence that a listener can follow.
- Suggested repair: Replace fragment bullets on slides 7, 9 and 11 with one concrete sentence or example each, enlarge the ASL captions on slide 12 and tie that slide to the deployment section.
- What would change the score: Cut the duplicate agenda on slide 4 versus 5, and fix the pixelated illustration on slide 10 and the loose title and bullet alignment, with the title indented off the bullet marker on slides 7 and 9.

**Text**

Constitutional AI

● Define principles
● Revise responses accordingly
● Reinforcement learning

**Layout**

Single-column text block on the left with the title and a three-item bullet list left-aligned at about 17.9% from the left edge, the title at about 23.9% from the top and the bullets spanning 41.7% to 67% of the height. A line drawing of a scroll sits in the upper-right corner, boxed by the slide edge at the top and right; there is no footer and no slide number.

**Visuals**

One hand-drawn line illustration at upper right: an unrolled scroll with black ink outlines and curled ends, drawn on a cream paper with a light-cream fill. On the scroll face a network diagram of five salmon-coloured (#D98A73 approx.) filled circles joined by straight salmon lines: a central larger node with branches to an upper-left node, an upper-right node, a lower-left node and a lower-down node. A hand with a pen (black outline, no fill) writes on the scroll from the right, the pen tip touching the lowest node. The illustration has no shadow, is not cropped and is not boxed; it is set directly on the background, with its right edge near the slide's right margin.

**Design**

Background: Solid cream/beige #EBDBBC.

**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 [#EBDBBC]: Full-slide background fill (Solid rectangle covering the whole slide, no border.)
- illustration at x 66.3, y 0, w 33.7, h 57.1 [#D98A73]: Hand-drawn scroll with network diagram of salmon nodes and a hand holding a pen (Line art in black ink (about 0.5% stroke), salmon filled nodes and connecting lines, paper fill matches background, no border, no shadow, sits over the background at upper right.)
- title at x 17.9, y 23.9, w 35.1, h 7.4 [Lexend Medium, sans-serif, medium weight, regular case; font size 7.41% of slide height; #212121; left aligned]: Constitutional AI
- bullet list at x 12.9, y 41.7, w 60.3, h 25.3 [Lexend Medium, sans-serif, medium weight, regular case; font size 7.41% of slide height; #212121; left aligned]: Define principles / Revise responses accordingly / Reinforcement learning (Three filled black round bullets (ArialMT, #212121, about 2.5% wide column at x 12.9%) at x 12.9%, y 41.7%, 50.6%, 59.5%; item text starts at x 17.9%; line pitch about 8.9% of slide height; no bullet line spacing beyond that.)
- bullet at x 12.9, y 41.7, w 2.5, h 7.4 [ArialMT; font size 7.41% of slide height; #212121; left aligned]: ● (Filled circle bullet)
- bullet at x 12.9, y 50.6, w 2.5, h 7.4 [ArialMT; font size 7.41% of slide height; #212121; left aligned]: ● (Filled circle bullet)
- bullet at x 12.9, y 59.5, w 2.5, h 7.4 [ArialMT; font size 7.41% of slide height; #212121; left aligned]: ● (Filled circle 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: #EBDBBC (covers the slide)
- image at x 66.3, y 0.0, w 33.7, h 57.1: 656 by 656 pixels
- text at x 17.9, y 23.9, w 35.1, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Constitutional AI
- text at x 12.9, y 41.7, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 41.7, w 34.5, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Deﬁne principles
- text at x 12.9, y 50.6, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 50.6, w 60.3, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Revise responses accordingly
- text at x 12.9, y 59.5, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 59.5, w 48.2, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Reinforcement learning

### Slide 8

**Review notes**

- Worth keeping: Keep the earth-tone palette, the consistent illustration style, and the highlighted-agenda device (slides 5, 8, 10) that tells the audience where they are.

**Text**

Agenda
●
Alignment
●
Scalable oversight
●
Responsible deployment

**Layout**

Single text column left-aligned at about 16% from the left edge, with the title Agenda at the top of the column and three bullet items stacked below it with even spacing. The image sits in the upper-right corner, bleeding off the top and right edges, and there is no footer or slide number.

**Visuals**

One small raster illustration in the upper right, 82 by 82 pixels, showing a network diagram: about seven coral-orange dots (nodes) joined by thin salmon-coloured lines radiating from a central node, with a black hand-drawn outline of a hand and forearm reaching in from the lower right and overlapping the network. The illustration is low-resolution with visible pixelation and a faint light halo around the drawing. A tan rectangular highlight bar (#CC9B7A) sits behind the second bullet, 'Scalable oversight', spanning most of the text line width.

**Design**

Background: Solid light warm grey #E6E4E0 covering the full slide. Plain warm light-grey background (#E6E4E0) with no gradient or texture. Title and bullet text are near-black (#212121) in Lexend Medium, a geometric sans-serif, at regular/medium weight and not italic. Body and title text are about 8.9% of slide height (cap height about 6.2%). Bullets are solid black dots in Arial. The only accent is the tan highlight #CC9B7A behind the middle item, with square corners and no shadow. No lines or borders.

**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 [#E6E4E0]: Background fill covering the full slide (Solid fill, no gradient, no border.)
- illustration at x 66.8, y 0, w 33.2, h 59.1 [#E0826A]: Network diagram with coral nodes and a hand-drawn black hand outline overlapping it, low-resolution raster 82 by 82 px (Placed raster image bleeding off the top and right edges; nodes coral-orange, connecting lines salmon, hand outline black; faint light halo around drawing; no border or shadow.)
- shape at x 15.7, y 47.7, w 51.1, h 10.7 [#CC9B7A]: Tan highlight bar behind the second bullet item (Filled rectangle, square corners, no stroke, no shadow; sits behind the text 'Scalable oversight' and its bullet.)
- title at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend Medium, sans-serif, regular/medium weight, not italic, no capitals, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Agenda
- body at x 15.7, y 38.5, w 3, h 8.9 [Arial, sans-serif, regular, solid round bullet; font size 8.89% of slide height; #212121; left aligned]: ● (Bullet glyph for item 1.)
- body at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend Medium, sans-serif, regular/medium weight, not italic, no capitals, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Alignment (Bullet list item 1.)
- body at x 15.7, y 49.2, w 3, h 8.9 [Arial, sans-serif, regular, solid round bullet; font size 8.89% of slide height; #212121; left aligned]: ● (Bullet glyph for item 2, on the tan highlight.)
- body at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend Medium, sans-serif, regular/medium weight, not italic, no capitals, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Scalable oversight (Bullet list item 2, sitting on the tan highlight bar.)
- body at x 15.7, y 59.8, w 3, h 8.9 [Arial, sans-serif, regular, solid round bullet; font size 8.89% of slide height; #212121; left aligned]: ● (Bullet glyph for item 3.)
- body at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend Medium, sans-serif, regular/medium weight, not italic, no capitals, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Responsible deployment (Bullet list item 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: #E6E4E0 (covers the slide)
- fill at x 15.7, y 47.7, w 51.1, h 10.7: #CC9B7A
- image at x 66.8, y 0.0, w 33.2, h 59.1: 82 by 82 pixels
- text at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Agenda
- text at x 15.7, y 38.5, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Alignment
- text at x 15.7, y 49.2, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Scalable oversight
- text at x 15.7, y 59.8, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Responsible deployment

### Slide 9

**Review notes**

- Worth studying: Three parallel method names under one heading show how a single goal is broken into concrete approaches.
- Suggested repair: Replace fragment bullets on slides 7, 9 and 11 with one concrete sentence or example each, enlarge the ASL captions on slide 12 and tie that slide to the deployment section.
- What would change the score: Cut the duplicate agenda on slide 4 versus 5, and fix the pixelated illustration on slide 10 and the loose title and bullet alignment, with the title indented off the bullet marker on slides 7 and 9.

**Text**

Scalable oversight / Automated testing / Expert red teaming / Crowdsourcing

**Layout**

Two-zone composition on a single flat field: a left-aligned text block (title 'Scalable oversight' followed by a blank line and a three-item bullet list) occupies the left half, starting around 18% from the left edge and 24% from the top; a single illustration sits at the right, vertically centred in the lower two-thirds and extending to about 96% of the width. No title bar, footer, or slide number is present.

**Visuals**

One illustration: a hand-drawn line-art microchip (computer processor/CPU icon) drawn in heavy black strokes, with pins along the top, bottom and both sides, an inner square die containing a circuit trace with three small open circles at the top and bottom ends. It is placed over a soft peach/terracotta watercolour-style blob shape with a flame-like pointed top-right tip, fully visible (not cropped), not full-bleed, boxed only by its own transparent background. Lighting is flat; no shadow. Dominant colours: black line art on peach #E8A987-like blob on light grey field. Composition: centred on the blob, roughly square image of 656 by 656 pixels.

**Design**

Background: Flat solid colour #E6E4E0 (light warm grey) covering the entire slide. Flat light warm-grey background #E6E4E0 with no gradient or texture. Title and bullet text in near-black #212121, set in Lexend Medium (geometric humanist sans-serif, medium weight), no italics, no capitals, normal letter-spacing. Title text height about 7.4% of slide height (cap height roughly 5.2%). Body bullets identical size 7.41% of slide height, same colour, with generous line spacing (about 9% of slide height between baselines). Bullet glyphs are solid black dots set in Arial, same colour #212121. Illustration uses a warm peach accent (#E8A987-like) behind black line art. No lines, borders, shadows, rounded corners, or gradients on 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 [#E6E4E0]: Full-slide background fill (Solid flat fill covering the entire slide, no border or shadow.)
- title at x 17.9, y 23.9, w 38, h 7.4 [Lexend Medium, sans-serif, weight medium, normal case, no letter-spacing; font size 7.41% of slide height; #212121; left aligned]: Scalable oversight
- body at x 12.9, y 41.7, w 2.5, h 7.4 [Arial, sans-serif, regular; font size 7.41% of slide height; #212121; left aligned]: ● (Solid round bullet glyph, paired with the line of text to its right.)
- body at x 17.9, y 41.7, w 38.3, h 7.4 [Lexend Medium, sans-serif, weight medium, normal case; font size 7.41% of slide height; #212121; left aligned]: Automated testing
- body at x 12.9, y 50.6, w 2.5, h 7.4 [Arial, sans-serif, regular; font size 7.41% of slide height; #212121; left aligned]: ● (Solid round bullet glyph.)
- body at x 17.9, y 50.6, w 39.9, h 7.4 [Lexend Medium, sans-serif, weight medium, normal case; font size 7.41% of slide height; #212121; left aligned]: Expert red teaming
- body at x 12.9, y 59.5, w 2.5, h 7.4 [Arial, sans-serif, regular; font size 7.41% of slide height; #212121; left aligned]: ● (Solid round bullet glyph.)
- body at x 17.9, y 59.5, w 30.6, h 7.4 [Lexend Medium, sans-serif, weight medium, normal case; font size 7.41% of slide height; #212121; left aligned]: Crowdsourcing
- illustration at x 61.6, y 31.7, w 38.4, h 68.3 [#E8A987]: Hand-drawn black line-art microchip (CPU icon) with pins on all four sides and an inner circuit trace with circular pads, layered over a peach watercolour-style blob with a flame-like pointed top-right tip; no cropping, boxed by transparent PNG background, 656 by 656 pixels (Blob fill peach/terracotta #E8A987 with soft watercolour edges; chip outline and traces solid black #000000 at heavy stroke weight; no shadow, no opacity change, no rotation.)

**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: #E6E4E0 (covers the slide)
- image at x 61.6, y 31.7, w 38.4, h 68.3: 656 by 656 pixels
- text at x 17.9, y 23.9, w 38.0, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Scalable oversight
- text at x 12.9, y 41.7, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 41.7, w 38.3, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Automated testing
- text at x 12.9, y 50.6, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 50.6, w 39.9, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Expert red teaming
- text at x 12.9, y 59.5, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 59.5, w 30.6, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Crowdsourcing

### Slide 10

**Review notes**

- Worth keeping: Keep the earth-tone palette, the consistent illustration style, and the highlighted-agenda device (slides 5, 8, 10) that tells the audience where they are.
- What would change the score: Cut the duplicate agenda on slide 4 versus 5, and fix the pixelated illustration on slide 10 and the loose title and bullet alignment, with the title indented off the bullet marker on slides 7 and 9.

**Text**

Agenda
Alignment
Scalable oversight
Responsible deployment

**Layout**

A single left-aligned column of text sits in the left two-thirds of the slide, with the title Agenda at top and three bullet items below it on a shared left edge at about 16% of the width. A pixelated illustration is anchored to the top-right corner and bleeds off the top and right edges. The highlight bar behind the last bullet extends from the bullet column to just past the end of its text. No footer, slide number or logo is visible.

**Visuals**

One pixelated illustration in the top-right corner, about 33% of the slide width by 59% of its height, cropped by the top and right slide edges. It is a 41 by 41 pixel mosaic showing a dark grey outline of a face or profile in the upper right, with pink and peach blocks on its left and a pale grey-white field around it. The blocks are soft and low resolution with no overlay. Lighting is flat and the dominant colours are pale off-white, dark grey and muted peach. The only other visual element is a single filled highlight rectangle behind the third bullet, colour #CC9B7A, with no border.

**Design**

Background is a flat warm light grey #E6E4E0 with no gradient or texture. Title Agenda and all three bullet items are set in Lexend Medium, a geometric sans-serif, in near-black #212121, with no italics or capitals and normal letter-spacing. The bullet glyphs are set in ArialMT in #212121. Title and body text are the same size, about 8.9% of slide height (about 10 pt equivalent on a 7.5 in slide). The accent is the terracotta #CC9B7A highlight bar, which has square corners, no shadow and no stroke. Bullets are solid round dots, and lines are spaced about 10.7% of slide height apart. The image has no shadow or border.

**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 [#E6E4E0]: Full-slide background fill (Solid rectangle covering the whole slide, no border, no gradient.)
- photo at x 66.8, y 0, w 33.2, h 59.1 [#E6E4E0]: Pixelated 41 by 41 pixel illustration of a dark grey face or profile outline with pink and peach blocks, cropped at top and right slide edges (Mosaic image, square pixels, no border, no shadow, no overlay, bleeds off top and right edges, overlaps nothing.)
- shape at x 15.7, y 58.4, w 65.7, h 10.7 [#CC9B7A]: Highlight bar behind third bullet (Solid rectangle, square corners, no stroke, no shadow, sits behind the text Responsible deployment.)
- title at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend-Medium, sans-serif, medium weight, normal case, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Agenda
- body at x 15.7, y 38.5, w 3, h 8.9 [ArialMT, sans-serif, regular weight; font size 8.89% of slide height; #212121; left aligned]: ● (Solid round bullet glyph.)
- body at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend-Medium, sans-serif, medium weight, normal case, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Alignment
- body at x 15.7, y 49.2, w 3, h 8.9 [ArialMT, sans-serif, regular weight; font size 8.89% of slide height; #212121; left aligned]: ● (Solid round bullet glyph.)
- body at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend-Medium, sans-serif, medium weight, normal case, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Scalable oversight
- body at x 15.7, y 59.8, w 3, h 8.9 [ArialMT, sans-serif, regular weight; font size 8.89% of slide height; #212121; left aligned]: ● (Solid round bullet glyph, sits on the #CC9B7A highlight bar.)
- body at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend-Medium, sans-serif, medium weight, normal case, normal letter-spacing; font size 8.89% of slide height; #212121; left aligned]: Responsible deployment (Sits on the #CC9B7A highlight bar.)

**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: #E6E4E0 (covers the slide)
- fill at x 15.7, y 58.4, w 65.7, h 10.7: #CC9B7A
- image at x 66.8, y 0.0, w 33.2, h 59.1: 41 by 41 pixels
- text at x 16.2, y 17.2, w 19.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Agenda
- text at x 15.7, y 38.5, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 38.5, w 25.2, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Alignment
- text at x 15.7, y 49.2, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 49.2, w 45.5, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Scalable oversight
- text at x 15.7, y 59.8, w 3.0, h 8.9 [ArialMT; font size 8.89; #212121]: ●
- text at x 21.2, y 59.8, w 60.1, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Responsible deployment

### Slide 11

**Review notes**

- Suggested repair: Replace fragment bullets on slides 7, 9 and 11 with one concrete sentence or example each, enlarge the ASL captions on slide 12 and tie that slide to the deployment section.

**Text**

Deployment
Strong demand at launch
Data privacy matters!
Layered response to new attacks

**Layout**

Single text block left-aligned at about 17.9% from the left edge, with the title Deployment at the top of the content block and a bullet list of three items below it. A small illustration sits at the upper right, outside the text column. No footer, page number or logo is visible.

**Visuals**

One illustration at the upper right: a flat, hand-drawn style set of cream-coloured (#F5EEDC-ish, off-white on beige) rounded UI control icons arranged in a grid, including a horizontal slider bar with a knob at top, a vertical slider at right, a circular dial/knob, several square toggle-like tiles with circles and small rotary dials. Overlaying the grid is a black hand-drawn pointing index finger, drawn with a thick black marker stroke, pointing up and to the right at the control panel. The finger's tip sits over the upper-right of the grid and its base extends down-left toward the bottom-left of the illustration. The illustration is boxed on the beige background with no visible border, not cropped, and lit flatly with no shading.

**Design**

Background: Solid beige #EBDBBC covering the entire slide. Flat beige background #EBDBBC with no gradient or texture. Title and bullet text are dark charcoal #212121 in Lexend Medium, a geometric sans-serif, regular-to-medium weight, not italic, no capitals, no letter-spacing. Title Deployment is about 7.4% of slide height (about 8.3 pt equivalent by cap-height reading scale of the slide), body bullets are the same size and weight. Bullets are solid black filled circles (#212121) in ArialMT, set in a separate text box. No accent colour, no lines or 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 [#EBDBBC]: Full-slide background fill, solid beige (Covers the entire slide; no border, no shadow.)
- illustration at x 70.3, y 0, w 29.7, h 53.2 [#F5EEDC]: Hand-drawn cream UI control panel (horizontal sliders, dials, toggle tiles, vertical slider) overlaid by a black hand-drawn pointing finger (Image 656 by 656 pixels, placed as a box, not cropped; the panel is cream on beige; the finger is black #000000 with a thick marker stroke; no shadow, no opacity change, no rotation.)
- title at x 17.9, y 23.9, w 24.9, h 7.4 [Lexend Medium, sans-serif, medium weight, regular, no capitals, no letter-spacing; font size 7.41% of slide height; #212121; left aligned]: Deployment
- bullet list at x 12.9, y 41.7, w 54.8, h 7.4 [Lexend Medium, sans-serif, medium weight, regular, no capitals, no letter-spacing; bullet in ArialMT; font size 7.41% of slide height; #212121; left aligned]: ● Strong demand at launch (Solid round bullet glyph in separate text box at x 12.9, text at x 17.9; line spacing about 8.9% of slide height between items.)
- bullet list at x 12.9, y 50.6, w 50, h 7.4 [Lexend Medium, sans-serif, medium weight, regular, no capitals, no letter-spacing; bullet in ArialMT; font size 7.41% of slide height; #212121; left aligned]: ● Data privacy matters! (Solid round bullet glyph in separate text box at x 12.9, text at x 17.9.)
- bullet list at x 12.9, y 59.5, w 72.6, h 7.4 [Lexend Medium, sans-serif, medium weight, regular, no capitals, no letter-spacing; bullet in ArialMT; font size 7.41% of slide height; #212121; left aligned]: ● Layered response to new attacks (Solid round bullet glyph in separate text box at x 12.9, text at x 17.9.)

**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: #EBDBBC (covers the slide)
- image at x 70.3, y 0.0, w 29.7, h 53.2: 656 by 656 pixels
- text at x 17.9, y 23.9, w 24.9, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Deployment
- text at x 12.9, y 41.7, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 41.7, w 52.3, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Strong demand at launch
- text at x 12.9, y 50.6, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 50.6, w 45.0, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Data privacy matters!
- text at x 12.9, y 59.5, w 2.5, h 7.4 [ArialMT; font size 7.41; #212121]: ●
- text at x 17.9, y 59.5, w 67.8, h 7.4 [Lexend-Medium; font size 7.41; #212121]: Layered response to new attacks

### Slide 12

**Review notes**

- Review: Slide 12, the only information-bearing graphic (ASL levels), is the weakest legible page: tiny captions in a flat four-box row with an arrow, and it appears after the deployment slide with no framing and no link to the agenda.
- Suggested repair: Replace fragment bullets on slides 7, 9 and 11 with one concrete sentence or example each, enlarge the ASL captions on slide 12 and tie that slide to the deployment section.

**Text**

High level overview of AI Safety Levels (ASLs)
ASL-1
Smaller / models
ASL-2
Present / large models
ASL-3
Significantly / higher risk
ASL-4+
Speculative
Increasing model capability, / Increasing security and safety measures

**Layout**

Four equal-width vertical columns of rectangles sit in a row across the middle of the slide, with a left-aligned title above them and a left-aligned two-line caption below an arrow. Margins are about 4% on the left and right, columns have roughly 1% gutters, and there is no footer, logo or visible slide number.

**Visuals**

Four adjacent flat rectangles, left to right: ASL-1 near-white (#FCFCFA), ASL-2 pale sand (#EDDABB), ASL-3 light terracotta (#DCA079), ASL-4+ coral-orange (#DA7356). Each rectangle is about 22% of slide width and 50% of slide height, with square corners and no border or shadow. Each holds a bold label at its top-left and a serif description beneath it. Below them runs a thin black horizontal arrow with a solid triangular head pointing right, spanning about 85% of slide width; it indicates increasing levels from left to right. No photographs, icons or logos. Nothing is highlighted beyond the colour progression.

**Design**

Background: Flat light warm grey #F1F1EC, no gradient, photograph or pattern. Light warm-grey background (#F1F1EC). Title and caption text are near-black (#1A1A1A) in a geometric sans-serif, semibold for the title and medium for the caption. Column labels are near-black bold geometric sans. Descriptions are regular-weight serif in near-black (#1A1A1A). Title cap height is about 5% of slide height, so the font size is about 7% of slide height; column labels are about 6%; descriptions about 4%; caption about 4.3%. Shapes are flat solid fills with square corners, no gradients or shadows. The only line is the black arrow, about 0.5% of slide height thick.

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

- shape at x 3.9, y 20.7, w 22, h 49.8 [#FCFCFA]: Rectangle for ASL-1, plain fill (Flat fill, square corners, no border, no shadow, opacity 100%.)
- shape at x 26.8, y 20.7, w 21.9, h 49.8 [#EDDABB]: Rectangle for ASL-2, plain fill (Flat fill, square corners, no border, no shadow.)
- shape at x 49.7, y 20.7, w 22, h 49.8 [#DCA079]: Rectangle for ASL-3, plain fill (Flat fill, square corners, no border, no shadow.)
- shape at x 72.6, y 20.7, w 22, h 49.8 [#DA7356]: Rectangle for ASL-4+, plain fill (Flat fill, square corners, no border, no shadow.)
- title at x 5.4, y 7.5, w 80, h 6 [geometric sans-serif, semibold; font size 7% of slide height; #1A1A1A; left aligned]: High level overview of AI Safety Levels (ASLs) (Normal case, no italics, no letter-spacing.)
- subtitle at x 5.3, y 25.3, w 8.8, h 4.5 [geometric sans-serif, bold; font size 6% of slide height; #1A1A1A; left aligned]: ASL-1 (Sits on the ASL-1 rectangle.)
- body at x 5.3, y 33.6, w 9.5, h 8.2 [serif, regular; font size 4.2% of slide height; #1A1A1A; left aligned]: Smaller / models (Two lines, line spacing about 1.2, sits on the ASL-1 rectangle.)
- subtitle at x 28.5, y 25.3, w 9.1, h 4.5 [geometric sans-serif, bold; font size 6% of slide height; #1A1A1A; left aligned]: ASL-2 (Sits on the ASL-2 rectangle.)
- body at x 28.5, y 33.6, w 12.8, h 8.2 [serif, regular; font size 4.2% of slide height; #1A1A1A; left aligned]: Present / large models (Two lines, line spacing about 1.2, sits on the ASL-2 rectangle.)
- subtitle at x 51.3, y 25.3, w 9.2, h 4.5 [geometric sans-serif, bold; font size 6% of slide height; #1A1A1A; left aligned]: ASL-3 (Sits on the ASL-3 rectangle.)
- body at x 51.2, y 33.6, w 12.8, h 8.2 [serif, regular; font size 4.2% of slide height; #1A1A1A; left aligned]: Significantly / higher risk (Two lines, line spacing about 1.2, sits on the ASL-3 rectangle.)
- subtitle at x 74.3, y 25.3, w 11.3, h 4.5 [geometric sans-serif, bold; font size 6% of slide height; #1A1A1A; left aligned]: ASL-4+ (Sits on the ASL-4+ rectangle.)
- body at x 74.2, y 33.6, w 11.6, h 3.8 [serif, regular; font size 4.2% of slide height; #1A1A1A; left aligned]: Speculative (Single line, sits on the ASL-4+ rectangle.)
- line at x 5.2, y 77.4, w 85, h 1.5 [#000000]: Horizontal arrow pointing right, running under the four rectangles (Line about 0.5% of slide height thick, solid black, with a filled triangular arrowhead at the right end, no dash pattern.)
- body at x 5.4, y 82.5, w 50.2, h 10.5 [geometric sans-serif, medium; font size 4.3% of slide height; #1A1A1A; left aligned]: Increasing model capability, / Increasing security and safety measures (Two lines, line spacing about 1.5 of size; the first line ends with a comma.)

### Slide 13

**Text**

Let’s build safe AGI

**Layout**

Single centred composition on a plain field: an illustration sits in the upper-middle of the slide, with the title centred beneath it on the horizontal axis. No footer, slide number, or other elements; the lower half of the slide is empty.

**Visuals**

One line-drawn illustration above the title: a stack of five hand-drawn rounded stones (small stone on top, then two medium stones, then a large stone at right and a large stone at lower left, with a small pebble at lower right) balanced in a pyramid, with a hand drawn from the left reaching in to rest on the stack at its middle-left. The outlines are black, about 2-3 px wide, with a loose, sketchy hand-drawn quality. A translucent terracotta-coloured circle (about #D9826A at partial opacity) sits behind the hand, overlapping the stones and hand, acting as a soft highlight spot. The illustration is not cropped and is boxed only by the slide background, with no frame or border.

**Design**

Background: Solid colour #D4A27F (tan terracotta) covering the entire slide. Flat solid tan/terracotta background, #D4A27F, with no gradient or texture. Title text is near-black #212121 in a geometric rounded sans-serif (Lexend Medium), roughly medium weight, not bold, set in sentence case with no capitals except 'AGI', at about 8.9% of slide height for the cap-height line. The illustration uses black #000000 line strokes with a terracotta accent fill (#D9826A-ish, semi-transparent). No shapes, shadows, or corner radii other than the organic outlines of the drawn stones.

**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 [#D4A27F]: Full-slide background fill (Solid flat fill, no gradient, covers entire slide, back-most layer.)
- illustration at x 42.6, y 11.5, w 14.8, h 25 [#D9826A]: Hand-drawn line illustration of a balanced stack of five rounded stones with a hand resting on the middle-left stone, with a translucent terracotta circle behind the hand (Black hand-drawn outlines about 2-3 px, stroke colour #000000; translucent terracotta circle (about #D9826A, partial opacity) behind hand and stones; no frame, not cropped, transparent background so the slide colour shows through.)
- title at x 26.7, y 47.9, w 48.3, h 8.9 [Lexend Medium (geometric sans-serif, medium weight, not italic, no capitals transform, no letter-spacing change); font size 8.89% of slide height; #212121; center aligned]: Let’s build safe AGI

**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: #D4A27F (covers the slide)
- image at x 42.6, y 11.5, w 14.8, h 25.0: 256 by 245 pixels
- text at x 26.7, y 47.9, w 48.3, h 8.9 [Lexend-Medium; font size 8.89; #212121]: Let’s build safe AGI

## Transcript

Captions (automatic) of the recording "Lessons Learned While Building Anthropic's LLM Product, Claude, Ben Mann" (https://www.youtube.com/watch?v=GWXfpFchkMQ). Timestamps are minutes and seconds into the recording.

[00:05] hello everyone I'm honored to be here at the AI conference speaking with you all today I'm Ben man a co-founder and member of technical staff at anthropic an AI safety startup based in San Francisco as many of you know generative AI exploded onto the scene last year with chatbots like cloud. a systems like Dolly 2 stable diffusion anthropic debuted our chatbot cloud. in July of this year in the US and UK we've been serving customers like slack zoom and notion for almost a year via our API button button there we go but prior to that release we spent over two years researching AI safety to make Claude as helpful honest and harmless as possible today I'll talk about what we learned as we prepared to deploy and then deployed CLA at scale through rigorous

[01:09] research responsible deployment techniques focused on aligning AI systems with human values we can create artificial intelligence that benefits Humanity in a safe and ethical manner this belief guides our Approach at anthropic in my talk today I'll share the top lessons we learned building CLA including techniques for aligning an AI system with human values and establishing scalable oversight of model outputs before diving deeper into our techniques it's important to clarify what exactly we mean when we say we're developing AI that is helpful honest and harmless when we say helpful we mean the AI should make a clear attempt to perform a task or answer a question as concisely and efficiently as possible it should ask relevant follow-up questions when more information is needed and respond with appropriate sensitivity and discretion by honest we mean the AI should provide accurate information it should Express appropriate uncertainty and be honest about what it does and doesn't know the AI should not try to

[02:12] fake knowledge or falsely appear competent Beyond its abilities such as claiming that it will remember conversations or update itself and when we say harmless we mean that AI should not take actions that cause harm either through offensive or discriminatory Behavior or by aiding dangerous or illegal activities that the AI should recognize potentially sensitive situations and act with care and humility these criteria can sometimes conflict and there are edge cases and nuances to work out but at a high level an AI assistant that is helpful honest and harmless will act in accordance with human values and preferences ensuring systems meet these criteria is crucial as AI grows more advanced and influential deciding which criteria to prioritize in is an immense challenge given the vast base of potential issues at anthropic we focused on helpfulness harmlessness and honesty as top level goals based on several considerations first these aims Encompass many Downstream safety properties an AI assistant that is

[03:14] helpful honest and harmless will likely avoid problems like toxicity bias and unreliability second the criteria represent clear end goals rather than specific techniques this allows us to experiment with various methods to instill these virtues into AI systems and third helpfulness honesty and harmlessness reflect core human values that translate across cultures we believe these principles constitute a strong foundation for Value alignment as AI capabilities Advance through rigorous research and responsible development techniques focused on aligning AI systems with human values we can create beneficial AI how do we actually do this alignment let's dive into our training process we found that very simple prompting can go a long way in mid 2021 we wrote a prompt with 14 hypothetical example conversations between humans and an AI assistant we wrote the examples such that the assistant exemplified helpfulness harmlessness and honesty whether in simple or difficult conversations without any further

[04:17] training prompting large language models with these conversations alone significantly improved their performance on safety evaluations we had written to test for helpless helpfulness honesty and harmlessness however prompting alone has its limitations the assistant can only mimic the patterns in the prompt and May Fail on new situations not foreseen by The Prompt designer oops in 2022 we developed a breakthrough that allowed us to steer Model Behavior using natural language alone with no human labeling of responses we call this constitutional AI constitutional AI leverages the AI itself to provide critiques and revisions of its own potentially harmful outputs based on adherence to a set of natural language constitutional principles that we defined constitutional AI provides a principled approach for aligning large language models like Claude with human values rather than relying solely on human feedback during training constitutional AI makes the values guiding the AI explicit through a set of principles such as avoiding toxicity or

[05:21] respecting privacy we first trained Claude to critique and revise its own responses using these principles next Claude learned to identify harmful responses and rewrite them to be more ethical next Claud was trained via reinforcement learning to choose the safer response when given two options using the principles rather than human judgments the resulting model responds appropriately to dangerous inputs yet remains helpful without any human generated training data claud's Constitution which we published provides transparency into claud's goals and allows values to be adjusted over time we drew principles from sources like the UN Declaration of Human Rights right platform guidelines non-western perspectives and our own research while not a complete solution constitutional AI offers techniques for aligning llms with human values in a scalable manner the second key challenge we focused on was establishing scalable oversight of cla's behavior as conversational AIS become more advanced and widely used how can we ensure they

[06:24] behave appropriately at anthropic we recognized early on their traditional techniques like manual content moderation will not suffice reviewing every chatbot interaction does not scale and inevitably some harmful responses will slip through instead we take a multi-pronged approach combining automated testing red teaming and crowd sourcing I'll elaborate on each of these components first we automatically generate a large diverse set of test questions using language models themselves as described in our paper discovering language model behaviors with model written evaluations we create over 150 evaluations covering topics from social biases to privacy risks each test has over 1,000 examples like asking if the chatbot wants to avoid being shut down or whether it exhibits racist beliefs we generate the data efficiently by instructing language models to produce examples of this unsafe Behavior then we use other models to filter for Quality the pipeline lets us rapidly test for Myriad potential

[07:28] failures at a scale and speed that would be challenging to reproduce with human workers however model generated evaluations are not foolproof clever edge cases and exploits often require human creativity to discover that's where Red teaming comes in red teaming complements automated testing by finding more subtle issues through human Ingenuity at anthropic we have teams of employees act as adversaries to uncover flaws they before they impact external users however for Frontier Technologies like like AI red teaming requires investments in specialized expertise over the last 6 months we conducted an intensive biological red teaming project with top biocurity experts we spent over 150 hours working with these experts to probe our models capabilities to Output harmful biological information the experts used a secure interface with our standard Safety Systems disabled we observed the model exhibiting concerning capabilities sometimes producing sophisticated accurate knowledge at an expert level this does not happen

[08:32] frequently today but we found indications that larger models are more capable models gaining access to tools could also Advance their capabilities in biology taken together we believe unmitigated models could accelerate efforts to misuse biology compared to just having internet access models may enable tasks not otherwise achievable we estimate these risks could arise as soon as 2 to 3 years from now as models advance on the bright side discovering risks enables developing mitigations we found ways to update our training procedures that reduce harmful outputs by improving model's ability to distinguish between beneficial and dangerous uses classifier filters also make it harder to get the chained expert insights needed for harm we have deployed these in our public model and identified further mitigations to experiment with this work confirms Frontier red timing is crucial And Timely current models show early warning signs presenting a window to evaluate emerging risks and mitigate them before

[09:35] acute threats emerge longer-term risks like deception also warrant proactive red teaming by anticipating future dangerous capabilities models should lack we can measure risks build alignment and inform mitigation efforts before these issues become severe finally we leverage the wisdom of crowds external crowdsourcing provides diverse perspectives to catch biases crowd workers chat with our a AI assistant and flag concerning responses aggregating many viewpoints helps reveal biases that are automated methods Miss in summary scalable oversight of industrial chat Bots requires evaluating from every angle automated model written tests surface a breadth of potential fitf Falls at massive scale red teaming addresses gaps through targeted human creativity crowdsourcing collects diverse subjective viewpoints while research provided the critical groundwork seeing our methods succeed with millions of users has further validated our approach enough about research let's talk about putting these models into the

[10:39] world we began collecting human feedback data in April 2021 and began to feel confident in our safety methods in 2022 but our API wasn't yet ready to reliably serve users at scale so we rote our software stack from scratch to ensure it could handle scaled user demand from day one most companies take a launch then iterate approach but we guess that as soon as we open for business we would be flooded with traffic this prediction proved accurate and thanks to our Enterprise Partnerships we're on track to be one of the fastest growing companies by annualized Revenue in history we heard from our partners that a key barrier to leveraging llms like Claude is protecting their proprietary data especially in regulated Industries but they lack the capability to monitor what kind of data was flowing through their API keys so we structured our data use agreements to ensure we would never train generative models on user data only trust and safety classifiers which cannot leak private information we then surface specific warning flags showing Partners things like how many of their requests contain spam misinformation or violence so they can take mitigating

[11:44] actions we frequently see spikes of abusive content as users discover new ways to circumvent our safety systems such as creating fictional roleplay scenarios as many of you know these exploits are often called jailbreaks when a Jailbreaker occurs we take a multi-layered approach designed to quickly patch the attack factor and then follow up with a more comprehensive fix in a day or two over time our goal is to improve both the speed and efficacy of these patches and fixes I'm proud to cite a recent Carnegie melon paper stating that while they achieved up to 84% attack success on open AI models 66% attack success on Google's Palm 2 uh on the contrary claud's robustness resulted in just a 2.1% success rate this demonstrates the strength of our safety methods though of course we still have work to do to bring that last 2% down looking further ahead we believe the stakes are enormous if AI alignment fails as progress continues rapid advances suggest human level AI may arrive much sooner than expected raising crucial questions around how to safely

[12:48] steer such powerful systems to proactively address risks from advancing AI anthropic has developed a framework called AI safety levels modeled after bios safety protocols it defines escalating safety requirements as AI becomes more capable under this framework Claude is rated AI safety level two or asl2 meaning it shows signs of dangerous capabilities but lacks reliability in our responsible scaling policy we have committed to increasingly strict standards before progressing to asl3 and beyond for example we will not deploy asl3 systems if red team testing reveals meaningful risks of catastrophic misuse anthropic goal is to channel competitive Dynamics towards solving alignment challenges in order to safely unlock further progress We Believe Empirical research with more advanced models is essential to prepare for human level AI understanding how to train and test near human systems will provide critical

[13:52] Insight in closing an empirical iterative approach combining Theory and Hands-On research search offers the most promising path ahead progress will require bold ideas rigor and deliberation around human values if we work together proactively AI can Empower Humanity while avoiding catastrophic risks anthropic will continue driving progress toward helpful honest and harmless AI benefiting everyone we urge others to join us in this Mission and welcome partnership in charting a course toward a flourishing and just AI powered future I'm looking forward to the rest of the conference and to discussing these important issues with all of you thank you
