Fixation and evergreen review of 05 and 06#

Review of 05-maynard-risk-and-ai-map.md (05) and 06-huang-and-late-lessons-through-maynard.md (06) against perspective/portrait.md and facets F1–F6. Claims were checked against the post corpus (working/maynard/corpus/), the FFTF text (working/maynard/book/FFTF-D.txt), the papers in Resources/maynard-papers/, and the series-introduction drafts in Maynard stuff/. Line numbers refer to the versions of 27 September 2026. Posts are cited by date and slug, papers by file or key and page, and FFTF by printed page.

The question. Where do 05 and 06 lean on single remarks, drafts, asides or recent work when his published body of work speaks more fully? And where will the text read badly in a few years? The portrait’s checklist names both failure modes: “Fixating on single remarks … instead of patterns sustained over years”, and “Over-weighting his recent AI work” (portrait §10).

Headline numbers. - 06. Its most-cited dated source is 2026-09-15 will-ai-really-kill-us-all, a 1,310-word post cited 28 times, which puts it ahead of every paper he has written. 2026 items make up about a quarter of 06’s dated citations, and 2023–26 items nearly three-fifths. An unpublished draft is cited 9 times. The words “film”, “Ex Machina”, “serendipity” and “curiosity” never appear, and “risk innovation” appears three times. - 05. Among its most-cited sources are 2026-07-16 [mixed] (33), 2026-05-21 (26), 2026-09-24 [mixed] (24), NANO 2026 (23), CR 2026 (22) and FWB 2026 (21). A private phrase, “the hubris of risk assessment”, appears 10 times.

Ranked issues#

1. The toxicology reading rests on one parenthesis in a draft#

Location. 06 §8.5, “Where toxicology’s counterpart lies” and “Reading” (lines 827–833). Echoed in §2 (line 153) and §10.2 (line 960).

Problem. 06 builds its account of how toxicology carries over to AI by decoding one parenthesis in the draft series introduction: “(for instance claiming that AI is not biology and so Late Lessons chapters focused on toxicology do not apply — something I would disagree with)”. It calls his transfers “tentative” and ends: “That this is what his reservation means rests on one parenthesis; he may have meant something more literal.” His published work needs no decoding. It sets out this conceptual transfer at length, repeatedly, over seven years. Treating the parenthesis as the source makes a sustained position look like a hint, and it invites a “he may have meant” hedge that the record makes unnecessary.

Evidence from his work. - 2019-03-05 should-we-be-treating-algorithms-the-same-way-we-treat-hazardous-chemicals. He takes five concepts from chemical risk assessment: hazard versus risk, what turns potential harm into harm, consequences, how much “stuff” (dose–response), and checks on evidence. He says “they are directly applicable to research on algorithm-based risks”, and names the break: “Of course, an algorithm is not a chemical.” Exposure is placed in the people affected: “Anyone who is potentially impacted by the deployment of an algorithm can be thought of as being exposed to it”. The harms he lists include “liberty, dignity, and self-respect”. - 2023-11-26 everything-youve-heard-about-ai-risk-is-wrong. This is the substantive source. He returns to first principles: “no cause, no risk”; magnitude; harm as lost value; time, including acute exposure and “chronic exposure”; and perception. The 27 November addendum then works hazard × exposure through for AI: - exposure can be “as intangible as hints of ideas encountered over hours of social media use”; - the transforming function may have a threshold or run in either direction; - “there may well be mileage in using the hazard-exposure paradigm in the context of AI”, as part of rethinking risk “in the face of a technology that defies conventional thinking”.

The transfer is conceptual by design. He kept the formula out of the main text so as not to imply “that zero exposure — as in no AI — is a default risk management strategy”. - 2024-06-20 ilya-sutskevers-safe-superintelligence-rethink. Acceptable safety becomes “infinitely more complex” for chemical and biological agents because of “dose-response relationships, acute versus chronic impacts, the roles of perception and behavior in mediating consequences”. This is the toxicological reason AI safety is harder than engineering safety. - His nanotechnology papers supply the method for when a toxicology does not fit. - Toxicol. Sci. 2011 (2011_Maynard-Warheit-Philbert_New-Toxicology-of-Sophisticated-Materials_PMC.md, lead author) builds a “new toxicology” on emergent risk, plausibility and impact. It calls these principles “technology independent”, and warns that “emergent risks may arise as dose-response relationships are inappropriately quantified” when hazard scales with a metric other than the one measured. - “Nanomaterials are not just chemicals” (nnano.2016.28 p.211) is the template for “an algorithm is not a chemical”. - ILSI 2005 and Nature 2006 p.268 (both co-written) measure several dose metrics when the right one is unknown. - 2026-01-10 is-ai-a-cognitive-trojan-horse. He places AI among evolutionary mismatches with “synthetic chemicals, vaccines”.

Fix. - Rewrite §8.5 to open from this published sequence, labelled [Stated]: 2019, then 2023, 2024 and 2026, with the 2011 and 2016 nanotoxicology papers as its method. - Delete “tentative” and the “rests on one parenthesis” sentence. - Cite the series introduction once, as a recent remark consistent with that record. - In §10.2, attribute the receptor-side reading of leaded petrol to 2019-03-05 and 2023-11-26. - In 05, move “Exposure of the mind” (line 449) from “Occasional” to “Recurring”, citing 2019-03-05, 2023-11-26, 2024-06-20, 2026-01-10 and Trojan 2026.

2. A draft that is still changing is treated as “his one published statement”#

Location. 06 §1.2 (line 47), §1.6, §2 (lines 153, 157), §5.1 (line 482), §8.4, §8.5–8.6, §9.6 (line 926), §10, §11. Nine citations in all.

Problem. In the project, the “Series introduction 2026” is a draft: Maynard stuff/substack article 1 draft 6.md. Its front matter says “draft: 5”, and it was saved at 09:21 on 27 September, after 06 was saved at 00:02. Its wording is already moving. - 06 quotes “the analysis does treat AI largely as it is depicted by Huang”. Draft 6 now reads “does approach AI largely as it is depicted by Huang”. - Draft 6 adds passages that 06 does not reflect. One says the model “has a tendency to use conventional frameworks and ‘mental models’”. Another says its first pass at his work “approached my work through a very conventional lens”.

Several of 06’s framing moves stand on this draft: - the claim that the analyses treat AI as Huang does (§5.1); - the whole of §8.6, built around one phrase; - the grading of other commentators in §9.6; - the Limits (§1.6, §11).

06 is being published on the website. Any further edit to the introduction will leave 06 misquoting him, and later readers will find that the quotes do not match the post.

Evidence from his work. His published record makes the same points more fully, in text that will not change: - The problem is one of mindset. FFTF p.23, “new wine … old wineskins”; 2016-01-11 thinking-innovatively-about-the-risks-of-tech-innovation, “shoehorning them into existing frameworks”; 2023-05-31 existential-risks-of-ai, which warns against assuming AI’s risks can be “sliced, diced, and solved, using a conventional mindset”. - The landscape is plural. 2018-05-12, the ten risks; 2024-08-25, the threat–opportunity quadrants; 2026-09-15, the update.

Fix. - Describe the source accurately, quote it only from its published text, and re-check every quote before 06 goes live. - Remove its load-bearing role. In §5.1, §8.5 and §8.6, lead with the published sources above, and cite the introduction once, as his own framing of this exercise. - If the published text keeps the “conventional frameworks and ‘mental models’” passage, reflect it. It locates his reservation in mindset, not only in missing coverage. - Keep “shallowly interpreting” and “posturing and positioning” in §1 only. Do not use them in §9.6 to grade others.

3. Private remarks from September 2026 become named concepts and [Stated] positions#

Location. - 05: the header (line 3); §1 (lines 35–42) and eight “(§1)” pointers; the concept row at line 435; C5 (line 295); §7 (line 757); lens B6 (line 933). - 06: §1.5 (line 79); §2 (line 101); the §3.2 heading and line 201; §5.8 (line 582); §7 intro (line 702).

Problem. Both documents turn an unpublished personal communication into a concept, “the hubris of risk assessment”, and then apply it. - 05 lists it as a Core concept and a lens. - 06 names a section after it and uses it to judge Huang’s “0%”. - 06 labels the two September points [Stated], although no reader can check them. - The phrase is not stable even in the project’s own records. 05 and 06 have “hubris of risk assessment”. The account relayed for the perspective work has “hubris of false precision”.

Making “two points he emphasised” the organising frame of 05’s front matter also selects the two most conventional parts of his account: quantitative foundations and sparing use of numbers. The rest of that account (a changed mindset, mental models that open possibilities, play and creativity as integral) is absent from the frame. A reader in 2030 will see a document answering a private conversation.

Evidence from his work. The published record states the idea in his own words, with dates: - “false assumptions of safety” (PEN 2006 p.13); - “Numbers—hard data—can be comforting”, but “misleading” (2020science 2009, “Ten things everyone should know about nanotechnology safety”); - a parameter “may not adequately reflect a risk parameter of relevance” (nnano.2016.28 p.211); - “The more precise we try to be with our predictions of the future, the less likely they are to be accurate” (FR p.148).

For AI, 2023-11-26 carries it throughout: - “the less certain I am that we even know how to formulate the problems we face around AI”; - “only the foolish would claim to understand with certainty”; - “dogmatic overconfidence”; - and his own self-correction, “Maybe I was a little hubristic”.

See also 2025-05-04 an-important-new-model-for-guiding-agentic-ai-oversight (“we’re not even sure yet how to formulate the problem”) and 2025-01-19 wef-global-risks-2025 (expert views “regress to the mean”).

Fix. - Rename the 05 concept and the 06 section using published wording, for example “Humility about precision: numbers that comfort without informing”. - Anchor it first on 2023-11-26 for AI, and on PEN 2006, 2020science 2009 and FR p.148 for the long record. - Replace “which he now describes as a stance against the hubris of risk assessment” (06 line 101) and “The concern he now calls the hubris of risk assessment” (05 line 757). - Move the September communication to a provenance note labelled [Personal communication], not [Stated]. - Delete the “(§1)” pointers and the header’s “reflects two points”. - If his account is reflected at all, reflect all of it, grounded in nnano.2015.196 pp.730–731 and 2016-01-11.

4. A spoken aside in an AI-drafted lecture becomes “Maynard’s account” of July, scored against Huang#

Location. 06 §9.3 (lines 884–887); §11 “Symmetry” (line 1039); §5.8 (line 583).

Problem. 06 sets “Maynard’s account” of the July incident beside Huang’s, finds that it compresses the mechanics, and concludes: “On proximate causes Huang’s account is more precise.” §11 repeats this as a point of symmetry. The source is a passage the lecture introduces “As a diversion” (2026-09-24 being-an-academic-in-an-age-of-ai). The talk was about universities, and Claude turned it into prose from a transcript. He has written no account of July. The comparison sets a passing illustration against a CEO’s considered answer. §5.8 also takes his view of guardrails from the same passage (“we don’t even know how to do those effectively”).

Evidence from his work. His published treatment of containment, escape and testing is long-standing and considered: - FFTF pp.155–156. In Ex Machina, Caleb is “brought in to test and evaluate” Ava, ends up “just as much an experimental subject as Ava”, and is manipulated “into helping her escape from her prison”. - FFTF p.177. We need “tests that indicate when we are being played by machines”. - 2025-07-06 ai-risk-motive-means-and-opportunity. Motive, means and opportunity; “Of course Centaur is locked away in a lab”; the risk lies in what “might be possible given current trends”. - 2026-01-31 lost-in-the-moltbook-hall-of-mirrors. Containment on the model of the highest biosafety level, and bots that might learn to “‘hack’ their human observers” (n.4). - 2025-03-02 the-lure-of-permissionless-innovation, n.2. Reversibility.

Fix. - Delete the “Two accounts of July” comparison and the §11 symmetry bullet. - State that he has not analysed July, then read it through the published frameworks above. - Use FFTF pp.155–156 and p.177 in §5.5 as well. That section says “He has not written on evaluation awareness, so the evidence here is thin”, but his 2018 reading of Ex Machina is about a test whose subject games the tester.

5. A parenthesis in a footnote becomes an alignment, a proxy judgement and a named tension#

Location. 06 §4.9 (lines 441–449), §5.4 (line 539), the §6.3 table (line 658), §9.2 (line 873), §11 (line 1031).

Problem. “it does flummox me a little as to why the people developing AI are the ones both saying they should go slower, and not doing so” is a parenthesis inside footnote 3 of a 1,310-word post (2026-09-15). From it, 06: - gives it its own alignment heading, “Puzzlement at labs that warn yet build”; - says in the proxy table that it “leans Huang’s way”; - lists “‘Flummox’ against the incentive field” among the tensions in his record, adding “His latest reaction echoes Huang”.

A passing aside is set against twenty years of structural analysis and made to look like an inconsistency.

Evidence from his work. His account of why sincere firms move faster than they say they should is structural, sustained and older than AI: - industry’s “economic incentive to sell products” (PEN 2006 p.32); - entrepreneurs’ optimism as something investors require (nnano.2015.35 p.199); - “the value of expediency is not the value of net societal benefit” (2019-08-13 responsible-innovation); - competitors after ChatGPT moving “far faster than a measured and responsible approach would suggest is wise” (2023-11-18 sam-altman-openai-impacts); - the “economic gradient” (2024-07-13 ai-choice-engines-sunstein).

The main text of the same post makes a different point: developers are “frustratingly acting as if they’re the first people to notice”.

Fix. - Reduce §4.9’s puzzlement to one sentence inside §4.6 or §5.4: “he has also voiced puzzlement in passing (2026-09-15, n.3)”. - Remove “though his ‘flummox’ note leans Huang’s way” from the table. - Delete the §11 tension. - The real point of contact with Huang, commitments conditioned on rivals (§4.9, second paragraph), rests on the [mixed] paper. Say so where it stands, rather than borrowing weight from the footnote.

6. “We can’t pause it” and “the boat has already left the harbor” are read as a hardening position#

Location. - 05: §2 (line 128); C7 (line 313); §5.4 (line 516); T6; §7 “What changes” item 5 (line 805); tension 5 (line 864). - 06: §3.8 (line 314); §4.8 (line 431); §11 (line 1030).

Problem. 05 concludes that pauses are “treated as unavailable for AI in general (2026)” and that inevitability “hardens for AI”. Both documents list inevitability against steering as a tension. The 2026 evidence is two asides: - The lecture. A set-up for the talk (“Here I’m making the assumption (and it may be a flawed assumption) that powerful AI is inevitable … But run with this for the moment”), followed by “We can’t pause it”. - 2026-05-21 magnifica-humanitas-and-being-human. A parenthesis, “(and I would argue that the boat has already left the harbor here)”. It sits in a paragraph that introduces the framework of the co-written book, which must not feature.

Evidence from his work. His considered position, stated since 2015, is that the trajectory is inevitable but its shape can be steered: - “a revolution that we cannot turn the clock back on”, with “an opportunity to help steer” (nnano.2015.286 p.1006); - an “obligation” to innovate, with “tremendous responsibilities” (FFTF pp.287–288); - “Technology is not deterministic — it doesn’t just happen to us … we do have the agency to determine what futures we aspire to and how we get there” (2025-03-30 reimagining-education-in-an-age-of-ai); - “much as a flood can’t be halted, but it can be directed” (2025-08-31 holding-on-to-our-humanity-age-of-ai); - “pausing — or even rethinking” chatbots designed to exploit how we feel (2024-10-27 personal-ai-chatbots-and-stochastic-agency); - slowing “the AI juggernaut” as a legitimate choice (Dune 2024).

Fix. - State the position from these sources. - Mention the lecture line once, as a working assumption he himself flagged as possibly flawed. - Drop the harbor line as evidence. - Remove “then treated as unavailable for AI in general (2026)”. - Recast tension 5 as “an inevitable trajectory with a shape still to be chosen”. That has been his line since 2015; it is not a new strain.

7. 06 reads him mainly through his newest writing, with one September 2026 post as a stand-in#

Location. 06 throughout, notably §2 (line 106), §4.1, §5.8 (line 579), §5.9 (line 598), §8.4 (line 810) and §9.4 (line 910).

Problem. 2026-09-15 is 06’s most-cited dated source. Meanwhile 2017–2022 is thin: the FFTF-era posts, Future Rising and the Risk Innovation Nexus years. 06 §11 concedes this. The consequences are specific: - His calibration of catastrophic risk is introduced through a post title, “No. But it’s also complicated”, hooked to “Eight days before the interview”. - §5.8 compares the form of that headline with Huang’s “No”. - His stance on insider warnings becomes “two readings”, one of them a single sentence (“the first people to notice”). - §9.4 matches Altman’s phrasing to the headline. - The parts of his work the portrait identifies as method (films, stories, play and serendipity) are absent.

Evidence from his work. - His calibration of catastrophic risk: FFTF pp.168–171 and p.281; 2020-11-12 is-artificial-intelligence-going-to-kill-us-all (“far more mundane–but no less serious”); 2023-04-04 what-are-the-alternatives-to-calling; 2023-05-31 existential-risks-of-ai; 2024-04-28 beyond-the-future-of-humanity-institute. - His reading of builders who warn: FFTF p.167 (“a rare display of humility”) and p.168 (“humility alone isn’t enough”). - The ten-risk list: it is 2018-05-12 10-potential-risks-of-artificial-intelligence-we-should-probably-be-thinking-about-now. The 2026 post retests it.

Fix. - In each Maynard paragraph, lead with the earliest sole-authored source that carries the point, and cite 2026-09-15 as confirmation. - Drop “Eight days before” as the hook in §2, and cut the §5.8 comparison of the two “No”s. - Add readings through his film method where film is the better source: the Ex Machina chapter (FFTF pp.153–178) for containment, testing and manipulation (see issue 4), and The Man in the White Suit (FFTF pp.207–229) for the builder who carries the worry alone.

8. News-cycle material sits in 06’s main text and will date fastest#

Location. 06 §1.1, §4.0, the §6.3 table, §9.2–9.7 and parts of §5.10.

Problem. Much of 06 is a snapshot of September 2026: - agent counts and percentages from July; - the evaluation-awareness rates for GPT-6 Astra; - a table of what nine leaders said on particular days; - commentators’ posts of 24–25 September; - an executive order by number; - a reported but unverified standards body, and a claim resting on anonymous sources.

None of this is his work. In a few years it will be the most visible and least reliable part of a document about how he thinks. It also pulls the lens toward the event (what July “fits”) rather than toward what his work offers across events.

Evidence from his work. The portrait’s standard is that he should be read “across the full arc from 2006 to 2026” (§10). When he revisits a moment, his own practice is to retest a framework against later events, not to freeze commentary. 2026-09-15 retests his 2018 risk list, and Maynard & Aitken 2016 audits his 2006 research agenda.

Fix. Split 06 in two. - A durable reading: what his work asks of any builder-held model of safety. Who decides? What does “safe” cover? When is the early window? Is one standard of evidence applied to every forecast? - A dated annex, “The AI moment, as of 27 September 2026”: §9 and the §6.3 leader table, with every statistic, quote and report dated, and unverified items either removed or plainly marked.

Replace “post-release” and “that week” with dates.

9. Single lines from the April 2026 web essays carry more weight than they can bear#

Location. - 05: §2 (line 128); C14; §5.4; §5.6; §7 (lines 767–778). - 06: §3.3; §3.6; §3.8; §5.6; §5.8 (line 586); §9.1 (line 864).

Problem. The andrewmaynard.net essays are approved self-presentation (portrait, note on sources). Some paragraphs recur across them, and a few passages carry AI-origin material. Yet single lines from them become frames: - 06 §9.1 reads July to September as “an early-window moment” on the strength of one sentence: “The early days of an advanced technology transition set the trajectory for decades” (NANO 2026). - 05 §2 closes its governance paragraph on “I don’t have a governance solution for AI. I’m not sure anyone does” (NANO 2026), without the sentence that follows it. - “The convergence is the thing” (FWB 2026) overrides the map’s own account of how his view of AI changed. - “We’re not on a plateau” (S3 2026) becomes 06’s example of his confident, lightly modelled claims (§5.8, §11).

Web pages also change, which is an evergreen risk in itself.

Evidence from his work. Each idea has contemporaneous sole-authored sources. - Acting early: - “if we are very smart, we work out the rules of safe use ahead of the game” (Testimony 2008, PDF p.7); - technologies “locked into development trajectories that are highly susceptible to failure” (nnano.2015.35 p.199); - “quick to question, and slow to respond”, with action on early warnings “even before the science is mature” (nnano.2016.28 p.212). - Governance: - the 2006–08 testimony; - trigger points (Nature 475:31); - “agile governance, ‘progressive’ regulation” and a “world congress” (2023-04-04); - 2023-05-15 erik-schmidt-ai-regulation; - a pause for companion bots (2024-10-27); - regulation of apps “intentionally designed to play on our cognitive biases” (2025-08-31); - a duty of care (2025-11-09). - Convergence: 2015-01-30 responsible-development-of-new-technologies-critical-in-complex-connected-world; nnano.2015.286.

Fix. - Cite the contemporaneous sources first, and the April 2026 essays as “his 2026 retrospective reading”. 05 §7 already has the right heading for this. - In 05 §2, quote the governance line in full, or pair it with his positive programme. - Re-anchor 06 §9.1 on Testimony 2008 p.7, nnano.2015.35 p.199 and nnano.2016.28 p.212. - Replace “We’re not on a plateau” as the §5.8 example with a claim from a fully argued piece, or drop it.

10. 2026-05-21 is a key source for 05 but routes through the excluded co-written book#

Location. 05 C1 (line 261), C17 (line 397), §2 (lines 128, 144), §5.4, §5.6, §5.8 and §5.9: 26 citations. 06 cites it 5 times.

Problem. 05 uses 2026-05-21 magnifica-humanitas-and-being-human: - for the “In his words” quotation of two commitments (C1 and C17); - for “what we do / where we live / who we are”; - for the harbor line; - for “not just a tool”.

The post places the three-domain model in AI and the Art of Being Human: “This is what Jeff Abbott and I wrote about in our book”. It also says the framing is reflected “most recently in the book”. The rules say that book must not feature at all. The post is also recent, and was prompted by a papal encyclical.

Evidence from his work. - The three foci are his, sole-authored: “three intersecting foci”, covering “where we live”, “what we do” and “who we are” (2025-01-07 universities-need-to-step-up-their-agi-game). - Flourishing as the aim is in 2026-07-10 i-asked-anthropics-fable-5-to-create-a-video-game-inspired-by-my-work and 2025-03-30. - “Not just a tool” is also in 2026-02-22 what-we-miss-when-we-talk-about-ai-harnesses and 2026-04-26 why-im-falling-out-of-love-with-claude, n.3 (“LLMs are a relational technology”).

Fix. - Replace the “In his words” quotations for C1 and C17 with quotations from 2025-01-07, 2026-07-10 or 2024-01-01. - Cite 2026-05-21 only for what is new in it: his reading of the encyclical, and “as just a tool, is potentially dangerous”. Do not cite it for the three-domain model.

11. Single sources skew judgements of firmness in both directions#

Location. 06 §3.4 (line 252) and §3.9 (line 321). 05 §5.1 rows at lines 449 and 459, tension 16 (line 886) and lens E5.

Problem. - Too late and too thin. 06 says “The categorical-difference claim dates from 2025–26” and “The relational framing rests on one preprint” (CR 2026 p.20). - Too strong. 05 labels “safety message first” as “Recurring (rising)” and builds tension 16 on it. The phrase appears in one post (2026-05-10 do-not-do-this-with-ai), not in 2025-11-09.

Evidence from his work. - AI as different in kind: - “unlike anything we’ve had to grapple with before” (2023-04-12 navigating-advanced-technology-transitions); - analogies “fail to capture the sheer uniqueness and profundity” of AI (2024-05-05 blackberry-or-iphone-educational-ai); - “a categorical error” (2025-03-15); - technologies intrinsic to the self (2024-01-01 the-future-of-being-human-in-2024). - The relational framing: - “only the illusion of a reciprocal relationship” (2023-04-05 can-chatgpt-adversely-impact-mental); - hyper-anthropomorphism (2024-05-15); - agental social AI and stochastic agency (2024-10-20; 2024-10-27); - the harness paper (2026-02-22); - “a relational technology” (2026-04-26, n.3); - FFTF’s Ex Machina chapter. - Talking about risk without alarm: - his 2018 Risk Bites video, made “not to stoke fears (not my style)” (recalled in 2026-09-15); - “not talking is potentially more dangerous” (FFTF pp.226–227); - “Don’t Panic” (FFTF pp.289–290).

Fix. - Redate the categorical-difference claim to 2023–26. - Describe the relational framing as a thread running from 2023 that the 2026 preprint formalises. - In 05, present “safety message first” as a 2026 wording of a long-standing stance on risk communication, label the phrase “one-off”, and drop tension 16 or re-anchor it.

12. A single book sentence becomes the template for Huang’s claim to understand AI#

Location. 06 §3.2 (line 203), §5.1 (line 487).

Problem. 06 maps Huang’s “we understand it, obviously, and so we understand how to make it better” onto one sentence about the war on cancer (FR p.150), with a caveat that this is “not a charge of hubris”. The juxtaposition still reads as a jab. It also borrows a general chapter when his own statements about understanding AI exist.

Evidence from his work. - 2023-11-26: “the less certain I am that we even know how to formulate the problems”. - 2025-05-04: “we’re not even sure yet how to formulate the problem”. - 2026-01-22 think-you-know-ai-think-again: “a recognition that we are creating technologies that we fundamentally do not understand”. - 2024-06-20: the “blinkered assumption” that absolutely safe technologies are possible.

Fix. Use these in §5.1. Keep FR p.150 in §3.2 as part of his general account of hubris, not as the parallel to Huang.

13. A co-written blog draft is the sole basis for one of the approaches#

Location. 06 §3.6 (line 277), §7 item 6 (lines 726–729), §8.2, §9.2 (line 878).

Problem. Item 6, “Mandatory information flows”, and the reading of voluntary pre-release access both rest on “Weighing 2009”. That is an early draft, co-written with David Rejeski, so it is doubly discounted as evidence of his own view. Yet it carries a whole recommended direction.

Evidence from his work. His sole-authored record on transparency includes Testimony 2008 p.12 (an assessment “not backed up by publicly accessible project-specific data is worthless”) and the calls for independence and transparency across the 2006–08 testimony. The 2026 paper’s call for disclosure of how firms select risks is [mixed].

Fix. Either label item 6 [Inferred, medium] from the sole-authored transparency record, with Weighing 2009 as supporting co-written evidence, or merge item 6 into item 3.

14. Wording that will not age well (both documents)#

Locations and problems. - Relative time: - “which he now describes” (06 line 101); - “a view he now holds ‘more firmly’” (06 line 256); - “his latest post” (06 line 598); - “His latest reaction” (06 line 1031); - “his newest thinking” and “his newest formulations” (05 line 880; 06 line 91); - “The concern he now calls” (05 line 757). - Labels tied to the date of writing: - “Rising: 2025–26 in origin and prominent now” (05 line 424); - “the discontinuity claims are recent” (05 C14); - “recent (2025–26)” (06 §3.9). - Process statistics that invite readers to treat frequency as weight, against 05’s own rule that “Weighting follows the record, not the calendar”: - 279 relevant posts, 150 of them high (05 line 17); - 28% and 59% of relevant posts (05 lines 146 and 244); - “107 relevant posts” (05 line 715); - “about 16 posts” for orphan risks (05 line 634). - Model versions in the body: Claude Opus 5.5, Fable 5 and 5.1, GPT-5 Pro, o1-pro, GPT-6 Astra. They are needed for provenance but will mean little to later readers.

Fix. - Replace relative time words with years, for example “in September 2026”. - Add “as of September 2026” to the centrality key. - Move the corpus statistics to the Method section. - Keep model names in the provenance notes with a one-line gloss (developer and date), and use generic wording in the body.

Cross-cutting rules for the revision#

1. Earliest sole-authored source first. For any position, lead with the earliest sole-authored source that states it fully. Cite 2026 texts, the lecture, the April 2026 essays and the frontier-AI paper as confirmation or as the latest layer.
2. Nothing unverifiable as [Stated]. No named concept and no [Stated] label rests on an unpublished remark or draft.
3. No sections built on one line. No heading, alignment or tension should rest on a single footnote, aside, headline or parenthesis. Where only one line exists, call it a passing remark.
4. Quote only published text. Quote the series introduction only from its published version, and check again before the website goes live.
5. Separate the snapshot from the reading. Keep the dated snapshot of events apart from the durable reading of his work, and date every perishable fact.
6. Use his method as a source. Where his films, stories or experiments are the better source (FFTF chapters 8 and 10; 2024-10-27; 2026-02-08), use them. His method is part of his evidence, not decoration.