Late Lessons, Jensen Huang and AI
Markdown version

Fidelity review of 06: attributions to Maynard#

Working file. Reviews 06-huang-and-late-lessons-through-maynard.md (890 lines, 26 September 2026 text) for fidelity to Maynard’s own texts: quotations, citations, pages, labels, provenance and proportion. Line numbers (L) refer to 06.

Method#

Overall verdict#

Quotation accuracy is high. No quotation is invented. Wording matches the sources, and nearly all pages are right. The problems are of four kinds: 1. Context dropped in ways that move Maynard towards Huang (items 1, 4, 5, 6) or understate his position against builder-held decisions (item 3). 2. Other people’s views, or arguments he set out, attributed to him (item 2). 3. Miscitations (items 7, 8). 4. Labels calibrated too high, mostly in In brief (items 9, 10).

Two of these (items 1 and 2) were flagged in the M3, M7 and M9 fidelity checks and have come back during integration.

Ranked issues#

HIGH#

1. “Apparent agency is ‘illusory’” misstates the source and contradicts his stochastic-agency concept. - Location. §4.6, L373 (Maynard paragraph under “Sincere builders, harm without intent, and no machine will”). - 06 says. “Apparent agency is ‘illusory’ (2026-01-31).” - Source. 2026-01-31: “there are already whispers of an exponential surge toward AI self-awareness … I must confess that I am skeptical of this. Much of what we’re seeing is, I suspect, illusory, as it’s rooted in the unique abilities of large language model-based AI’s to emulate very human behavior while not being in any sense self-aware. That said, there are very real risks here, as bots learn from each other how to exploit vulnerabilities…” - The object is apparent self-awareness, not agency. - The claim is hedged (“Much of”, “I suspect”). - Maynard affirms functional agency elsewhere: “stochastic agency” (2024-10-27), and agentic “organoids” that “can wreak havoc as if they are [alive]” (same post, quoted correctly at L215). - Why it matters. This sentence is the main support for aligning him with Huang’s “no willpower… just electrical power”. It makes him more absolute, and closer to Huang, than he is. The M3 and M7 checks flagged the same flattening. - Fix. “He suspects that apparent self-awareness is largely ‘illusory’ (2026-01-31), while treating functional, ‘stochastic’ agency as real (2024-10-27).” Move stochastic agency into the Limits list below it.

2. Two views on open weights attributed to “his record” are others’ views that he reported. - Location. §9.2 table, L749 (“two strands of his record pull apart”); §7 item 8, L638 (2023-07-12 cited as the basis for reversibility). - Source, first strand. 2025-07-23: open-weight support is “something that will be welcomed by many who worry about the corporate control and associated vulnerabilities associated with powerful closed models”. He reports other people’s worry. He adds that researchers will applaud it, and notes “strings attached”. - Source, second strand. 2023-07-12: “If this is the case for frontier AI models, then open sourcing these models could lead to capabilities that, once out of the bag, are very hard to put back in. Of course, whether you think that this is a good thing or a bad thing depends on your philosophy.” This sets out the Frontier AI Regulation paper’s argument against Howard’s. He remains even-handed and concludes that “we’re going to have to hash this out together”. - Why it matters. The row presents a tension in Maynard’s thinking that his texts do not show. The M9 check flagged the 2025-07-23 misattribution. - Fix for L749. “His record on open weights is thin. He noted that open-weight support ‘will be welcomed by many who worry about the corporate control’ of closed models (2025-07-23). In 2023 he set out, without adopting, the argument that open-sourced capabilities are ‘very hard to put back in’, and treated the choice as depending on ‘your philosophy’ (2023-07-12).” - Fix for L638. Drop 2023-07-12 from the basis; 2025-03-02 n.2 carries the reversibility principle. Label as [Inferred] for the open-weights application.

3. “I don’t have a governance solution for AI” is quoted without the sentence that follows it, and that sentence bears directly on Huang. - Location. §3.6, L258 (“His confidence in remedies has fallen”); §7 introduction, L603; related claim at §8.4, L705. - Source. NANO 2026: “I don’t have a governance solution for AI. I’m not sure anyone does. But I do think the nanotech experience offers something valuable: evidence that inclusive, transdisciplinary governance — however messy and slow — produces better outcomes than leaving decisions to the people who happen to be building the technology.” The next paragraph ends: “‘together’ means considerably more people than are currently in the room.” - Why it matters. - The omitted sentence is his clearest recent statement against a builder-decides model, which is exactly the divergence in §5.3. - Without it, “confidence in remedies has fallen” reads as more pessimistic and less directional than he is. - §8.4 (L705) then says his case for engagement “rests on process, not measured outcomes”. In fact he claims outcome evidence, even though he does not measure it. - Fix. - Quote both sentences at L258 and L603. - At L705: “He claims ‘evidence’ that inclusive governance ‘produces better outcomes’ (NANO 2026), but does not present measured outcomes; 01 rates participation’s outcome benefit as suggestive (G6).”

4. The coordinated-pause alignment omits his call, in the same post, for urgent coordinated action. - Location. §4.8, L399–403; In brief, L110 (“doubt that a coordinated general pause would work”); tension “Inevitability against steering”, L843. - Source. In 2023-04-04, the post in which he declines the pause letter, he also: - calls for “agile governance, ‘progressive’ regulation … And we need this, like, yesterday!”; - proposes “a world congress” to produce “implementable consensus guidelines and a roadmap”, with public–private partnership; - concludes: “the risks of LLMs … are potential[ly] catastrophic … And yet the biggest risk is not taking action or, worse, assuming no action is needed.” - Why it matters. “He sits nearer Huang than the pacing advocates” is defensible only for the instrument. His 2023 alternative was more coordinated, faster collective action, which is further from Huang’s “Apply [existing law]” than the section conveys. That breaches clarification (1) on absoluteness in the other direction. - Fix. - Add to Limits: “His objection was to the instrument, not to coordination. In the same post he called for an international convening and ‘progressive’ regulation ‘like, yesterday’, and wrote that ‘the biggest risk is not taking action’ (2023-04-04).” - In In brief, write “doubt that a general training pause is the right instrument (partial; section 4.8)”.

MEDIUM-HIGH#

5. Hedges on the tail are dropped where 06 aligns him with Huang’s anti-doom stance. - Location. §4.1, L320; In brief, L113; §4.1 Limits, L325. - (a) FFTF p.281. 06 quotes “more an act of faith than of reason” for superintelligence scenarios. The same paragraph says: - “It’s not that gray goo or superintelligence don’t have some probability of occurring”; - Occam’s Razor gives “a lower probability … but not a zero probability. And this leaves the door open to more complex, more fanciful possibilities being plausible”. Placed beside Huang’s “0%”, the truncation hides the very contrast 06 draws in §5.8. - (b) 2023-05-31. “Catastrophic risk … ‘not so small’” is used as a limit on alignment with Huang’s tail view. But Maynard defines catastrophic risk broadly: events “where large numbers of people risk losing something that is deeply valuable to them — whether this is their life, their health, their dignity, autonomy, purpose…”. It is not comparable with Huang’s “end of the world”. - Fix. - Add at L320: “…though, he adds, ‘not a zero probability’, which ‘leaves the door open’ (FFTF p.281).” - At L113 and L325, gloss his definition: “catastrophic risk, which he defines broadly as large-scale loss of what people value, including dignity and autonomy, is ‘not so small’”.

6. “Huang’s almost word for word” overstates the containment convergence and sits under a [Stated] label. - Location. §4.4, L352; repeated in §11 Symmetry, L853 (“Maynard’s 2025 reasoning about containment matches Huang’s“). - Source. 2025-07-06 goes on: “At the moment, it feels that ‘opportunity’ is the weakest part of the link… But the risk here isn’t what is currently possible, but what might be possible given current trends”, as agents gain “autonomous write … access”, and “the opportunity … will only increase”. - Why it matters. For Maynard, containment is a temporary condition that trends are eroding. For Huang, it is the remedy. That is a real difference, and it is hidden. The comparison itself is the report’s reading, not his statement. - Fix. - Relabel the comparison [Inferred, medium]. - Replace “almost word for word” with “shares Huang’s premise”. - Add: “but he treats lab containment as temporary: ‘the risk here isn’t what is currently possible, but what might be possible given current trends’ (2025-07-06).” - At L853, write “overlaps with” in place of “matches”.

MEDIUM#

7. The “careers” claim is miscited, and its co-authorship is not flagged. - Location. §6.4, L594: “Careers built around assumed risk carry interests too, as Maynard said of his own field (NN 2014-03).” - Source. NN 2014-03 does not mention careers or interests; it speaks of a “rut” and of researchers digging for “assumed-to-exist evidence of novel risk”. The careers point is in Maynard & Aitken 2016 p.999 (co-written): “as careers and funding pathways are built around assumptions of substantial nanomaterial-specific risk, it will become increasingly difficult to make…”. - Fix. Cite “Maynard & Aitken 2016 p.999, co-written; cf. NN 2014-03 p.160”. Consider “as he and Aitken warned of their own field” in place of “said”.

8. A 2026 [mixed] quotation is cited to 2019. - Location. §8.3, L694: “show them ‘a threat to something it values’, alongside the ‘crude boundaries’ of rules (2019-08-13).” - Source. “you show it a threat to something it values” and “mission-driven” occur only in 2026-07-16 [mixed]. That paper presents them as “the lesson that has stayed with me” from the Garbee work. The 2019 text has “crude boundaries” and the mutual-worth framework, not this wording. - Fix. “show them ‘a threat to something it values’ (2026-07-16 [mixed], recalling the Garbee work), alongside the ‘crude boundaries’ of rules (2019-08-13)”. §3.7 (L272) already does this correctly.

9. In brief states some items more strongly, and labels them higher, than the body supports. - Location. L102 (header “[Implied, from stated positions]”) and L104–111; L126. - Detail. - The list inherits [Implied], but the body gives [Inferred, medium] for the pause (§4.8), [Inferred, medium-high] for students and purpose (§4.10), and [Inferred, low-medium] for liability (§4.7). - “Scepticism of incumbents seeking antitrust relief” (L109) goes beyond the source. In 2023 he raised the “cynical move” possibility and set it aside (“I don’t think this is happening — at least, not yet”; Altman “comes across as genuinely sincere”, 2023-05-17). He has not commented on the waiver. §4.7 itself says only that his work “would treat the waiver as a legitimate object of scrutiny”. - “Rejection of doom built on … eminence” (L104) needs a qualifier. He treats eminence as a reason to engage the reasoning: of Bengio, “he knows his stuff. And so when he writes about … ‘rogue AI’ it’s worth paying attention” (2023-05-25). - “The refusal of liability relief” (L126) appears among values without the low-medium confidence of §4.7. - Fix. - Change the header to “[Implied or Inferred; section 4]”. - Change L109 to “willingness to scrutinise whether incumbents’ calls for rules serve incumbents”. - Change L104 to “rejection of doom built on extrapolation, and of eminence as a substitute for reasoning”. - Mark the liability item “(low-medium)”.

10. Several labels are set too high, and one claim carries two different labels. - §4.6, L376. “[Implied, high] On motive his work sits nearer Huang than Klein’s ‘the profit motive… the desire to cut corners to be first’.” His structural record overlaps Klein’s “cut corners to be first”: - “sheer fantasy” that good intentions suffice (2023-12-15); - “the value of expediency is not the value of net societal benefit” (2019-08-13); - competitors moving “far faster than a measured and responsible approach would suggest is wise” (2023-11-18); - “go fast and bugger the consequences” (2025-07-23). Fix: [Inferred, medium]. “Nearer Huang on personal motive; nearer Klein on structural pressure.” - §5.1, L434. “[Implied, high] He disagrees with that depiction itself.” The source says “I’m not sure I fully agree with” the assessment. Fix: “[Implied, medium-high] He regards that depiction as incomplete (see 2026-01-22; CR 2026).” This fits clarification (1). - §5.8, L525. “[Implied, high]” applies a 2019 claim about entrepreneurs’ good intentions and codified approaches to the evidential standard for “I know they know how to fix it”. That is a structural transfer. Fix: [Inferred, medium]. - Same claim, two labels. The control-banding reading of “two out of three rights” is “[Inferred, medium] … conceptual transfer” at §4.5 (L366) and “[Implied, high]” at §6.1 item 3 (L553). Fix: use [Inferred, medium] in both places. - Labels in the wrong scope. - §4.3, L344: “[Implied, medium-high]” is attached to a claim about Huang’s meaning. Unlabel it or cite 02. - §5.7, L515: “[Stated] He is closer to Huang on cars than the contrast suggests” is an inference. Move [Stated] to the Waymo quotation. - Label too low. §4.2, L336 gives [Inferred, medium] to the claim that acting on implausible forecasts has costs. That is Stated at FFTF pp.205–206 (“if investors and consumers avoid certain technologies because they’ve bought into a narrative that belongs more in science fiction…”). Only the radiology application is inferred.

11. The provenance of the April 2026 essays is not disclosed, although wording from them carries a key claim. - Location. §1.4, L62–67 (only three [mixed] items listed); §3.1, L166; also L186, L203–204, L258, L284, L811 (30Y, NANO, HNS, FWB and S3 are cited about 25 times). - Source. S6 finds the series is his own ideas, published and endorsed, but: - paragraphs are duplicated across essays; - four passages present AI-generated or others’ material as his; - “Exact wording and a few attributions are less secure”. - Specific problem. L166 rests the claim that “his own summaries sometimes sound more absolute” on 30Y wording. It also truncates that wording: the full phrase is “from something to be minimized to something to be navigated creatively in pursuit of value”, and the next sentence is “I’m skeptical of both the safety absolutists and the move-fast-and-break-things crowd”. That sentence is not absolute. - Fix. - Add a §1.4 bullet: “The April 2026 andrewmaynard.net essays (30Y, NANO, HNS, FWB, S3) are his self-account under his byline; their ideas are weighted as his, but exact wording is less secure (repeated paragraphs, a few AI-origin passages). None of the passages quoted here comes from those passages.” - At L166, quote the full phrase or soften to “a 2026 essay’s shorthand”.

12. The series introduction is cited as published, but the source is a draft. - Location. L47 (“The Future of Being Human, 27 September 2026”). It is quoted at L88, L143, L434, L702, L709, L781 and L827. - Source. Maynard stuff/substack article 1 draft 3.md (“draft 3”, with “[add url]” placeholders and typos). All quotations match draft 3. - Fix. Before release, re-check every quotation against the published post. Keep the date only if publication happens as planned.

13. The footnote on hacking human observers is misread. - Location. §9.3, L761: “agents acting through people, ‘learning to “hack” their human observers’, who would then be ‘already beyond being contained’ (2026-01-31, n.4) [Stated]”. - Source. “the possibility of bots on Moltbook learning to ‘hack’ their human observers… And here, they are already beyond being contained.” - The subject is the bots, not the observers. - The tense is present (“already”), not conditional. - The claim is about Moltbook specifically. - “His most distinctive concern” is the report’s characterisation, not his. - Fix. “…the possibility of Moltbook bots ‘learning to “hack” their human observers’, where, he writes, ‘they are already beyond being contained’ (2026-01-31, n.4) [Stated]; that this is a distinctive concern is [Implied].”

LOW#

14. Page and slug errors. - “exhilarating” is on FFTF p.166, not p.163 (L274). - “systemic instabilities” is on NN 2015-12 p.1006, not p.1005 (L698). - Four slugs are truncated and will give broken mirror links: - L171 life-on-mars-astrobiology should be …-and-thinking-differently-about-risk; - L279 responsible-development-of-new-technologies should be …-critical-in-complex-connected-world; - L291 bounded-infinities should be …-quantum-tunneling-and-the-future-of-education; - L670 if-elon-musk-is-a-luddite should be …-count-me-in. - L83 says keys follow 05 Appendix C, but “Weighing 2009” (L253, L629, L686, L748, L875) is not in it. Define it in 06 as Maynard and Rejeski’s 2009 draft of a Nature commentary (co-written), which he posted on 2020science on 8 July 2009, or add it to 05.

15. Quotations clipped or softened; each moves the reading slightly. - L567: “Severity floors make sense as triage” drops “But they also carry hidden consequences” (2026-07-16). Add it. - L568: OpenAI “was fast to admit” failure. The source says it was fast to admit “a potential risk” (2025-08-31). Change “failure” to “the risk”. - L373: “trying hard” drops “they still lack the breadth of vision and understanding that’s necessary to succeed here” (2025-01-07). Add it. - L686: “ethics boards as ‘smoke-and-mirrors’”. Source: “At worst, ethics advisory boards can easily become a smoke-and-mirrors attempt”. He also says they have “a valuable role to play”. Add “at worst”. - L570: the contrast in Harness 2026 p.5 is with Anthropic’s Constitutional AI approach (Anthropic 2022), not the 2026 constitution. Change to “Anthropic’s Constitutional AI approach”. - L747: “his 2023 worry about rules favouring ‘first movers’”. He raised it and judged it was not happening, “at least, not yet”. Change to “the possibility he raised in 2023 (and then set aside, ‘not yet’)…”. - L251: “Industry cannot lead risk research”. PEN 2006 p.32 says “one would not expect them to conduct more general and basic research” and that findings “might be considered suspect” without independent support. Change to “should not be relied on to lead strategic risk research”. - L251: Testimony 2007 p.30 is a weak citation for the general promoter–overseer principle (it says the NNI is “not an ideal organization… more attuned to stimulating… than providing science in support of oversight”). Add Testimony 2008 p.8: safe nanotechnologies will not emerge “if the promoters of the technology are calling all the shots”. - L722: “for AI it ‘may be linear. It may have a threshold’”. In 2023-11-26 the sentence describes the transforming function in general. Change to “the transforming function ‘may be linear. It may have a threshold’”. - L403: “the broad-remit federal body he proposed”. The source (2023-05-17) “underscore[s] the need for something at the Federal level”. Change to “the broad-remit federal capacity he called for”.

16. Co-authorship and [mixed] flags are applied unevenly. - These co-authored items are unflagged where cited: Toxicol. Sci. 2011 (3 authors; L186, L197), Nature 2006 (14 authors; L495, L621) and ILSI 2005 (14 authors; L696). This goes against L68 (“treated as shared positions and flagged”). - The [mixed] tag is applied to Trojan 2026 passages beyond the four mechanisms (p.11 boundary conditions, L452; p.14 research directions, L454 and L624; p.12, L722) without saying so. §11 (L834) then calls them “the Trojan mechanisms”. - Fix. Either widen the §1.4 Trojan bullet (“and, conservatively, its boundary-condition and research-direction passages”) or drop [mixed] from those passages. Correct L834 either way.

Verification log (88 items checked in context)#

Legend: OK = wording, page and meaning verified; OK* = verified, with the caveat in the issue cited.

# 06 location Attribution Source checked Verdict
1 L88, L827 “naive and misguided”; “shallowly interpreting…” Series intro draft 3 OK* (12)
2 L143, L709 “rigor and balance”, “not sure I fully agree”, “broader landscape”, “depicted by Huang”, “struggled…toxicology” Series intro draft 3 OK* (12)
3 L702 “a particular frame to bear” Series intro draft 3 OK
4 L160 control banding “decision-making based on incomplete information”; not “a substitute” AOH 2007 p.10 OK
5 L702 “highly hazardous until proven otherwise” AOH 2007 p.10 OK
6 L161 “the risk assessment paradigm remains relevant” Toxicol. Sci. 2011 OK
7 L186 plausibility “a crude but effective filter” Toxicol. Sci. 2011 OK* (16)
8 L162, L164 “Important as evidence-based…”; “existing or future ‘value’” NN 2015-09 pp.730, 731 OK
9 L683 “important and implementable” NN 2015-09 p.731 OK
10 L164, L199 “a useful starting point”; “not just chemicals” NN 2016-03 p.211 OK
11 L501 “quick to question, and slow to respond”; “even before the science is mature” NN 2016-03 p.212 OK
12 L164 “dignity, belonging, identity, belief…” FFTF p.23 OK
13 L164 “extends conventional thinking rather than replacing it” 2018-12-13 OK
14 L166 “minimized… navigated” 30Y 2026 OK* (11)
15 L166 “not just about technical hazards to be minimized” NANO 2026 OK
16 L169–170 “always… social and political”; “social construct”; “blinkered assumption”; “standards and policy” 2024-06-20 OK
17 L724 “infinitely more complex… chemicals and biological substances” 2024-06-20 OK
18 L171, L702 “bear the brunt”; precaution for “complex, uncertain, and potentially catastrophic risks” 2020-07-30 OK
19 L174 “through fear and uncertainty… severe blow” Testimony 2006 p.52 OK
20 L174, L521 “zero exposure”; “brave person”; “dogmatic overconfidence” 2023-11-26 OK
21 L721–722, L626 exposure “hints of ideas”; “gnarly”; “beginnings of a framework” 2023-11-26 addendum OK* (15)
22 L175, L181 “blatant disregard”; “as-yet-to-be-discovered risk” NN 2014-03 p.160 OK
23 L181, L684, L251 “false assumptions of safety”; “appropriate surveillance”; “economic incentive to sell products” PEN 2006 pp.13, 14, 32 OK* (15)
24 L181, L254, L551, L685 “methodology for strategy”; “good intentions are not enough”; “science in the service of safety”; “no substitute for progress” Testimony 2007 pp.21, 16, 9, 16 OK
25 L342 10% of federal nano R&D Testimony 2007 pp.4, 14 OK
26 L181, L183 “Numbers—hard data…”; “When the data run out – innovate!” 2020science 2009 OK
27 L181, L683 “The harder challenge is working out what we should be measuring” NN 2015-06 p.483 OK
28 L181, L552, L676 “may not be as high”; “process of science is working”; controls “in the main… work effectively” Maynard & Aitken 2016 p.999 OK
29 L183, L342, L501, L462, L615 $68m to $13m; “need not be, nor should be”; “very smart”; “on their own”; “worthless” Testimony 2008 pp.12, 5, 7, 11, 12 OK
30 L183, L610 trigger points “must be flexible” Nature 2011 OK
31 L555 “I have changed my mind” Nature 2011 OK
32 L197 “not by the technological labels…” Nature 2011 draft OK
33 L183, L236, L645 “explorations, not findings”; “before the answers arrive”; SCOPUS seven; “before it locks in” HNS 2026 OK
34 L186 “when make-believe is treated as plausible reality” FFTF p.205 OK
35 L334 speculation steers investors and consumers away FFTF pp.205–206 OK* (10)
36 L186, L203, L232, L811 “massively amplifies uncertainties”; “technology had changed dramatically”; “more firmly”; “closer to the bone” FWB 2026 OK
37 L187, L113 “vanishingly small”; catastrophic “not so small” 2023-05-31 OK* (5)
38 L188, L399 “(I do)”; “have the intended effect” 2023-04-04 OK* (4)
39 L256 “there are no silver bullets” 2023-04-04 OK
40 L102, L113, L238, L320, L352, L409, L521, L536, L540 title; ten-risk list; “risen in significance”; “devoid of details”; “more likely… informed eye”; n.1; n.3 “flummox”; n.5; “freaking out”; “first people to notice” 2026-09-15 OK
41 L189 AI 2027 “speculation — no more”; “sliver of truth” 2025-04-06 OK
42 L190, L216, L248, L288, L401, L756, L818 “informed speculation”; “might happen… irrelevant”; “critically important… ‘Solve it for us’”; “can’t pause it… flawed assumption”; “somebody else will”; July account; “potential is profound” 2026-09-24 [mixed] OK
43 L195, L495 fibre-shaped particles and asbestos; records “interpreted in the light of new knowledge” Nature 2006 p.268 OK* (16)
44 L196, L721 “directly applicable”; “not a chemical”; exposure to algorithms 2019-03-05 OK
45 L196 “is not directly applicable… But the concept is.” AOH 2007 p.10 OK
46 L199 “a rather unreliable indicator” NN 2014-06 p.410 OK
47 L199 “cloud-based AI”; “not remotely the right shape” 2016-01-11 OK
48 L201, L212, L221 “defy the analogies”; “calculators on steroids”; “Rather, they are different”; constitution quote 2026-01-22 OK
49 L202, L232, L433, L228 “substantial scaling”; “temporal cognitive processes”; “If AI is a tool…”; “just a tool” CR 2026 pp.2, 3, 2, 20 OK
50 L208, L258 “foreshortening”; “who decides who will suffer” 2023-10-19 OK
51 L212 “AI wasn’t even on my radar” (2008) FFTF p.168 OK
52 L215 Moltbook quotes 2026-01-31 OK
53 L373 “Apparent agency is ‘illusory’” 2026-01-31 Error (1)
54 L352 BSL-4 containment 2026-01-31 OK
55 L761 “hack their human observers” 2026-01-31 n.4 Misread (13)
56 L217, L373, L352, L760 “a reason for doing something”; “outweighs ethical considerations”; Centaur; “weakest part of the link” 2025-07-06 OK* (6)
57 L220, L454 “cannot comprehend”; mastery question open 2026-04-11 OK
58 L222, L240 “just a tool, is potentially dangerous”; “shaking things up” 2026-05-21 OK
59 L224, L240, L536 “working with a machine”; “illusion of learning”; “safety message first” 2026-05-10 OK
60 L226, L448, L570 Harness pp.1, 3, 5, 8 Harness 2026 OK* (15)
61 L234, L236, L452, L454, L688, L722 Trojan pp.1, 2, 3, 11, 12, 14 Trojan 2026 OK* (16)
62 L232 “prolonged interactions with intelligent machine[s]” 2020science 2014 OK
63 L232 “far more plausible, and far scarier” FFTF p.159 OK
64 L232, L289, L373 “stochastic agency”; “pausing — or even rethinking”; n.2 2024-10-27 OK
65 L232 “economic gradient”; “may not be intentional or even malicious” 2024-07-13 OK
66 L240, L684 “duty of care”; “very small tip” 2025-11-09 OK
67 L245, L246, L262, L221 “abdication of responsibility”; “complete autonomy”; myopic benevolence; “genuine commitment” FFTF pp.288, 227, 218–227, 221 OK
68 L247, L797 “can’t get AI governance right on its own!”; “everyone has the right” 2023-05-15 OK
69 L253 “reticence of industry” Weighing 2009 (co-written) OK* (14)
70 L256, L503 deepfake letter (signed); “agile and responsive” 2024-02-25 OK
71 L256, L454, L745 “intentionally designed…”; “safety checks”; “better-managed, but probably not eliminated” 2025-08-31 OK
72 L568 “was fast to admit” 2025-08-31 Clipped (15)
73 L256, L389, L399 “at this point”; “cynical move… not yet”; agency doubts 2023-05-17 OK* (9, 15)
74 L638, L749 “very hard to put back in” 2023-07-12 Misattributed (2)
75 L749 “corporate control” 2025-07-23 Misattributed (2)
76 L258, L603 “I don’t have a governance solution…” NANO 2026 Truncated (3)
77 L268, L272, L477 “expediency”; “remain good intentions”; “crude boundaries”; “loosely coupled”; “death knell” 2019-08-13 OK
78 L694 “a threat to something it values” as 2019 2019-08-13 / 2026-07-16 Miscited (8)
79 L270, L482, L562, L567, L603, L661, L649 incentive field; “land on every organization”; “not equally open”; “surprisingly diligent”; “augmentation”; “yet to be shown”; “past 2028” 2026-07-16 [mixed] OK* (15)
80 L264, L274 “isn’t necessarily reckless”; “single innovator”; “kindling”; “rare display of humility”; “humility alone isn’t enough” FFTF pp.162, 167, 168 OK
81 L274 “exhilarating” FFTF p.166 (cited p.163) Page (14)
82 L320 “more an act of faith than of reason” FFTF p.281 Hedge dropped (5)
83 L282 reversibility quotes 2025-03-02 main text, n.2 OK
84 L284, L670, L790 “not fully entrenched”; “patchy”; “rooted in chemistry”; “applying them effectively enough”; “sealed processes” Hansen et al. 2008 pp.444–447 OK
85 L287, L698 steering; “systemic instabilities” NN 2015-12 p.1006 (cited p.1005 at L698) OK / Page (14)
86 L594 careers built on assumed risk NN 2014-03 (actually Maynard & Aitken 2016 p.999) Miscited (7)
87 L248, L477, L797 public role vs expert drafting; “take the brunt” Handbook 2010 pp.583, 579 OK
88 L842 “can thus be seen to favor the enterprise” BMI 2019 p.6 OK

Also verified, with no issue: - 2015-01-30 (“until, suddenly, it isn’t”; “in isolation”) - 2021-04-09 - 2020-11-05 - 2018-12-15 (Luddite; “Being cautious ≠ smashing the technology”) - 2023-05-25 (Bengio, engaged on reasoning) - 2024-02-18 (“far less accountable”) - 2023-12-15 (“sheer fantasy”) - 2023-11-18; 2023-11-09; 2024-10-06 (“[sic]” correct); 2024-10-13 (both quotations) - 2024-09-01 (system card, main text and n.1; “simple test”) - 2025-03-15; 2025-08-10; 2025-03-30; 2025-06-01; 2024-03-31; 2024-05-15; 2026-01-17; 2024-01-18; 2023-10-02 - S3 2026 (“not on a plateau”) - NN 2016-06 p.491; NN 2014-09 p.659; NN 2015-03 p.199 - Rethinking Risk 2017 p.200; WEF 2008 (“independently of the funders”); Hyun et al. 2024 p.591 - 2026-01-10

INTERNAL (not for publication)#

Questions for Maynard, to settle items above: 1. Open weights. Was the 2025-07-23 line on open weights (“welcomed by many who worry about the corporate control…”) your own view? Do you hold a position on open weights and irreversibility? (Items 2 and 13; also asked in the M9 check.) 2. Pause letter. In 2023 you declined the pause letter but called for a “world congress” and “progressive” regulation “like, yesterday”. Would you describe your 2026 position on the “Pacing the Frontier” proposals as nearer Huang or nearer the pacing advocates, or as neither? (Item 4.) 3. April 2026 essays. Were the April 2026 andrewmaynard.net essays drafted with AI assistance? Should their wording be quoted as yours, or paraphrased? (Item 11.) 4. Series introduction. Will the published series introduction change the sentences quoted from draft 3, especially “not sure I fully agree”, “struggled to apply conceptual rather than literal comparisons” and “This was intentional”? (Item 12.) 5. Catastrophic risk. When you say catastrophic risk is “not so small”, is that the value-based sense of 2023-05-31 (large-scale loss of dignity, autonomy, purpose) or mass physical harm? The answer changes how 06 should set it against Huang’s “0%”. (Item 5.)