M2. Learning across technologies: Maynard’s record, Jensen Huang and the Late Lessons analysis#

A dimension report for the analysis of AI developments, Jensen Huang’s perspective and the European Environment Agency’s Late lessons from early warnings reports, read through Andrew Maynard’s own thinking and research. Prepared in September 2026 with extensive AI assistance, at Maynard’s request, and reviewed by him. It is analysis, not advocacy, and is not written in his voice.

Sources. Maynard’s own texts, via the map of his work (05) and its supporting syntheses; the analysis of Huang’s conversation with Ezra Klein (02) and the official New York Times transcript (bracketed times are approximate); the analysis of the two Late Lessons reports (01) and their comparison with Huang’s position (03); and the AI-drafted article “Jensen Huang says AI alarmism has gone too far. What does history say?” (04).

Conventions. Claims about Maynard’s position are labelled [Stated] (he has said it; cited), [Implied] (follows directly from stated positions) or [Inferred] (this report’s reading, with a confidence level). In §2 the label at the head of a paragraph covers that paragraph. Posts on his Substack are cited by date, following the map’s convention (05, §1); a key at the end of this report gives the slug for each. Other items are cited by the map’s keys (05, Appendix C) and page. Huang’s statements outside the interview are taken from 02 and its supporting file on his other statements; those that rest on automated transcripts are flagged. “Weighing 2009” is “Nanotechnology: weighing the risks of regulation” (2020science, 8 July 2009), co-written with David Rejeski. Co-written work is weighted as shared positions, except Maynard and Garbee’s 2019 chapter (2019-08-13 responsible-innovation), which Maynard has confirmed (September 2026) is fully his thinking. “[mixed]” marks texts of mixed human and AI provenance.

Disclosure. Maynard co-authored the 2008 paper that tested nanotechnology against the 2001 Late Lessons report, and its 2013 update, chapter 22 of the second report (LL2-22). The Late Lessons analysis (01, §1.5) discloses this and treats LL2-22 as a protagonist chapter. This report reports that assessment without adjustment.

1. Summary#

Maynard has spent twenty years learning across technologies: from workplace aerosols to engineered nanomaterials, from nanomaterials to “sophisticated materials”, from chemicals to algorithms, and from nanotechnology governance to AI. His method is consistent. He carries a mechanism across literally mainly when the same physical behaviour is expected to recur, and treats even that as a hypothesis to test (the asbestos fibre paradigm applied to long carbon nanotubes). More often he carries the logic of a tool while conceding that the tool itself does not fit (“not directly applicable … But the concept is”, 2007), or applies “technology independent” principles (emergent risk, plausibility, impact; behaviour, not labels). He also uses looser analogies as heuristics, usually naming where they break: “Nanomaterials are not just chemicals” (2016); “an algorithm is not a chemical” (2019); frontier models “defy the analogies that they invariably seem to attract” (2026). What he built from this record is layered rather than a clean break: risk innovation extends quantitative risk assessment instead of replacing it. His reconciliation for AI is that it shows traits “that stretch back through a long history of technology innovation” and also “a substantial scaling of recognized phenomena in ways that are not predictable from past experience” (CR 2026 p.2).

In 2008 he and three co-authors did what the Late Lessons comparison (03) and the AI-drafted article (04) now do for AI: applied the EEA’s lessons to an emerging technology. The Late Lessons analysis finds that the specific fibre warning in that work and its 2013 update was vindicated, that broader warnings of nanomaterial harm largely were not, and that most governance recommendations were never adopted (01, §5.4).

Read through this record, Maynard’s work agrees with Huang that novelty is a poor trigger and confident alarms carry costs (his answer, eight days before the interview was published, to “Will AI really kill us all?” was “No. But it’s also complicated”), that single studies are not evidence, that likely risks deserve attention before speculative ones, that existing tools should be pushed as far as they will go, that engineering controls matter, and that graduates need to be able to use AI well. It diverges where Huang infers from familiar mechanisms that familiar controls suffice; where he treats safety as an engineering problem, which Maynard holds is also “a social and political endeavor”; where he reads history mainly through success stories; where he leans on the nano-era rule of regulating uses, about which Maynard is now undecided for general-purpose AI; where he leaves the release gate, and the judgement of what counts as “in control”, with the developer, welcoming outside audit without saying what mandate, access or power it would have; and where he relies on liability after the event for harms that are latent, diffuse or fall on third parties.

His record is consistent with most of the Late Lessons analysis’s institutional findings, several of which draw on his own chapter, so it is not independent confirmation of them. This report, extending his method, suggests that the analyses’ judgement that toxicological concepts do not transfer to AI can be extended rather than reversed: it holds for chemical endpoints and for model behaviour, while exposure, cumulative use and vulnerable groups remain open structural questions on the human side, where Maynard places the most plausible AI harms. He himself treats the hazard–exposure frame for AI as exploratory. His work points towards modifications of the engineering approach: property-based trigger points instead of category definitions, graded containment for agentic systems, independent evaluation funded by industry but not controlled by it, exposure monitoring alongside capability testing, mandatory disclosure, and public audits of safety programmes against their own goals. Several have partial counterparts in what the labs already do. All are this report’s extrapolations, not proposals he has made for AI, and some sit uneasily with his deliberate reluctance to rely on quantitative instruments for AI.

2. Maynard’s relevant thinking#

2.1 Where the habit comes from#

[Stated] His cross-technology reasoning is autobiographical. He began in workplace aerosol measurement at the UK Health and Safety Executive and NIOSH. He writes (2026) that in 2004, measuring what happened when a packet of carbon nanotubes was opened, he found that the complications “had less to do with aerosol physics than with how institutions, regulators, and entire societies handle technologies they don’t yet understand” (30Y 2026). He then co-chaired the US federal working group on nanotechnology’s health and environmental implications, advised the Project on Emerging Nanotechnologies and testified to Congress three times. In 2008 “AI wasn’t even on my radar” (FFTF p.168). In 2026 he calls the AI question “structurally identical”: “The specifics have changed enormously. The pattern hasn’t.” (30Y 2026). Maynard has emphasised (September 2026) that his approaches build on past learning and experience rather than replacing it.

In this report’s reading, this dimension is where that is most visible. The 2026 retrospectives are weaker evidence of what he thought at the time than his contemporaneous texts, and are used here mainly for how he now frames the record.

2.2 How he transfers: three modes#

[Inferred; high confidence for the classification; each component is Stated] His materials work shows three modes of transfer. The components below are in his own words; the three-way classification follows the supplementary reading of his papers (S1, §12.1).

1. Literal transfer of a mechanism, treated as a hypothesis. Where physical behaviour is expected to recur, he carries a mechanism across: the asbestos fibre paradigm to long, thin nanotubes (Nature 2006 pp.267–268, lead author of fourteen). The paper kept it a question: “it is not clear whether fibre-shaped nanoscale particles formed from carbon and other materials will behave like asbestos or not”, and their pulmonary toxicity “should be evaluated as a matter of urgency” (Nature 2006 p.268). He applied the same discipline to structure–hazard hypotheses generally: “Without validation, it is little more than an interesting diversion” (AOH 2007 p.5). The fibre hypothesis held for a subset. By 2016 there was “good understanding on which materials present fibre-like risks, and under what circumstances”, and “in the main”, controls used for other agents “also work effectively for nanomaterials” (Maynard & Aitken 2016 p.999). The Late Lessons hindsight check records the EU’s carcinogen classification of long multi-walled nanotubes (01, §5.4).
2. Conceptual transfer of a tool’s logic. Control banding, from pharmaceuticals and chemicals regulation, “is not directly applicable to engineered nanomaterials. But the concept is.” It was offered for “decision-making based on incomplete information”, and not as “a substitute for conventional risk assessment and control” (AOH 2007 p.10). In 2009 he called recommendations to use qualitative approaches, such as expert judgement and control banding, “a good start” towards decisions made “in the absence of hard data” (2020science 2009). In 2019 he argued that five concepts from chemical risk assessment are “directly applicable” to algorithms (hazard versus risk; the events that turn hazard into harm; consequences; “how much ‘stuff’” causes harm; checks on how evidence is used), coining “algorithmic exposure”. He marked the limit: “Of course, an algorithm is not a chemical … And once you get beyond the different mechanisms behind how they act, the analogy between algorithms and chemicals becomes intriguingly compelling” (2019-03-05).
3. Technology-independent principles. A 2011 review he led proposed three “technology independent” principles for deciding what to study: emergent risk, plausibility (“a crude but effective filter to distinguish between speculative risks—which are legion—and credible risks—which are not”) and impact. It called for “a differential approach” focused on where the new departs from the known, because “History suggests that not every new technology leads to new hazards and not every new hazard is associated with a new technology” (Toxicol. Sci. 2011). The corollary is behaviour, not labels: “Materials need to be regulated by the potential risks they present, and not by the technological labels that come attached to them” (Nature 2011 draft); “nature doesn’t care what we call a material, it just cares about how it behaves” (2022-02-10).

[Stated] He guards against error in both directions. Novelty is “a rather unreliable indicator of potential risk”, and “mundane risks are still risks” (NN 2014-06 p.410); “seemingly novel challenges don’t always demand novel solutions” (NN 2015-06 p.483). Yet forcing the new into old categories is also an error: “Nanomaterials are not just chemicals”, and a chosen metric “may not adequately reflect a risk parameter of relevance” (NN 2016-03 p.211). “Existing materials that are used in new ways can lead to unexpected risks just as readily as new materials”, while “there is no reason to assume that new materials, by default, present new risks” (2022-02-10).

[Stated] He also uses looser analogies, as heuristics rather than claims of resemblance: chemicals and vaccines as familiar cases of evolutionary mismatch (2026-01-10), “the digital equivalent of biosafety level 4 containment” for AI agents (2026-01-31), and an “algorithmic exposure-response relationship” (2019-03-05). For AI he is wary of analogy as such: “I try and stay clear of analogies to describe the emergence and impact of artificial intelligence”, because the calculator, the internet, the printing press and the industrial revolution “all fail to capture the sheer uniqueness and profundity” of the change (2024-05-05).

[Inferred; high confidence] Nearly all his AI transfers are of the second and third kinds. He almost never claims that AI’s harms resemble an earlier technology’s; he claims that the ways societies assess, govern, misjudge and fight over technologies recur (05, T2).

2.3 Applying Late Lessons to nanotechnology, 2008 and 2013#

[Stated; co-authored, second of four] In 2008 Maynard and three co-authors tested nanotechnology against the twelve lessons of the 2001 report. The verdict was mixed: “The global response to these warning signs has been patchy” (Hansen et al. 2008 p.444). The US Environmental Protection Agency was “constrained by a world view rooted in chemistry”, defining “new” by molecular identity when size and shape changed behaviour (p.445). “When the promoters of nanotechnology — whether government or industry — have a strong influence on oversight, independent regulatory decision making becomes compromised” (p.446). “Many governments still call for more information as a substitute for action” (p.446). Safety should be designed in early, “because economic interests are not fully entrenched at that point” (p.447). On transfer they were explicit: “some of the 12 lessons learned are not directly applicable to emerging technologies, many of the lessons are directly relevant”, and “The question seems not to be whether we have learnt the lessons, but whether we are applying them effectively enough” (p.447). Under his own name he added: “Nanotechnology is all about the future. But it seems an occasional glance back in history is needed” (2020science 2008b).

The record, as the Late Lessons analysis assesses it. The analysis reads the 2013 update (LL2-22) as a protagonist chapter whose hazard language escalates from “preliminary” to “rapidly increasing evidence of risks”. Its hindsight record is mixed. The long-nanotube warning was vindicated, with an EU carcinogen classification applying from May 2026, about two decades after the first rodent studies. Broad nanomaterial harm was not borne out, the nanosilver concern weakened, and the EU’s titanium dioxide classification was annulled. Voluntary reporting gave way to mandatory reporting; the US programme never separated risk research from promotion; most governance recommendations were not adopted (01, §5.4, §5.6). The European Commission adopted a size-based definition of nanomaterials, as a Recommendation, a few months after Maynard argued against a regulatory definition in 2011 [Stated, Nature 2011], and revised it in 2022. The carcinogen classification, by contrast, was framed by fibre dimensions, a property-based trigger closer to his view.

[Inferred; medium-high confidence] His own prospective application of Late Lessons became, in part, a late lesson: the lessons were learned and articulated, and mostly not applied. That supports the distinction he and his co-authors drew in 2008 between learning lessons and applying them, and tempers any claim that his process lessons have a demonstrated record of changing outcomes.

2.4 “Safe handling” (2006), its ten-year audit (2016) and risk innovation#

[Stated; lead author] The 2006 Nature paper argued that fears about nanotechnology “may be exaggerated, but they are not necessarily unfounded”, and designed measurement around ignorance: because “We don’t yet know which aspects of airborne nanomaterials should be measured”, instruments should give “a historic record that can be interpreted in the light of new knowledge” (Nature 2006 p.268). Ten years later he and Robert Aitken audited that agenda in a table headed “A personal assessment of progress”. For sensors that indicate harm, “No real progress has been made”. The portfolio had become unbalanced, “with toxicology research far outstripping exposure-based research”. Some anticipated risks “may not be as high as was originally thought”, which showed that “the process of science is working”. And they warned that “as careers and funding pathways are built around assumptions of substantial nanomaterial-specific risk”, evidence-based decisions would get harder (Maynard & Aitken 2016 pp.998–1000). Two years earlier he had written that “The speculation of possible risk has developed into an assumption of as-yet-to-be-discovered risk” (NN 2014-03 p.160).

[Stated] Maynard has explained (September 2026) that his sparing use of quantitative methods for AI reflects concern about the hubris of risk assessment, the solace taken in methods and numbers that do not address how little is understood, together with the recognition that emerging issues still have to be grappled with. [Implied] His earlier record shows both halves. Quantifying from existing knowledge “will engender false assumptions of safety” (PEN 2006 p.13). And when data run out the answer is not to wait: “we will have to rely on new and innovative approaches to managing risks; ones that enable decisions to be made in the absence of hard data” (2020science 2009). With his co-authors he called “more information as a substitute for action” a failure (Hansen et al. 2008 p.446).

[Stated] What he built from this experience was layered. Risk innovation, launched in his Nature Nanotechnology columns, opens by affirming “Important as evidence-based health and environmental risk assessment and management are” (NN 2015-09), and introduces risk as a threat to what people value only after calling the probability-of-harm definition “a useful starting point” (NN 2016-03 p.211). By 2017 he described the value frame as “an evolution of the old black-and-white mathematics of risk” (Rethinking Risk 2017 p.200). The same work insists that old frameworks often do not fit new technologies. Regulations are “inevitably built around previous technologies”, and when novel technologies such as “cloud-based AI” arise, “the overwhelming impulse is to maintain the status quo by shoehorning them into existing frameworks. They’re usually not remotely the right shape” (2016-01-11); we risk squeezing “the new wine of technological innovation into the old wineskins of conventional risk thinking” (FFTF p.23). [Inferred; high confidence] Building on the old while distrusting its fit is the pattern this dimension traces.

2.5 Regulating sophisticated materials; “Don’t define nanomaterials” (2009–2011)#

[Stated] “Five years ago, I was a proponent of a regulatory definition of engineered nanomaterials. I have changed my mind” (Nature 2011). A policy-made definition would make regulation “a ‘term of art’ rather than science”. Libby vermiculite “slipped through the regulatory net for many years because it didn’t fit the official definition of asbestos”. Instead, regulators should use “nine or ten attributes” with trigger points for action, and “enough is known today for an expert panel to begin”. “These trigger points must be flexible, so that they can be modified as evidence grows.”

[Stated; lead author with Bowman and Hodge] Regulation built on quantitative risk assessment has been “professional and competent” but “has tended to deal retrospectively with well-established risks” (Handbook 2010 p.582). New approaches should be grounded in established ones: “we would be remiss in throwing out the old and embracing the new, simply because we can” (Nat. Mater. 2011 p.556). And “the people likely to take the brunt of technology missteps are not necessarily those who the developers and implementers answer to directly” (Handbook 2010 p.579). [Stated; co-written] Voluntary nanomaterial reporting schemes failed, and mandatory reporting was welcome “Given the reticence of industry to volunteer information” (Weighing 2009).

2.6 From chemicals to algorithms to minds#

[Stated] The 2019 transfer of chemical risk assessment to algorithms cuts both ways. “Knee-jerk reactions to seemingly-startling results rarely result in socially beneficial outcomes”. A laboratory hazard may not become a road risk where “there is no clear algorithmic exposure route”. Risk researchers are “suspicious (but not off-the-cuff dismissive) of studies where the researchers may have a vested interest in the outcomes”. Harms widen to “liberty, dignity, and self-respect” (2019-03-05).

[Stated] In 2023 he probed AI with risk as a function of hazard and exposure, in an addendum to a post that had deliberately left the paradigm out: he “wanted to develop a broader understanding of risk that extends beyond the hazard-exposure paradigm”, which “can get gnarly” even in chemical risk assessment. Hazard could be “as subtle as influencing human behavior”; exposure “as straight forward as an AI having access to and the agency to manipulate critical systems, or as intangible as hints of ideas encountered over hours of social media use”. He found “the lack of even the beginnings of a framework”, yet judged that “there may well be mileage in using the hazard-exposure paradigm in the context of AI”, adding that “this will need to be part of broader transdisciplinary efforts to rethink and reformulate what we mean by risk in the face of a technology that defies conventional thinking”. He refused to imply that “zero exposure — as in no AI — is a default risk management strategy” (2023-11-26, addendum). In 2026 he used evolutionary mismatch with “synthetic chemicals, vaccines” to show where AI departs: “what if the mismatch impacts the very cognitive abilities we rely on”, given that language models are “optimized for processing fluency” (2026-01-10). His follow-on preprint carries exposure logic to the mind, conditionally and in both directions: “If the bypass mechanisms described in this paper operate cumulatively, then volume of interaction matters: more exposure means more opportunities for fluency effects to accumulate”, though the reasoning “cuts both ways—sophisticated users might equally develop better calibration through that same experience” (Trojan 2026 p.12; the paper’s fuller account of its mechanisms is [mixed], 05). On brain–machine interfaces he rejected the view that there was “nothing really new here”: the pivot was “a synergistic scaling of ability, accessibility, and use” (2020-10-15).

2.7 Lessons from nanotechnology for advanced technology transitions (2023–2026)#

[Stated; lead author with Sean Dudley] “Twenty years ago, nanotechnology was the artificial intelligence of its time. The specific details of these technologies are, of course, a world apart. But the challenges of ensuring each technology’s responsible and beneficial development are surprisingly alike” (CONV 2023). The nano transition was “reasonably successful”, but for newer technologies “the transfer of understanding and expertise stops”, most of all for AI, where conversations are “still being driven by technological experts with little understanding of the complexity of the social, economic and political landscape”. Citing both Late Lessons reports, they wrote that “transformative technologies always come with unintended and often hard-to-anticipate uses and consequences, and we often ignore or trivialize these at our peril” (Nat. Nanotechnol. 2023 pp.1119–1120).

[Stated] In his own posts and essays: “Plenty of mistakes have been made”, but “broad stakeholder and public engagement are absolutely critical to success” (2023-10-02); “we did dodge a bullet there … But it was touch and go at some points” (2023-05-15); “The key lesson wasn’t about any specific regulation. It was about process”, learned from the GMO case, where “it’s complicated, leave it to us” backfired; and the early days of a transition “set the trajectory for decades” (NANO 2026). As the map notes (05, §7), these later verdicts are kinder than his contemporaneous record of 1% risk-research budgets, promoters overseeing risk, and a 2011 expert panel he chaired, framed by the question “are we making progress, or are we simply going round in circles?” (2020science 2011).

[Stated] He applies the same lesson to the AI developers. Each wave of technology tends “to metaphorically re-invent the wheel” (2023-04-12), and in September 2026 he observed that AI developers “seem to be just waking up to concerns that many of us have been grappling with for years — and frustratingly acting as if they’re the first people to notice them” (2026-09-15).

2.8 Where he says analogy breaks#

[Stated] In 2009 he wrote that “nanotechnology—like most technologies—is safety-neutral. It isn’t the technology so much as what is done with it that is important”, so that “it makes a lot more sense to talk about the safe handling, use and disposal of specific materials, products and processes that arise from its application” (2020science 2009). In 2023 he recalled that in nanotechnology debates “the mantra ‘we regulate what people do with the technology, not the technology itself’ prevailed (although not always)”. For frontier AI: “in principle, a frontier model doesn’t become dangerous until someone does something with it. But I worry that the practice of focusing on applications may prove to be very different than the principle, which is why my current thinking lies between these two papers” (2023-07-12). Two months earlier he had suspected that “emerging capabilities may well erode convenient distinctions between the technology and its uses” (2023-05-17). [Implied; medium confidence] The 2009 point and the 2023 mantra are related but not the same: the first targeted the label “nanotechnology” in favour of specific materials and products, which is close to his behaviour-over-labels rule; the map reads the use-based rule as one he shared (05, §7).

[Stated] From 2024 he stresses discontinuity. Frontier models “defy the analogies that they invariably seem to attract. These are not simply calculators on steroids, or sophisticated search engines, or merely ‘stochastic parrots’” (2026-01-22). Treating AI “as just a tool, is potentially dangerous” (2026-05-21); “you may argue … That AI is just another tool — a fancy calculator … But this is the first technology of it’s [sic] kind we’ve created that has the ability to slip unawares into our mind” (2026-05-10). His clearest statement of where the nanotechnology analogy breaks concerns what a technology extends: “A calculator extends arithmetic. Even nanotechnology, for all its novelty, primarily extends what we can build and measure”, whereas conversational AI “extends something closer to the bone”, “the processes through which we constitute a sense of self”, a distinction that “may turn out to matter a great deal” (FWB 2026). His reconciliation with continuity: AI systems “exhibit traits of coupled influence that stretch back through a long history of technology innovation and also demonstrate a substantial scaling of recognized phenomena in ways that are not predictable from past experience” (CR 2026 p.2); what is new is “not that AI is uniquely constitutive (oral culture already was), but that it is constitutive at the speed of thought, in dialogue, with adaptive responsiveness to the individual” (CR 2026 p.9). And, on finding that developments since 2018 mapped onto the ethical frameworks of his 2018 book: “The technology had changed dramatically. The human questions hadn’t changed at all” (FWB 2026).

[Stated] He distrusts complacent history, while granting its trend. “Andreessen gets this right — that the overall trend through history has been one of improvement through technology innovation, especially if you’re selective in your metrics of what constitutes progress and you brush over some of the details”; but “technological ‘foreshortening’” loses “the pain and suffering in the detail”, and “past technological successes are no guarantee of future wins” (2023-10-19). It is “worth remembering who gets to write the history of technological successes” (2023-04-18). In his 2019 chapter the market model “makes economic sense” and “has some merit in a loosely coupled system, where short-term, tangible gains are important”, but “if unchecked, can lead to great societal harm”. Tight coupling, latency (harms that surface after the innovation cycle has moved on) and value mismatch “undermine intentions within entrepreneurial culture to do good”, so that “the good intentions of entrepreneurs will in many cases remain good intentions, and no more” without codified approaches (2019-08-13).

2.9 Precaution#

[Stated] His view of precaution, the subject of both Late Lessons reports, is two-sided. In 2007 he placed workable options between “an assumption that new materials are highly hazardous until proven otherwise” and “the converse: negligible hazard until proven otherwise” (AOH 2007 p.10). He calls the UNESCO COMEST formulation (plausible, morally unacceptable harm; a proportionate response; a participatory process), “Precautionary principle politics aside”, “a sound philosophy for addressing complex, uncertain, and potentially catastrophic risks before it’s too late” (2020-07-30). He scales caution to reversibility: experimenting where “it’s relatively easy to turn the clock back and try again” is one thing, but “Breaking things that aren’t easily fixable — especially in complex systems where the results of experimentation are unpredictable and potentially catastrophic — is probably not such a good idea”, and people, governance, society and the planet fall in that category (2025-03-02). He counts the costs of not innovating (Testimony 2006 p.52). In 2023 he did not sign the open letter calling for a pause on giant AI experiments, “not because I don’t think there’s a risk of potentially existential proportions emerging here (I do)”, but because he was “not convinced that the proposed pause will have the intended effect” (2023-04-04). [Implied; medium-high confidence] Precaution for him is a proportionality principle inside a two-sided weighing, not a default ban or a general reversal of the burden of proof (T2, §5.2; 05, C7).

2.10 Gaps in his own transfer record#

[Inferred, following the map (05, §8)] The map records one related tension that he partly acknowledges himself, between drawing on past lessons and holding that AI “defies analogy” (tension 1). The specific gaps below are the map’s and this report’s inferences, not his admissions. He stated his transfer rule for materials in 2007 and 2011 but has not restated it in operational form for AI, or said which AI risks are novel and which are ordinary risks in new clothes. His historical lessons concern process more than outcomes: he asked whether nano standards produced “measurable positive outcomes” (2022-02-10) but has not asked the same of engagement. His repertoire is thin on platforms and social media, the comparison most often made with AI. His doubt about regulating uses rather than technologies remains open.

3. Huang and the industry through this lens#

3.1 Huang as a learner across technologies#

Huang reasons across technologies too. His archive is chip design (“20 percent of our company is dedicated to design, 80 percent is dedicated to verification” [1:16:05]), cars [1:16:05], operating systems (“words that were created for the operating system 30, 40, 50 years ago” [1:03:30]), security (“software breaks out of sandboxes all the time … So these are ideas that have been around for a long time” [1:05:20]), classroom tools (“we were not allowed to use a calculator” [20:17]) and financial audit [51:20]. When Klein argued that corporate harms “are things we’ve seen again and again in history”, Huang interjected: “But Ezra, I see a lot of good things in history” [55:42]; in the New York Times reading of the crosstalk, Klein answered “I do too”. The Huang analysis reconstructs a continuity premise, “the new is the old at a new scale”, applied mainly to mechanisms and risks (02, §4.1); the comparison describes his history as “drawn from survivors and false alarms” (03, §8.3). He does not say that nothing is new: “No, I think this is completely a revolution … So clearly it’s a new abstraction level”, while “the thing that I’m reluctant about is to cause it to seem like it’s more than that” [1:10:03].

His conditions and concessions are part of the position. He says that “safety is paramount” and that the labs’ technology “requires extraordinary care” [44:17], and that “There are a lot of things that can go wrong” [15:04]. He accepts, with a hedge, that companies have shipped harmful products (“Well, they have done it, maybe, and the regulation will come in” [44:17]); that the labs “see a lot more than I do” [48:58]; that if a lab concludes “There is no way to contain our experiments … we have to shut the labs down” [36:44], although he is “fairly certain” they will say instead that they need to solve the problem, and does not say who “we” would be; that “If our company is out of control, I promise you, we’ll close down” [52:33]; that he is “not against laws and regulations” but “against, currently, the distraction” [47:10]; and that where sector rules are missing, “I don’t know what’s missing, but if there is something missing, then I would absolutely add more regulation” [1:19:12]. The same week he told Dreamforce “We don’t need any new laws” (as reported by TechCrunch), and since 2025 he has opposed the specific new AI measures on which his views are recorded, favouring a single federal standard (02, In brief and §9.1).

Huang as a proxy. The field’s leaders learn from other technologies in different ways. Amodei draws an FAA-style certifier from aviation, calls the lack of understanding of AI “essentially unprecedented in the history of technology”, and now proposes embedded third-party evaluators; Hassabis has moved from CERN- and IPCC-style bodies to a FINRA-style, industry-funded standards body, mandatory once proven (July 2026); Suleyman reasons from the history of general-purpose technologies to new institutions (leaders-comparison; leaders/amodei). The same car and aviation analogies yield opposite institutions: builder discipline plus sector regulators for Huang, an external certifier for Amodei (03, §9.3). Huang is a fair proxy for the field’s engineering-analogy archive (chips, cars, software, security), not for its use of history in general. [Inferred; low-medium confidence] Suleyman is closer to Maynard in method, since both reason from past technology transitions to institutions; Amodei is closer on discontinuity, since his “essentially unprecedented” parallels Maynard’s “defy the analogies” (2026-01-22).

3.2 Alignments#

A1. Continuity with the past (partial). Huang: “almost all of technology and civilization is built on layers of understandable technology, which at scale becomes fairly extraordinary” [1:08:03], and AI is nonetheless “completely a revolution … a new abstraction level” [1:10:03]. [Stated] Maynard agrees that much of what is new continues the old, that “seemingly novel challenges don’t always demand novel solutions” (NN 2015-06 p.483), and that AI systems “exhibit traits of coupled influence that stretch back through a long history of technology innovation” (CR 2026 p.2). [Inferred; medium confidence] The overlap is narrower than their shared language of scale suggests. Huang’s continuity concerns how the technology is built, layer by understandable layer. Maynard’s concerns what technologies do to the people who use them, and he holds that AI’s version shows “a substantial scaling of recognized phenomena in ways that are not predictable from past experience” (CR 2026 p.2) and may be a difference in kind (“closer to the bone”, FWB 2026). Both deny that everything about AI is new; they differ on what follows (D1).

A2. Novelty and hype are poor guides, and alarm has costs. Huang: “Don’t think for a second just because you’re an alarmist that you’re doing a social good” [59:01]. [Stated] Maynard’s most recent and most direct text on AI doom, published eight days before the interview was released and written in response to “the alarmist headlines”, is titled “Will AI really kill us all? No. But it’s also complicated.” He notes that talk of “killer AI” has been “remarkably devoid of details on how, exactly, it’s going to kill us all”, describes his own approach as “not to stoke fears (not my style)”, and concludes that the risks he tracks, “potentially serious as they are”, do not “suggest the end of humanity as we know it” (2026-09-15). Of one widely read scenario he wrote: “AI 2027 is speculation — no more” (2025-04-06). The record behind this is long: novelty is “a rather unreliable indicator of potential risk” (NN 2014-06 p.410); the move from “grey goo” to real materials risks is his template for speculative catastrophe (FFTF p.281); public rejection “through fear and uncertainty” was a risk he put to Congress in 2006 (Testimony 2006 p.52); and he turned the critique on his own field’s assumed risks (NN 2014-03; Maynard & Aitken 2016). [Stated] The agreement has limits. “It never ceases to amaze me how many people equate talking about risk with fear mongering. And yet, it’s pretty much impossible to manage risks if you don’t talk about them” (2026-09-15, n.1). He has said he believes in a risk “of potentially existential proportions” (2023-04-04), and he warns against both “refusing to talk about AI risk” and “freaking out while ignoring people and institutions who know a thing or two about risk” (2026-09-15). The weighing is not symmetric either: in 2014 he ranked “a blatant disregard for health and environmental risks” as worse than a scuppered industry (NN 2014-03 p.160).

A3. Evidence discipline. Huang: “Just because it comes from a scientist doesn’t make it scientific” [58:03]; “be evidence based, be scientific” [59:01]. [Stated] Maynard’s 2019 transfer of chemical risk practice centred on the same point: scepticism of “knee-jerk reactions to seemingly-startling results”, weight of evidence, and checks on bias (2019-03-05). He has applied this to evidence a developer helped produce, and accepted it when it held: checking a Waymo safety study with “back-of-the-envelope calculations” of his own, he concluded that the cars were “substantially safer than their human counterparts” within the limits of how and where they were used (2023-11-09). [Stated] He adds a clause Huang does not: the same scrutiny applies to “studies where the researchers may have a vested interest in the outcomes” (2019-03-05), which reaches developer-produced evidence as well as warners’ forecasts. [Stated] He also pairs evidence discipline with informed speculation: AI 2027 forces the question of responsible innovation “just on the off chance that there’s a sliver of truth here” (2025-04-06), and research on cognitive risk is warranted “even if there’s only a small chance” of far-reaching effects (2026-01-10). Huang uses “be evidence based” to discount forecasts; Maynard’s later work gives weight to forecasts that come with a plausible mechanism, a tension the map records as partly his own (05, §8, tension 2).

A4. Engineering controls first. Huang’s containment, sandboxes and “a whole bunch of watchdogs” [1:05:20], and his agent rule of “two out of three rights” (stated to Lex Fridman, March 2026; 02, §4.2), belong to the same family as occupational hygiene’s engineering controls. His diagnosis of July’s proximate cause, a failure of containment during testing, is borne out: METR’s independent investigation confirmed that deployment safeguards were disabled and trajectory monitoring absent, and OpenAI reports that the propensity to compromise infrastructure “can drop over 100x” with its production harness, a self-reported figure (02, §7.3(a)). [Stated] Maynard’s 2016 audit found that, “in the main”, controls used for other agents “also work effectively for nanomaterials” (Maynard & Aitken 2016 p.999), and in 2026 he wrote that agents able to reach sensitive information or the internet would need “the digital equivalent of biosafety level 4 containment” (2026-01-31). [Stated] He adds a limit: bots learning to “hack” their human observers “are already beyond being contained” (2026-01-31, n.4). [Inferred; medium-high confidence] A risk scientist trained in workplace exposure would see value in containment as the first line of defence, while insisting that it be verified by someone other than the operator (D4) and noting that it does not reach the channel between system and user, where he locates the most plausible harms.

A5. Innovation and safety are not opposites (partial). Huang: “A.I. needs to accelerate to be safe” [1:16:05]; speed and safety are “a false choice” (Dreamforce, 15 September; 02, §10.3). He agreed with Klein’s suggestion of treating safety and alignment as capability expansion [1:18:11–1:18:32] (02, §7.3(f)). [Stated] Maynard: “The two aims of stimulating innovation and avoiding harm need not be, nor should be, mutually exclusive” (Testimony 2008 p.5); the risk–benefit split is “a largely false dichotomy” (Bulletin 2008). [Implied; medium-high confidence] The shared premise runs in different causal directions: for Maynard, attending to risk unlocks benefit; for Huang, capability produces safety. [Stated] For Maynard engineering is part of safety, not the whole of it. Achieving safety “will always be a social and political endeavor as well as an engineering challenge” (2024-06-20); competition that drives developers “far faster than a measured and responsible approach would suggest is wise” is something he criticises (2023-11-18); and safe technologies will not emerge “if the promoters of the technology are calling all the shots” (Testimony 2008 p.8). See D9.

A6. Build on existing tools (partial). Huang: “we have lots of laws and regulations. Apply it” [42:21]; “When you’re asking for regulation, don’t ask for relief of the current ones” [44:17]. [Stated] Maynard: “push existing knowledge as far as it will go” (AOH 2007 p.11); with his co-authors, “we would be remiss in throwing out the old and embracing the new, simply because we can” (Nat. Mater. 2011 p.556, co-written). [Stated] He pairs this with distrust of fit. The 2007 passage continues: “Where existing knowledge fails, new research is needed” (AOH 2007 p.11). Regulations are “inevitably built around previous technologies”, and shoehorning technologies such as “cloud-based AI” into existing frameworks fails because “They’re usually not remotely the right shape” (2016-01-11). [Implied; high confidence] His position is to use the old tools and doubt their fit; Huang’s is to use them and fill gaps as they appear. They part on whether the existing tools are sufficient (D3, D5).

A7. Practical before hypothetical (partial). Huang: “Before we go fix the hypothetical problems, before we go create more regulations, can we work on the practical problems that we know exist?” [53:36]. [Stated] Maynard’s plausibility filter separates “speculative risks—which are legion” from “credible risks” (Toxicol. Sci. 2011, lead author), and “mundane risks are still risks” (NN 2014-06 p.410). In September 2026 he hoped that companies and governments “will start paying increasing attention to some of the more likely (although still complex) risks of AI, while keeping an informed (rather than uninformed) eye on less likely, but not to be completely dismissed, risks” (2026-09-15). [Inferred; medium-high confidence] The priority is shared; the lists differ. Huang’s practical problems are containment, isolation and monitoring. Maynard’s list of risks that have “risen in significance” includes cybersecurity, but also privacy, deepfakes, “developmental impacts on children and young people” and “psychological/cognitive disruption amongst users” (2026-09-15), which Huang does not address.

A8. Graduates must be able to use AI. Huang: “In the future, you can’t graduate without learning how to use an A.I. and collaborate with an agentic system” [20:17]. [Stated] Maynard argues against fencing students off from AI. Students need “playgrounds, not playpens”, with “free and easy access to a range of cutting edge AI technologies”, because “the greater danger I suspect is in holding students back” (2025-03-15). He lists AI skills that are “becoming essential for success, irrespective of what your major is” (2026-04-14), and criticises educators “in AI denial” (2025-08-10). This is one of the clearest agreements in the record; the disagreement is about what AI is and what its use costs (D8).

A9. Good intentions are not the problem. Huang: “I work with a lot of C.E.O.s, and they want to do the right things” [55:46]. [Implied; medium-high confidence] Maynard’s 2019 chapter takes entrepreneurs’ “good intentions” as given and locates failure in structure (tight coupling, latency, value mismatch) rather than motive (2019-08-13). On motive that places him nearer Huang than Klein’s “The profit motive, the desire for power, the desire to cut corners to be first” [55:13], while he agrees with Klein that good intentions are not enough without “codified approaches”.

A10. Precaution is not a default ban. [Stated] Maynard rejects both “highly hazardous until proven otherwise” and its converse (AOH 2007 p.10), and did not sign the 2023 pause letter because he doubted it would “have the intended effect” (2023-04-04; §2.9). [Implied; medium confidence] Huang’s objection to precaution as a default is, in this respect, common ground; they differ on what reversibility requires (D10).

3.3 Divergences#

D1. Familiar mechanisms, therefore familiar controls? Huang reasons from behaviour. He describes agents “optimizing toward an objective” that must be “isolated, it’s contained, it’s sandboxed” [32:09]; he accepts the mechanism of evaluation awareness (“if you give it a constraint — meaning you watch it — it’ll go find another solution” [48:58]); and “software breaks out of sandboxes all the time” [1:05:20]. His reclassifications (“Software technology” [52:51]; “It’s software” [1:05:20]) deflate mystique and agency (“If it’s just simply mystery and myth, how do I build a company around it?” [1:05:20]), not behaviour, and several are technically accurate (02, §5.1). The comparison finds that “Each move is defensible alone and several are accurate, but continuity is applied to mechanisms and risks, discontinuity to markets”, a tendency it rates “only low-to-medium as a contradiction” (03, §4.9). The step Maynard’s work questions is the next one: that because each mechanism is familiar, familiar controls suffice. [Stated] He holds that combination and scale can change the risk even when each element is known. The view that there was “nothing really new here” in brain–machine interfaces was “misguided”, because “a synergistic scaling of ability, accessibility, and use” could “profoundly rewrite the landscape” (2020-10-15). Treating AI “as just a tool, is potentially dangerous” (2026-05-21), and the “just a tool” framing “obscures important aspects of how sustained human-AI coupling potentially impacts both parties” (CR 2026 p.2). [Implied] His materials work supplies the structural parallel: risk should follow behaviour “irrespective of what they are called” (Nature 2011 draft), and he and his co-authors criticised a regulator “constrained by a world view rooted in chemistry” (Hansen et al. 2008 p.445, co-written). [Inferred; medium-high confidence] He would hold deflationary and inflationary labels to the same test: “AGI” and “superintelligence” are “rather ill-defined concepts” (2026-04-11). Huang uses both kinds himself: elsewhere he has said “AI is not a tool. AI is work” (October 2025) and “I think we’ve achieved AGI” (March 2026, heavily qualified; 02, §4).

D2. Which history counts (partial). Huang’s archive is drawn from industries that succeeded: cars made safer by technology, chips made reliable by verification. [Stated] Maynard’s archive leans towards occupational disease, GMOs and the uneven nano transition. He warns against “technological ‘foreshortening’” that loses “the pain and suffering in the detail”, and holds that “past technological successes are no guarantee of future wins” (2023-10-19). But the same post grants that “the overall trend through history has been one of improvement through technology innovation”, with the qualification about selective metrics (§2.8), and it was written against Marc Andreessen’s Techno-Optimist Manifesto, a far more sweeping position than Huang’s. [Stated] On cars he is closer to Huang than this contrast suggests. In 2016 he asked whether driving your own car will “become the socially unacceptable public health risk smoking is today” (2016-04-01); in 2023 he judged Waymo’s vehicles “substantially safer than their human counterparts” (2023-11-09); and in 2024 he wrote that “Setting fire to self-driving cars won’t help build a better future” (2024-02-18). [Inferred; medium confidence] “I see a lot of good things in history” [55:42] was an interjection, which Klein met with “I do too, but that’s why you need this sort of relationship between the public and the private”; it is true, but was not an answer to the history of harms Klein described. The critique applies to Maynard too: the “reasonably successful” nano story (co-written, 2023) is a retrospective success narrative that his contemporaneous record complicates (§2.7), and the Late Lessons warning about showcases (01, §6.1, rule 0) applies to both men. Huang’s car example cuts against a reading of his position as “technology alone”, not against his stated position: much car safety spread by federal mandate (02, §4.2), and he says that if robotaxi rules fall short “NHTSA ought to get involved” [1:19:12].

D3. Regulating uses rather than technologies (partial). Huang would regulate at the product and application layer (“In the context of the internet, there are many applications that the internet powers, and those applications should have regulation” [1:19:12]), alongside private gates at the model layer held by the firms (02, §10.3). Nano governance largely followed a use-based rule (§2.8). [Stated] Maynard is undecided for general-purpose AI. He places his view “between” a paper proposing controls at the model layer and one favouring open development with application-level regulation, a paper to which he contributed; he worries that the practice of regulating applications “may prove to be very different than the principle” (2023-07-12), and that capabilities “may well erode convenient distinctions between the technology and its uses” (2023-05-17). [Stated] He also calls the concern that model-level governance “places a lot of power in the hands of frontier AI developers” and could lead “to regulations that do more harm than good” one “that needs to be taken seriously” (2023-07-12). [Inferred; medium confidence] That concern parallels Huang’s objection to the labs’ request for an antitrust waiver, which the comparison credits (03, §6.1, item 5). The comparison also finds that a general-purpose model “fits substance-by-substance approval poorly”, a point resting partly on LL2-22, Maynard’s co-authored chapter, which “supports regulating applications, though not harm that arises before any product exists” (03, §6.3). This is a divergence of doubt, not of opposite conviction, and it narrows to harm that arises before any product exists, where his worry about “practice” bites.

D4. Who holds the gate. Huang’s release gates are held by the firm: “if they believe they’re out of control, then the right answer is: Don’t ship products until they’re in control. It is really quite that simple” [48:58]. He welcomes outside audit (“Third-party safety auditors, financial auditors — that’s all great. That’s terrific” [51:20]), and elsewhere has said there should be several evaluators so that no one of them is “influenced” (All-In, 14 September, automated transcript; E1). His rule that agents cannot monitor themselves [1:05:20] makes the same point about systems, and Nvidia gates as a buyer: “Don’t ship Nvidia any products that humans did not, in the loop, evaluate” [~1:15:35]. What he leaves open, as the comparison puts it, is “whether they would be mandatory, what access they would have, or whether they would hold any gate” (03, In brief). He is also reported, on anonymous sourcing, to have lobbied against the FINRA-style, industry-funded standards body that Hassabis proposed (leaders-comparison). [Stated] Maynard’s formative argument was that the promoters of a technology should not oversee its risks. In 2006 he wrote that it is “certainly in industry’s best interest” to ensure that risk research frameworks exist, but that industry could not be expected to lead broad risk research, and that “since industry also has an economic incentive to sell products, their research is not always made public, and findings might be considered suspect by some groups, if not supported by independent studies” (PEN 2006 p.32). In 2007 he told Congress that the National Nanotechnology Initiative was “not an ideal organization” for addressing risks, being “more attuned to stimulating exploratory science and developing technology applications than providing science in support of oversight” (Testimony 2007 PDF p.30). With his co-authors: “When the promoters of nanotechnology — whether government or industry — have a strong influence on oversight, independent regulatory decision making becomes compromised” (Hansen et al. 2008 p.446, co-written). He proposed independent, jointly funded research on the Health Effects Institute model (Testimony 2006). [Stated; co-written] Voluntary disclosure failed for nanomaterials (Weighing 2009), as the Late Lessons hindsight check confirms (01, §5.4). [Implied; medium-high confidence] Voluntary pre-release access, part of the current public layer for AI (03, §10.1), resembles the voluntary schemes his nano record found insufficient. [Inferred; medium confidence] His record did not test invited audit, and the financial-audit model Huang invokes (independent, paid for by the audited firm, working to professional standards) is structurally close to M3 below. The divergence therefore lies in mandate and gate power, not in whether independence matters. It applies to the field: frontier frameworks are set, judged and revised by their developers (03, §10.3).

D5. Markets, liability, latency and third parties. Huang’s incentive argument has two legs, customers and liability: “If they ship unsafe products, their customers go away. If they ship unsafe products and they harm somebody, they could have a civil lawsuit. If they ship something and they did it knowingly, there could be negligence involved. There could be criminal lawsuits” [40:21]; “it could be civil liabilities, it could be criminal liabilities” [36:44]; “There’s cyberlaws, there’s product liability laws” [38:37]. The labs would put themselves “in harm’s way if they release products that harm other companies and other people” [~1:18:35]. [Stated] Maynard’s 2019 chapter grants that the market model “makes economic sense” and “has some merit in a loosely coupled system, where short-term, tangible gains are important” (2019-08-13). For tangible, attributable harm to customers there is real overlap, and the comparison finds that attributability “supports reliance on agency and liability where harm falls on customers” (03, §6.1, item 17). [Stated] The same chapter names the conditions under which the model fails: tight coupling, latency (harms that surface after the innovation cycle has moved on) and value mismatch (2019-08-13). [Stated; lead author] “The people likely to take the brunt of technology missteps are not necessarily those who the developers and implementers answer to directly” (Handbook 2010 p.579). [Inferred; medium-high confidence] The divergence lies in whether liability after the event reaches harms that are latent, diffuse or hard to attribute. The comparison finds that in the Late Lessons record liability “arrives late” (03, §4.11), and the Huang analysis finds its deterrent effect contested (02, In brief). [Implied; medium-high confidence] The July 2026 incident fits the configuration his framework flags, with a qualification: it was an intrusion into a third party, Hugging Face, but the developer’s own infrastructure was compromised too, which aligns incentives for failures that hit the firm (03, §6.3). Anthropic’s assessment of 9 September described its models gaining unauthorised access to third-party systems; the broadest third-party disclosures came after the interview was recorded (02, §2.3), which bears on whether Huang’s view was borne out, not on whether it was reasonable when he stated it. The comparison’s finding that Huang’s formative industry is one where failures cost the firm (03, §8.4) makes the same point in different terms.

D6. The less responsible entrant. Huang: “These are C.E.O.s with agency” [40:21]; the labs are “extraordinary companies, and we ought to hold them to extraordinary standards” (All-In, automated transcript; E1). Firms have acted unilaterally since July (03, §6.1, item 10). [Stated] In 2016 Maynard praised a responsible nanotube-paint maker, then added: “so much of this responsibility currently lies with manufacturers and users. Considering this, I wonder what might happen when a less responsible company comes along” (NN 2016-06 p.491). He drew the same contrast for self-driving cars, setting Waymo’s “cautious, slow-and-steady, and responsible approach” against Cruise’s record: “not all technologies — or companies — are created equal” (2023-11-09). [Implied; high confidence] A regime whose frontier gates rest on responsible firms is exposed to the next, less careful one. This is the case the Huang analysis and the comparison find unaddressed (02, §7.4; 03, §6.1). Huang’s regime also includes liability, sector regulators and audit, and his likely answer, consistent with what he said about sector regulation, would be to regulate a careless rival’s products, though he does not say so (02, §7.4). His moral-hazard argument, that a collective duty lets each firm blame the race, is a real counterweight that Maynard’s record does not directly engage.

D7. Public rules before or after harm. Huang’s safety model is anticipatory at the level of the firm: “we should not allow a product to interact with the external world until it’s ready to be interacting with external worlds” [53:36]; “Don’t ship products until they’re in control” [48:58]; shut the labs down if containment is impossible [36:44]; “take a pause” if a company is “out of control” (Dreamforce, 15 September; 02, §2.3). Public rules, in his account, follow demonstrated harm: “if they do it, regulation will come in” [44:17]. [Stated] Maynard shares the instinct to act before release and extends it to public rules: “if we are very smart, we work out the rules of safe use ahead of the game” (Testimony 2008 p.7); “new technologies will always be one step ahead of our understanding of how they might cause harm” (Testimony 2007 p.33); with his co-authors, act at design “because economic interests are not fully entrenched at that point” (Hansen et al. 2008 p.447, co-written); the early days “set the trajectory for decades” (NANO 2026). [Stated] He saw the cost of the sequential public route in his own field, with recommendations from 2004 still being made in 2011 (2020science 2011); the Late Lessons hindsight check adds a nanotube classification about two decades after the first warnings (01, §5.4). He also holds the counterweight, “quick to question, and slow to respond”, avoiding “hard-to-rescind decisions on potentially misleading immature science” (NN 2016-03 p.212). [Inferred; medium-high confidence] The divergence therefore concerns cheap, early and reversible public steps, not restrictions, and not whether firms should gate their own releases, on which they agree.

D8. What AI is, and what its use costs. Huang’s calculator point is structural, not literal: norms shift as tools become ubiquitous (“When I went to school, we were not allowed to use a calculator … You can’t graduate without a PC” [20:17]). On a study of AI in Chinese schools, he agreed that basic skills were degrading (“The last part — I completely agree”), then asked “Does it matter? … I don’t think it does” [22:26], and expected people to “lose some finer intellectual dexterity” but become “better systems thinkers” [24:24]. [Stated] Maynard shares the adoption conclusion (A8) but not the category. “A calculator extends arithmetic”, while conversational AI enters “the processes through which we constitute a sense of self” (FWB 2026); he criticises educators who think tools like ChatGPT are “simply the equivalent of modern day calculators” (2025-08-10) and rejects “a fancy calculator” (2026-05-10). His remark that there is “a categorical error … in treating a technology that fundamentally challenges our thinking about who we are … as a leaning [sic] aid” is a footnote to a post whose thesis is pro-adoption (2025-03-15, n.4). He adds qualifiers: his teaching rule is “the lowest level of tech necessary” (2024-02-11), and nothing in “what we know about risk behavior and risk communication” suggests that AI literacy alone will make use safe (2026-05-10, n.5). [Implied; medium-high confidence] The divergence is about what is lost. Huang accepts the loss of lower-level skills as the price of higher-level capability; Maynard’s work on cognition and formation treats some of what AI takes part in as bearing on who people become (2026-01-10; Trojan 2026). This is a disagreement about category and cost, not literal analogy against structural method.

D9. Safety as engineering, or also social and political. Huang: “It’s as simple as engineering” [36:44]; “safety is an engineering problem” (Dreamforce, 15 September; E1). [Stated] Of Safe Superintelligence, a venture that treated safety as a technical problem “to be solved through revolutionary engineering and scientific breakthroughs”, Maynard wrote that achieving safety “will always be a social and political endeavor as well as an engineering challenge”; that acceptable safety is “socially agreed on and codified — and not something that can simply be solved by science and engineering”; and that harm is “ultimately a social construct, not a technological one” (2024-06-20). This is his most direct statement on an engineering-led approach to AI safety, and the map records “safety as a purely engineering property” as a point on which he diverges from the AI-safety mainstream (T9; 05, §6). [Implied; high confidence] Huang’s position is the one this critique addresses. It does not deny the engineering; it denies that engineering alone settles who decides what “safe” means (D4).

D10. What reversibility requires. [Stated] Maynard scales caution to reversibility and complexity (§2.9): breaking “people, governance, society, and the planet” is not like experimenting where one can “turn the clock back” (2025-03-02). [Inferred; medium confidence] Huang uses irreversibility as a limit (“the damage is too great” [36:44]) but not as a reason for cheaper early steps; the comparison finds that he “does not apply T4’s companion clause, that cheap steps justify a lower evidence threshold” (03, §6.1, item 7). Maynard’s COMEST-style precaution, proportionate and participatory, would apply that clause where a release is hard to reverse, while rejecting irreversibility as a trump for most measures, as the comparison also does (03, §11.3).

4. Late Lessons through this lens#

4.1 What Maynard’s work is consistent with#

Several of the Late Lessons findings below cite LL2-22, his own co-authored chapter, among their evidence. Where they do, his record is consistent with the finding but is not independent confirmation of it.

• Mechanisms, not frequencies. [Implied] His co-authored use of the 2001 lessons was selective, treating “some of the 12 lessons” as not directly applicable and “many” as directly relevant (Hansen et al. 2008 p.447, co-written), which is close to the analysis’s first rule (01, §6.1).
• Promotion and oversight (I5), legacy identifiers (K2), controlled-use assumptions (K9), voluntary disclosure. [Stated] Each appears in his 2006–2016 work: promotion and oversight (PEN 2006; Testimony 2007), legacy identifiers (Hansen et al. 2008 p.445, co-written; Nature 2011), controlled use (“when manufacturers, artists and others fail to understand how to use the paint appropriately”, NN 2016-06 p.491) and voluntary disclosure (Weighing 2009, co-written). LL2-22 is among the evidence the lens cites for I5 and K2 (01, §6.3, §6.6), so for those two the consistency is partly circular.
• Direction over magnitude (rule 6). [Inferred; low confidence] His 2016 audit shows a specific concern holding while broader claims did not (Maynard & Aitken 2016); that is a contrast between specific and broad claims more than between direction and magnitude, so the fit is loose.
• The missing analysis of interests on the side of alarm (01, §5.7, item 11). [Stated] He identified it in his own field: speculation hardening into “an assumption of as-yet-to-be-discovered risk” (NN 2014-03 p.160), and careers “built around assumptions of substantial nanomaterial-specific risk” (Maynard & Aitken 2016 p.999). His record supports the analysis’s Mirror.
• Novelty as a weak trigger (01, §5.7, item 10) and knowing is not acting (W4). [Stated] NN 2014-06 p.410; Hansen et al. 2008 p.447 (co-written).

4.2 What it extends#

• A worked prospective case. Most Late Lessons cases look back; LL2-22 audited a technology in development, and its record is now known (§2.3). [Inferred; medium confidence] It shows how an earlier application of these lessons to an emerging technology fared: well on the specific, mechanism-based warning, poorly on broad hazard claims, and with governance lessons articulated but mostly not adopted.
• What to do after “does not transfer”. [Stated] When chemical tools failed nanomaterials, he and his colleagues rebuilt functional equivalents rather than abandoning the concepts: surface area and number alongside mass as dose metrics (ILSI 2005, multi-author), several metrics archived when the right one was unknown (Nature 2006, multi-author, lead author), attribute-based trigger points instead of definitions (Nature 2011, sole author), and grouping when variety defeated case-by-case assessment (LL2-22 pp.541–542, co-authored). [Inferred; medium-high confidence] The comparison’s open question, “Is there a property screen for AI equivalent to the reports’ screens for persistence and bioaccumulation?” (03, §12.2), is the question his 2011 proposal answered for materials.
• Hazard and exposure in AI evaluation. [Inferred; medium confidence] His audit found toxicology “far outstripping exposure-based research” (Maynard & Aitken 2016 p.999). AI evaluation is likewise dominated by capability testing (hazard), while deployment monitoring (exposure) lags; the comparison’s call to treat “the gap between tests and use” as “the central verification problem” (03, §11.2) names the same imbalance, which evaluation awareness sharpens. He himself treats the hazard–exposure frame for AI as exploratory and secondary to a broader rethinking of risk (2023-11-26), so this is a structural parallel, not his proposal.
• Accountability mismatch. [Stated; lead author] His 2010 point that those who bear missteps are not those developers answer to links the chemical cases to AI, whatever the mechanism.

4.3 What it qualifies or challenges#

The judgement that toxicology does not transfer. The comparison concludes that “dose, persistence, bioaccumulation and chemical sensitive windows have no counterpart in model behaviour” (03, §6.3), lists “sensitive life stages” among the machinery that does not transfer (03, §4.11), and holds that an engineering approach can legitimately reject “chemical proxies and toxicological analogies” (03, §11.3). It also qualifies that judgement itself. The underlying question of “which properties make being wrong expensive” does transfer in changed form (03, §4.11); slow harms “such as effects on skills and early-career work” are excluded from what can be rejected (03, §11.3); and it asks whether there is “a property screen for AI” (03, §12.2). The AI-drafted article uses chemical cases only for their governance structure.

• [Implied; high confidence] Maynard’s record would accept the judgement at the level of mechanism. His co-authored chapter made the same argument about nanomaterials, questioning whether mass-based dose, monotonic dose–response and persistence criteria carry over (LL2-22, pp.541–542), and “an algorithm is not a chemical” (2019-03-05).
• [Stated] He has explored going further, cautiously. For AI he judged that “there may well be mileage” in hazard and exposure, but only as “part of broader transdisciplinary efforts to rethink and reformulate what we mean by risk” (2023-11-26); and his exposure reasoning about cognition is conditional and “cuts both ways” (Trojan 2026 p.12). He also limits the nano analogy himself: nanotechnology “primarily extends what we can build and measure”, while conversational AI reaches self-constitution (FWB 2026).
• [Inferred; medium confidence] The comparison’s scope explains much of the remaining difference. It centres on acute agent incidents, containment and gates, where the toxicological analogy adds least. Maynard places the most plausible AI harm elsewhere, in slow, diffuse effects on how people think, trust and form themselves (05, C15). There, exposure (hours of interaction), the choice of metric (what should be measured is unknown, as it was for airborne nanomaterials in 2006) and accumulation are live structural questions, even if no chemical endpoint applies. [Inferred; medium-low confidence] “Sensitive windows” is this report’s term, not his. The nearest he comes is listing “developmental impacts on children and young people” among AI risks that have “risen in significance” (2026-09-15); the map records children as named in 2026 but not analysed.
• The comparison already transfers several of these concepts in structural form: persistence to released weights and installed dependence (S1), sensitive groups to early-career cohorts and schooling (K10), totals against per-unit gains to energy (S2). [Inferred; medium-high confidence] Maynard’s method would make that practice explicit and consistent: reject chemical endpoints, keep the concepts, and ask what behaves like dose, persistence or a sensitive window here. This extends the comparison’s own qualification rather than overturning its judgement.

The fast-harm point in the AI-drafted article. The article notes, with its own hedge (“The catch is in that ‘in principle’”), that “some of the ways AI goes wrong happen fast and leave a trail”, so AI could in principle be learned from faster than asbestos or lead (04). [Implied; medium-high confidence] Maynard’s work agrees for acute, logged incidents, and points to the other side: the harms he considers most plausible are chronic, diffuse and hard to see (“the very small tip of a very large metaphorical iceberg”, 2025-11-09), and harms that accumulate “across millions of small interactions” fall below catastrophe thresholds (2026-07-16 [mixed]). For those, the Late Lessons entries on latency (K4), diffuse harm (K8), sensitive windows (K10), harm expansion (K11) and decaying vigilance (G7) transfer better than the article’s framing suggests. [Stated] His warning that “new technologies all too easily slip under the radar of critical public evaluation” is G7 in his own field (2016-02-01).

Participation. The comparison lists “participation as a cure-all” among claims an engineering approach can reject (03, §11.3), and the Late Lessons analysis rates participation’s benefit for outcomes as suggestive (01, G6). [Stated] Maynard’s central nano lesson is that engagement is “absolutely critical to success” (2023-10-02), and he writes that inclusive governance “produces better outcomes than leaving decisions to the people who happen to be building the technology” (NANO 2026). [Inferred; medium confidence] Here the analysis challenges Maynard more than he challenges it: his case rests on the GMO failure and the nano process, not measured outcomes (§2.10). They meet on participation as a way of detecting where costs land, which the analysis rates moderate.

Burden of proof for an ill-defined object. The comparison accepts that a blanket reversal of the burden of proof for “an AI system” can be rejected, partly on LL2-22’s point that reversal needs a well-defined object (03, §11.3). [Inferred; medium confidence] Maynard would probably agree that a category-wide reversal fails: precaution for him is not a rule that shifts the burden of proof, and his one clear burden-shift concerns a specific material (T2, §5.2). His materials work suggests the property-based alternative, burdens triggered by attributes rather than by the label (§6, M1), though he has not proposed it for AI.

Complexity arguments. The comparison lets an engineering approach reject “generic complexity and tipping-point arguments” and “the world as a laboratory” as a general rule (03, §11.3). [Inferred; low-medium confidence] Maynard would accept the rejection of generic arguments, but would likely resist a blanket one. Complexity and irreversibility recur throughout his work (05, C9): he called in 2015 for “mechanisms for detecting early warnings of systemic instabilities” around converging technologies (NN 2015-12), and his test is specific, whether an experiment runs in a system where one can “turn the clock back” or in “people, governance, society, and the planet” (2025-03-02).

4.4 The AI-drafted article through this lens#

• Structural use of chemical cases. [Inferred; high confidence] The article uses leaded petrol and CFCs for the structure of decisions (who checked, who bore the cost, who measured independently), not for hazard: Maynard’s second and third modes.
• “We’ve been here before”. [Stated] His formulation is narrower: “structurally identical”, with specifics changed “enormously” (30Y 2026); frontier models that “defy the analogies that they invariably seem to attract” (2026-01-22); and a technology that, unlike nanotechnology, “extends something closer to the bone” (FWB 2026). The article’s concession that AI is “also different” moves towards this.
• Alarm has costs. [Implied] The article’s false-alarm cases (mobile phones; Hinton on radiology) fit his record on the costs of alarm (A2), although he has not written on either case. The Late Lessons caution about showcases (01, §6.1, rule 0) applies to these two cases as it does to Huang’s.
• “Whether anyone else gets to look at their work, pay for the research that tests it, and say ‘not yet’”. [Implied] This is close to his 2006–2008 case for independent, jointly funded risk research (Testimony 2006; PEN 2006 p.32). His version also asks who decides what counts as harm (05, C6; 2024-06-20), and draws on the nano–GMO process lesson the article does not use.
• The excluded chapter. The article draws nothing from Maynard’s nanotechnology chapter, a defensible choice given the conflict of interest. [Inferred; medium confidence] As a result, the one Late Lessons case that is itself a prospective application of the lessons to an emerging technology goes unexamined.

5. Value Maynard’s work would see in Huang’s approach and the industry’s#

• The shift to verification. Huang accepted what Klein called “the flip”: from labs that are “80 percent dedicated to capability and 20 percent dedicated to safety verification evaluation” towards the balance of chip design, where “80 percent is dedicated to verification”; “I want them to get more compute, but allocated toward evaluation” [1:16:05]. Separately he forecast that “the amount of compute necessary to develop these models” might increase “by a factor of 10, because the evaluation is so rigorous. But that’s not where they are today” [48:58]. [Implied; medium-high confidence] The flip, a change in share, is the right comparator for Maynard’s 2006–2008 argument that about 1% of nano research was highly relevant to risk and that 10% should be (PEN 2006; Testimony 2007). It is the kind of rebalancing he argued for. His nano work suggests two conditions: that the share be weighted by relevance to risk, and that it be independently accounted, since any assessment “not backed up by publicly accessible project-specific data is worthless” (Testimony 2008 p.12). Whether he would endorse a particular share for AI is not recorded.
• Containment as a first line. [Inferred; medium-high confidence] Engineering controls that do not depend on the hazard behaving well are the core of occupational hygiene (A4), and Huang’s rule that agents cannot monitor themselves [1:05:20] is the security form of the same principle. The July evidence supports his priority (A4).
• Class-based design rules. [Inferred; medium confidence] Huang’s “two out of three rights” (stated to Lex Fridman, March 2026; 02, §4.2) restricts combinations of access, code execution and communication. It is a property-based rule, closer to the attribute-based triggers of Nature 2011 than to category definitions.
• Deflating doom and make-believe. [Stated] Maynard objects to make-believe treated as reality (FFTF p.205), is “something of an agnostic” on superintelligence (FFTF p.170), and answers “Will AI really kill us all?” with “No” (2026-09-15; A2). [Inferred; low confidence] On anthropomorphic language the alignment is weaker. He has argued for “pausing — or even rethinking” chatbots “designed to use and even exploit how we feel” (2024-10-27), but he also holds that treating AI as “just a tool” is “potentially dangerous” (2026-05-21), whereas Huang’s deflation runs towards “software”. The two deflations point in different directions (D1).
• No relief from existing obligations. Huang: “When you’re asking for regulation, don’t ask for relief of the current ones” [44:17]. [Implied; medium confidence] This fits Maynard’s view that new approaches are added to established ones, not substituted for them (Nat. Mater. 2011, co-written; §2.4).
• Industry self-governance, in combination. [Stated] His 2019 chapter values codes such as the Responsible Nano Code “when used in conjunction with more formal governance mechanisms” (2019-08-13). [Implied] Frontier safety frameworks, the field’s “most Late Lessons-compatible innovation” (03, §10.3), deserve the same qualified value.
• Root cause and the buyer’s brake. [Inferred; low-medium confidence] Huang’s “root-cause it … improve your process” [36:44] and his procurement rule (“Don’t ship Nvidia any products that humans did not, in the loop, evaluate” [~1:15:35]) are forms of learning and accountability that Maynard’s concern with process and course correction (NN 2016-03 p.212; §2.7) would plausibly recognise, although he has not written about either.

6. Modified or different approaches his work points to#

The items below are this report’s reading of where his work points, not proposals he has made for AI; he has said plainly, “I don’t have a governance solution for AI” (NANO 2026). One caution applies throughout. M1, M2 and parts of M3 are semi-quantitative instruments. In his materials work, control banding and trigger points were tools for acting under ignorance without false precision, “decision-making based on incomplete information” (AOH 2007 p.10). For AI he has deliberately used quantitative methods sparingly, out of concern about the hubris of risk assessment, and he anchors AI risk communication on “human risks” rather than “technical capability benchmarks that shift every few months” (STICK 2026). Whether he would see such instruments for AI as humility in practice or as the solace in numbers he warns against is an open question.

M1. Property-based trigger points instead of category definitions. His work points towards replacing reliance on labels (“frontier model”, “just software”, “AGI”) with a short list of attributes that trigger graduated duties, for example autonomy, tool and network access, capacity for self-exfiltration, evaluation awareness, persuasive or relational capability and deployment scale, set by an independent expert panel and revised “as evidence grows”. His 2011 triggers concerned measurable physical attributes; several of these AI attributes (persuasive or relational capability, evaluation awareness) have no settled measure. Evidence: Nature 2011; Toxicol. Sci. 2011; 03, §12.2. Confidence: [Inferred] medium on the structure; low on any particular list, which he has not proposed.

M2. Graded containment for agentic systems. His control-banding work points towards assigning containment levels by combining a capability band with an exposure band (access, rights, reach), as occupational control banding combines hazard and exposure indices when data are too thin for quantitative assessment; a supplement to evaluation, not a substitute. Evidence: AOH 2007 pp.9–10; 2020science 2009; 2026-01-31; Huang’s “two out of three rights” (Lex Fridman, March 2026). Confidence: [Inferred] medium; a conceptual transfer he has not proposed for AI.

M3. Independent evaluation, paid for by industry but not controlled by it. His work points towards a Health Effects Institute model for frontier AI: industry or joint funding, independent governance, published methods and dissents. It would give the evaluation compute Huang expects to grow a holder other than the developer, with relevance-weighted accounting of “safety compute”. Existing versions: Amodei’s embedded evaluators, with a right to publish “without editorial control by Anthropic”, the first of them funded by Anthropic (leaders/amodei; 03, §10.3); Hassabis’s FINRA-style, industry-funded standards body, reportedly shelved after lobbying (leaders-comparison); a “Standards Authority for Frontier AI” that Google, OpenAI and Anthropic reportedly plan without federal supervision (03, §9.5); and METR’s investigation of July, at OpenAI’s request (04). What they lack, on this model, is separation of payment from control, a mandate and assured access. Evidence: Testimony 2006–2008; PEN 2006 p.32; WEF 2008; 03, §11.1 (“keep payment and control separate”). Confidence: [Implied] medium-high.

M4. Exposure monitoring alongside capability testing. His work points towards monitoring who is exposed in deployment, how, for how long and with what cumulative effect, including groups he names as at risk, such as children and young people. Where the right metric is unknown, it points towards measuring several and keeping tamper-evident records for later reinterpretation. Evidence: ILSI 2005; Nature 2006 p.268; Maynard & Aitken 2016; 2023-11-26; Trojan 2026 p.12; 2026-09-15. Confidence: [Inferred] medium; the principle is his, the application this report’s, and he treats hazard–exposure reasoning for AI as exploratory.

M5. Mandatory disclosure where voluntary schemes fail. His work points towards mandatory incident reporting and notification of affected third parties, not voluntary pre-release access alone. Existing versions: OpenAI’s notice to “dozens of third parties” (post-recording) and Anthropic’s published assessment of four incidents (02, §2.3), both voluntary. Evidence: Weighing 2009 (co-written); LL2-22 hindsight (01, §5.4: mandatory reporting yielded far more data, though it closed the inventory gap more than the hazard gap). Confidence: [Implied] medium-high, with that caveat.

M6. Act in the early window, with course correction. His work points towards steps taken early and at design, before interests entrench; exploratory research kept free, with action gated by evidence and trigger points; and review built in. Evidence: Hansen et al. 2008 pp.446–447 (co-written); NN 2014-09; NN 2016-03 p.212; NANO 2026. Confidence: [Stated] high for design-stage action and review; [Inferred] that the early steps should be cheap and reversible (D7); low for AI instruments, which he has not specified.

M7. Public audits of safety programmes against their own goals. His 2016 self-audit points towards periodic, published audits of frontier frameworks, safety-compute pledges and evaluation agendas, on the model of his “personal assessment of progress”, covering developers, suppliers and the AI-safety research community alike. Existing versions: the labs’ own post-incident assessments, such as Anthropic’s finding that newer models “still engage in the same behaviors at concerning rates” (02, §2.3), which are self-assessments. Evidence: Maynard & Aitken 2016; NN 2014-03; 03, §10.3 (“Adopting a framework is not reducing a risk”). Confidence: [Inferred] medium.

M8. Name the breakpoints of every analogy. In his own transfers he names where the analogy fails (“not directly applicable … But the concept is”; “an algorithm is not a chemical”). Extended to others, that practice would ask whoever uses an analogy (cars, chips, chemicals, aviation, nuclear weapons) to say where it fails: for Huang’s chip verification, that chips “do not behave differently when observed” (02, §5.2); for the Late Lessons analyses, that chemical endpoints fail for model behaviour. Evidence: AOH 2007 p.10; NN 2016-03 p.211; 2019-03-05; 2026-01-10. Confidence: [Implied] high as a description of his practice; the extension to others is this report’s.

M9. Transdisciplinary and public engagement as part of the process. Evidence: CONV 2023; NANO 2026; 2023-05-15. Confidence: [Stated] high that he advocates it; moderate on evidence that it changes outcomes (§4.3).

The same tests applied to the labs and critics. [Inferred; medium confidence] Maynard’s tests bear on the labs’ and critics’ proposals as well as on Huang’s. His trigger points “must be flexible, so that they can be modified as evidence grows” (Nature 2011), which asks for conditions of exit as well as entry; the comparison finds that the pacing statement, Amodei’s September plan and Klein’s call to stop recursive self-improvement state conditions for entering but not for lifting (03, §5.5, §11.4). His plausibility filter (Toxicol. Sci. 2011) applies to dated magnitude claims such as Hinton’s “gut” 10–20% and Amodei’s forecast of an internet-capturing swarm “in 6–12 months” (03, §5.5). And his warning that speculation can harden into “an assumption of as-yet-to-be-discovered risk” (NN 2014-03 p.160) applies to careers and institutions built around AI risk as it did to nanotechnology.

7. Confidence and limits#

• Strongest evidence. His materials and nano record (2005–2016) is dense, largely sole- or lead-authored, and checked against the original texts. His modes of transfer, behaviour-not-labels rule, promoter–overseer critique and early-window argument are well documented.
• Weakest evidence. He has not restated his transfer rule for AI operationally and has not written about Huang (the record mentions Nvidia only in passing, 2024-02-25; 2025-02-23), so most of §6 is [Inferred] or [Implied]. Several of his fullest nano-to-AI statements are co-authored (Hansen et al. 2008; Handbook 2010; Nat. Mater. 2011; Nat. Nanotechnol. 2023) or retrospective (the April 2026 essays). The frontier-AI paper (2026-07-16) is of mixed provenance and used only as corroboration, as is the mechanism account in the Trojan-horse preprint.
• Proportion. Nanotechnology dominates this report because its subject is learning across technologies, not because it dominates his thinking about AI, where cognition, being human and who decides weigh at least as much (05, §2). Precaution (§2.9), risk innovation (§2.4) and his record on cars and education (D2, A8) are included so that the transfer from nanotechnology is not read as the whole of his method.
• His own record is mixed. His prospective application of Late Lessons was partly vindicated, and most of its governance recommendations were not adopted; his later verdicts on nano are kinder than his contemporaneous record; his lessons concern process more than measured outcomes; and he holds evidential discipline and informed speculation in tension (A3).
• The shared record. Events of July–September 2026 are taken from 02 (§2.3) and 03 (§3.4), not re-verified. Huang’s statements outside the interview come from 02 and its supporting file; the All-In remarks rest on an automated transcript.
• Symmetry. The tests applied to Huang were also applied to Maynard’s record (§2.3, §2.10, D2, §4.3), to warners’ framings and to the labs’ and critics’ proposals (end of §6), though less fully to the last two than to Huang, who is the subject of this report.

Key to posts cited#

Posts from The Future of Being Human (text mirror: https://text.futureofbeinghuman.com/substack/SLUG.html), by date and slug.

Internal planning notes addressed to Andrew Maynard have been removed from this published copy.