Late Lessons, Jensen Huang and AI
Markdown version

B31 perspective notes: 2026-07-16 to 2026-09-20 (9 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. I read all nine posts in full, including the whole of the roughly 11,000-word orphan-risks paper. None of the posts is a Modem Futura episode note. Quotes are exact, including his typos (“There’s a change of course” for “chance”, “a good enough reason as any”) and curly punctuation. Markdown italics are dropped.

Evidence rules applied

Context. Summer 2026, during his first ever sabbatical. He is running a run of public experiments with Anthropic’s Fable 5 and then Fable 5.1. He announces a move from ASU’s School for the Future of Innovation in Society to the Thunderbird School of Global Management, taking the Future of Being Human initiative with him. Bill Gates publishes “The turbulent AI era is here”. In mid-September there is a burst of AI-extinction alarm.

1. How he thinks here#

He makes himself the experimental subject, and publishes the result before he knows what it means#

The Fable writing series is a sequence of self-experiments run in public.

  1. The doubt comes first. The orphan-risks preamble (2026-07-16) does not open with the paper’s findings. It opens with a doubt about himself: “Was my version better in my eyes because I have a rather old fashioned and biased perspective on what an academic paper should be like?” He admits to a “crisis of identity around whether I can actually write papers any more in a world of AI”.
  2. The comparison is run with himself as one arm. In publish-or-perish (2026-07-19) he runs “an experiment (admittedly with an “n” of one)”. He pits his line-edited version against Fable’s, gives the two papers anonymised animal names (“Rabbit” and “Marmoset”), and asks four AI models to judge them. They all prefer Fable’s version. His response is “Ouch!”
  3. He sets out both sides of his uncertainty. He does not settle it quickly either way: - “lost in his own myopic hubris” - then “OK, so I actually don’t believe this”, with an appeal to “nearly 40 years of experience” - then again: “There’s a change of course that the LLMs are right and I’m wrong here.” - The subtitle states the result: “it’s left me more uncertain than ever.”
  4. The reader is part of the method. He supplies both papers as PDF and as Markdown “that is easy to give directly to an AI”, and asks readers to judge “and why (especially as “better” is such a subjective term)”.

This is how a physicist runs a small experiment, carried out in the open. The instrument includes his own judgement, and he treats that judgement as something that could be wrong.

He moves from “what” to “how it is made”#

The orphan-risks paper’s key move is to change the question. He accepts that AI risks have mostly been named already (the MIT repository’s “more than 1,600” of them). So “rather than just asking what the orphaned risks are, a more revealing question is how they are made. In effect, by what process does a known risk come to be nobody’s responsibility?”

He then traces the exclusion to its root, which is how risk is defined, not how well people behave: “it is actually working as designed. It’s just that the design itself may be flawed.” This is first-principles thinking applied to an institution. It explains a pattern by its design, not by bad faith.

He will not settle for a villain#

He works through the case that companies are acting cynically, and rejects it: “sincerity almost always operates inside an incentive field”. Sincere people “reasoning one reasonable compromise at a time, produce the same sort of drift that intentional bad actors might create on purpose.” His conclusion is practical: “remedies aimed at sincerity, such as exhortation or public shaming, are unlikely to have the desired impact.” Instead, remedies must “change what competition rewards”.

He looks for structural explanations because they point to levers you can pull. Blame does not.

He asks for the mechanism and tests plausibility#

In will-ai-really-kill-us-all (2026-09-15), what he notices first about the extinction talk is its lack of mechanism: it is “remarkably devoid of details on how, exactly, it’s going to kill us all.”

A footnote then diagnoses the psychology: “Being spooked by a general feeling of dread is a very human reaction when confronted by something that you can’t explain”. Imagination “is very happy to fill in the gaps”, but “acting on instinct is its own form of risk as it leads to decisions without understanding or reason.”

He is not dismissing fear. He is saying that fear without a causal account is a poor basis for decisions.

He goes back and checks his own earlier forecasts#

The same post looks back at his own 2018 video list of ten AI risks and asks how it has held up: “less has changed over the intervening eight years than might be imagined.” He then updates the list honestly. He names what has “risen in significance” (cybersecurity, water and energy, deepfakes, children, “psychological/cognitive disruption”) and admits that some are “definitely more prominently on my radar now”.

Public writing here works as a dated record he can hold himself to.

He builds things to find out#

Four of the nine posts are about something he made, or made with an AI, in order to think with it: - Hyperbubble, the game (“over 90 iterations”) - mull.chat, the reasoning parody site - the Fable 5.1 authorship experiment - a sealed, pre-registered project

The thinking comes after the making, and it surprises him. Of Hyperbubble he writes: “I’ve been surprised (although I shouldn’t have been) by how much my thinking around the importance of playful exploration has evolved”. It was “influenced by actually playing the game — which is something I wasn’t expecting.”

Serendipity sets the agenda#

Several posts begin with an accident, not a plan: - The Fable 5.1 post starts when “a couple of things unexpectedly dragged me down an AI paper-writing rabbit hole”. One was a Nature commentary citing a preprint he had forgotten to write about. - mull.chat started “with me being amused by some of the weirder stream of reasoning messages I was seeing, and idly wondering what a website might look like that parodied these.” - The Gates post starts from something missing: “in all the responses to it I’ve read so far, I haven’t seen much on what it implies for university leadership”.

He notices a gap or an oddity and follows it.

He reads on several levels at once#

Hyperbubble (2026-08-02) is read in layers. It is “just one more AI-created game”. It is also “a unique representation of my work”. It is “a surprisingly sophisticated way of exploring the often-complex tensions between technology innovation, risk, decision-making, and future flourishing”. And it is “an engine of delight”. Then: “it’s where all of these come together that I find that things get interesting.”

He then undercuts his own case: “Of course, I’m just messing around here, and am probably over-stretching the significance … But that, of course, is the point.” The self-mockery is not a disclaimer. It is part of the argument that play needs no justification.

He pairs rigour with play, both ways round#

Pre-registered-play-open-april-25 (2026-08-23) takes a research-integrity tool, pre-registration, and applies it to play. He publishes a sealed file on Zenodo with SHA-256 hashes “For the skeptics (and nerds)”. The justification pairs two values that usually sit apart: “But where’s the fun — or the accountability — in that?”

He says outright that the word “play” is chosen on purpose: “drawing on my ongoing work and thinking around the concept of navigating and thriving with frontier technologies through the metaphor of a playground rather than a playpen.”

In the other direction, the mull.chat post uses play to make a serious point about the “performative nature” of LLM reasoning streams.

2. What matters to him#

The human experience at the far end of the page. His objection to AI academic prose is not about accuracy. It is about what reading feels like. Fable’s paper “has studied the form of the “academic paper,” but has no idea what the experience of reading one is like to a real person.” The qualities the AIs praised (compression, no “throat-clearing”, a “mic drop” at the end of every paragraph) each “hinder the process of enabling the reader to get a glimpse into the mind of the writer” (2026-07-19). For him, writing is a way of connecting one mind to another. What he protects is the reader’s experience, and through it, human judgement.

How technologies reshape people. The deepest worry in the batch is that human standards drift towards the machine’s: “the AIs we have trained to “think” like us are now beginning to train us to think like them” (2026-07-19). The same concern with formation appears in a positive form in Hyperbubble. The game is designed to “contribute to the formation of a mindset that is attuned to thriving in a technologically complex world. But what a player takes away from it is uniquely theirs” (2026-08-02). He cares about who we become, not only about what the technology does.

Joy, delight and play as serious values. “I find “joy” a deeply under-appreciated metric of intellectual and academic achievement!” (2026-09-20). He shares mull.chat because “it brings me joy, and that seemed a good enough reason as any to share it.” He is frustrated that professional life treats these things “as trivial, immature, and not appropriate for serious people doing serious jobs”, and that there is “a lot of lip service” while “actions so often speak louder than words” (2026-08-02). He cites the Future of Being Human initiative’s principles as central to the mull.chat project: “Obsessive Curiosity”, “Radical Creativity”, “Grounded exuberance”, “Catalytic Serendipity” (2026-09-20).

Care and honesty in scholarship. He sets a demanding floor for AI-assisted work: “any paper that took less than 10-20 hours intensive human labor working with AI is, in my mind as an academic and researcher, highly suspect!” (2026-09-04). He checks citations “painstakingly”. He assigns authorship honestly even when it costs him credit. Fable is sole author because “I did not make a substantial intellectual contribution”. He then points out the institutional gap this exposes: “no straightforward mechanism for publishing papers with AI as author”.

Universities’ duty to the public. He is “someone who believes fiercely that universities have a deep responsibility to leverage their considerable freedoms and unique capacities for public good” (2026-08-30). He is frustrated that they are “guardians of the past more than leaders toward the future”, and that their culture is “ironically, deeply intolerant of people who do not play by an arcane set of rules.”

Privilege brings obligation. A sabbatical footnote admits “how privileged I am to be paid to think, to write, and to teach, with a level of autonomy and security that few other jobs afford. And, of course, the responsibilities and obligations that come with this.” He frames the sabbatical around “how I might continue to use the privilege of the position I have to impact others” (2026-08-02).

People with the least leverage. The orphan-risks paper names who a value-based approach might still miss: “Data workers in annotation supply chains”, “communities carrying the environmental costs of compute”, “people affected by systems they never chose to use” (2026-07-16).

What frustrates him: - AI developers “frustratingly acting as if they’re the first people to notice” risks others have worked on for years (2026-09-15) - the equation of risk talk with fear-mongering (2026-09-15) - “a cacophony of loud voices with very definite — if not always well-informed — ideas” (2026-08-30) - claims of hours-long AI paper pipelines: “I do not believe them!” (2026-09-04) - people trusting “a whole army of AI agent reviewers” over human judgement (2026-07-19)

What delights him: - being surprised by what a collaborator, even a machine, came up with. Fable’s “Today’s Future” mode “took me a few days to discover”, and there were strategies “I had to discover on my own” (2026-08-02) - LLMs parodying themselves “quite delightfully and unintentionally” (2026-09-20)

3. Risk as a way of thinking#

A redefinition that builds on the old foundations#

Much of the orphan-risks paper (2026-07-16) turns on redefining risk as a threat to value. He is explicit that this adds to probabilistic risk rather than rejecting it: “This does not abandon the idea of risk as involving the probability of harm. Rather, it widens what counts as harm”. It is framed as “not as an alternative, but as an augmentation”, and “nothing here argues that the catastrophic-capability apparatuses that are already in place should be loosened.”

The quantitative foundation stays in place. The new lens is laid on top of it.

What the lens makes possible#

The dignity example shows the redefinition working as a mental model, not a metric. Under probability-of-harm, “the conventional machinery of risk has little or nothing to run on”. But ask “who (or what) holds dignity in a given situation, what threatens this, and what protecting it would look like” and it “becomes something that can be acted on — even though nothing has been quantified.”

The general principle is that value “can be named, mapped and watched — even where it cannot be measured.” Drawing on Porter’s history of institutions retreating to numbers under scrutiny, he explains why the filters reward the quantifiable. He also concedes that “Numbers, of course, do not always uniquely capture the essence of what is relevant”. This is a scholar trained in quantitative risk showing where false precision leaves blind spots.

His vocabulary is consistently about navigation: - the “risk landscape” is “becoming increasingly hard to navigate” - the aim is to “navigate a complex risk landscape with open eyes” - the closing line: “running blind has never been a particularly good risk management strategy”

The most striking sentence reverses the usual question about risk: the risks most likely to blindside frontier AI “are the ones its institutions have organized themselves not to see.”

The landscape also contains opportunity. The goal includes “transforming vulnerabilities associated with orphaned risks into advantages”.

A lens that works on how organisations think#

The Maynard & Garbee lesson, which is his own, shows the value framing working on the culture of an organisation: “if you want a fast-moving organization to attend to a risk, you do not hand it a compliance duty; you show it a threat to something it values.” “The lesson that has stayed with me ever since.” He applies it to AI labs, which he describes as “mission-driven, often allergic to imposed process”. The framing is persuasive because it meets people where they already care.

In the same spirit he recasts founding charters as “assets” that “can be spent”. That makes a company’s own framework changelog “a leading indicator that the company is drifting”.

Designed for people with little time#

The risk innovation tools had to be “simple and intuitive”. The design principles “were driven by utility”, for people with “little time and even less money”. The Planner asks users to commit to “a handful of small actions that are completable within a few weeks, and then repeat.” For him, a concept is only as good as its use in a hurried, real organisation.

Holding himself to test#

He proposes three tests that could count against his own account. He names a result that “would count against the incentive-driven account argued here”. He calls the whole analysis “defensible, but has yet to be shown to be useful in practice.” This humility is structural: he builds falsifiability into a conceptual argument.

Risk as conversation, not fear (2026-09-15)#

Risk thinking you can feel (2026-08-02)#

He says Hyperbubble’s gameplay “is inspired by my work around navigating an increasingly complex risk-benefit landscape around emerging technologies”. In the game: - orphan risks can be “adopted” - moral panics “die down if you ignore them” but worsen if fed - there are “doom pits”, a “hype line”, “serendipity tokens” and “black swans”

There is no single right strategy. Fast and exuberant play works “as long as you develop your risk navigation skills”. Slow and cautious play also works. “(navigating the future is hard)”. The mechanics are Fable’s, but he endorses them. The endorsement matters because the game treats risk navigation as a skill learned through practice across many strategies. It does not treat it as compliance with a rule.

The same logic turned on himself (my interpretation)#

Pre-registered play (2026-08-23) applies to his own work the accountability he asks of AI companies in the orphan-risks paper. There he proposes an “aperture log” and an “orphan-risk register” that make scoping decisions public and versioned, so that a walk-back must be explained. Here he seals his own plan so that “I will not be able to wriggle out of what I’d originally set out to do, no matter how things went.” He does not draw this link himself. The parallel suggests that his accountability ideas are habits of mind he lives by, not just recommendations for others.

Unnamed orphan risks in the writing posts#

Publish-or-perish (2026-07-19) describes what his framework would call a threat to value. It is not catastrophic and not measurable by any capability test: the slow convergence of human standards on “an LLM-view of what good writing is”. He does not call it an orphan risk. But it has the accumulative, below-threshold shape that the orphan-risks paper argues current frameworks are built to miss.

4. Scholarship and public writing#

The Substack as a lab. A full academic paper runs inside a newsletter post, with a preamble about the process (2026-07-16). The follow-up turns the paper into a public experiment with reader participation (2026-07-19). The Fable 5.1 post reports on institutional friction around authorship: an earlier AI-authored preprint “was not accepted on arXiv (I suspect the AI thing was an issue)”. The AI-authored work goes to Zenodo with a CRediT-AI annex (2026-09-04). Scholarly norms are being worked out in public, as the work is done.

He writes academic papers as stories. His version of the paper opens with a narrative: “Sixteen months later, persuasion was gone.” It keeps a first-person voice (“I must confess that I am not optimistic”). It hedges openly (“These examples are, of course, anecdotal”). It lands on memorable lines (“A framework, it turns out, can be an excellent exhibit, and a weak instrument, both at the same time”).

This is the “narrative form” he says he restored to Fable’s compressed draft: “a sentence or paragraph can be technically accurate but narratively ineffective” (2026-07-19). The paper itself demonstrates what he is defending.

Transdisciplinary by construction. One argument draws on: - anthropology (Douglas and Wildavsky) - history of science (Porter) - sociology (Power, Vaughan) - risk analysis (Kasperson) - philosophy (Kasirzadeh) - responsible research and innovation (Stilgoe, Owen, Macnaghten) - enterprise risk standards (ISO 31000) - securities law (10-K filings) - his own fieldwork with entrepreneurs

He calls his definition “a deliberate fusion of approaches and framings”. It keeps enterprise risk management’s adoptability “while also widening the circle of those whose value counts.”

Evidence and expertise. He treats AI verdicts as data about AI, not as authority. The four models’ unanimous preference prompts his curiosity, not deference. He also treats his own expertise as fallible (“I may be an elitist curmudgeon”, 2026-09-04). He values expertise gained in practice: the “people and institutions who know a thing or two about risk” (2026-09-15), and “my own experiences working with entrepreneurs” (2026-07-16).

Accessibility as a design commitment. - Risk Bites is “very intentionally aimed at helping viewers from all backgrounds make sense of complex risks in a very short and accessible format”. It is “rather light on detail. But they do draw on deep expertise” (2026-09-15). - He makes fun of its stick figures (“no talent and even less time”, 2026-08-02). - He is experimenting with a new route to his ideas: asking readers to point their AI at his text mirror or at beinghuman.fyi, which is “designed to be used with AI” (2026-08-16). He admits in a footnote that this is “a bit of an icky AI shortcut I know” (2026-08-30).

Playful artefacts as scholarship. mull.chat is “a serious part of” a body of work that “uses play, creativity, and serendipity, to explore new ideas in unexpected and often deeply insightful ways” (2026-09-20). He admits that “like humor, such methods — and their results — can be a matter of taste”. He counts a game, a parody website and a sealed envelope as legitimate forms for a scholar’s thinking.

5. His role as he sees it#

A public scholar whose job is to help people navigate. The Thunderbird announcement (2026-08-16) states the mission: “navigating advanced technology transitions”. The aim is to “help equip people and organizations to navigate the coming age”. He will work on how “this and other “public scholarship” catalyzes thinking around new approaches to management and leadership in a future that has no precedent.” The phrase “a future that has no precedent”, together with “traditional approaches to equipping graduates and others for success are struggling to keep up”, is his reason for a new mindset. He shifts his focus to “the types of mindsets and skills that emerging leaders will need”.

Readers as co-investigators, treated with humour and candour. - He asks readers to judge the papers (2026-07-19). - He lets them “vote with their feet” (2026-09-20). - He jokes that anyone who unsubscribes may be “making a point” (2026-08-02). - He warns those who see no place for joy in professional life that they “may want to call it a day and stop reading here” (2026-08-02). - He praises his co-host above himself (“I’m talking about Sean of course, not me!”, 2026-08-16).

Industry: fair-minded and constructive. Companies are “surprisingly diligent”. The pattern of risk selection is “neither accidental nor, for the most part, cynical”. “I have no reason to doubt that the frameworks address what their authors believe to be relevant and important.” He offers tools that are “freely available for using and modifying” and designed “to complement existing risk machinery rather than replace it” (2026-07-16). He still names contradictions: it “does flummox me a little as to why the people developing AI are the ones both saying they should go slower, and not doing so” (2026-09-15).

Regulators: modest, realistic asks. He does not ask for mandated coverage of every risk, which “would be unworkable in practice”. Instead regulators should require disclosure of how risks are selected: “It would simply ensure greater visibility around who is deciding what matters, and on what grounds” (2026-07-16).

Tech leaders. Gates is read generously. The AI essay leaves him “surprised — and heartened” by how it aligns with his own work, and it becomes a prompt to question his own sector (2026-08-30).

An insider critic of universities. He is candid about his own institution type (“creatures of habit, of tradition, of fiercely-defended practices”), yet keeps “at least a sliver of faith”. He is honest about his experience: “Sadly, this has been my experience so far. But there’s always hope.” He poses three open questions and admits: “I honestly do not know the answers here” (2026-08-30).

What he refuses to do: - Stoke fear. The 2018 video was made “not to stoke fears (not my style)” (2026-09-15). - Dismiss. Existential risks “should not be dismissed” (2026-09-15). - Villainise. “the patterns observed here do not necessarily need “bad actors” to explain them” (2026-07-16). - Overclaim. The orphan-risks framework is “yet to be shown to be useful in practice”. Hyperbubble is “probably over-stretching the significance”. - Take credit he did not earn. Fable is named sole author (2026-09-04). - Preach. He worries that the game description sounds “a little educational-preachy” and praises it for being “anything but serious, preachy, or overtly “educational.””

Changes of mind, reported as they happen. - His view of AI prose has swung on “something of a roller coaster”. A year earlier he was “blown away by the seeming-eloquence of models like Anthropic’s Claude 4.5”. Now he finds the prose “superficially profound yet substantively hollow” (2026-07-19). - He reports that his thinking about playful exploration “has evolved” through building and playing the game (2026-08-02).

Willing to put his reputation aside to experiment. He will play anonymously “because doing this as myself would drag a whole lot of history, assumptions and perceptions along with it (and probably a fair bit of eye-rolling)” (2026-08-23). He links this in a footnote to his short story “Letters from the Department of Intellectual Craft”. He is prepared to risk embarrassment (“At least this way I shunt any embarrassment eight months down the road!”).

6. What is distinctive#

  1. Studying what institutions leave out, and how they come to leave it out. Most AI-risk commentary argues about which risks matter most. He asks how a known risk becomes “nobody’s responsibility”. He traces the answer to the definition of risk itself, and to incentive fields acting on sincere people. - This sits between two camps. The AI-safety community works with capability thresholds and catastrophe floors. Critical outside commentary attributes the pattern to bad faith. He is inside risk science, fluent in both enterprise and social traditions, and neither accuses nor excuses.
  2. A value-based redefinition designed to be adopted. He does not start from what is theoretically best. He starts from what a fast-moving, process-averse culture will actually take up: “show it a threat to something it values”. The redefinition works both as a concept that opens up new ways of seeing and as a lightweight tool. He keeps the quantitative machinery and adds a second layer.
  3. Calm about existential risk, grounded in a long record. He can point to his own 2018 list, dated and public, and say both that “we’ve known about this stuff for years” and that “new and innovative approaches” are still needed. Few commentators have a public record of their own that goes back that far to test against.
  4. Play as a research method, with rigour attached. He treats a game, a parody website and a sealed, hash-verified envelope as serious ways to think. He calls joy “a deeply under-appreciated metric of intellectual and academic achievement”. And he pairs play with accountability (“where’s the fun — or the accountability — in that?”). This combination is rare in academic AI discourse and almost absent from AI-policy discourse.
  5. Experimenting on himself, vulnerably, in public. He runs AI-versus-human comparisons with himself as the human, publishes the unflattering verdicts (“Ouch!”), and leaves the question open. His criterion is the human experience of reading, not measurable output.
  6. Concern with formation in both directions. He worries that AI is quietly retraining human standards (“train us to think like them”). He also designs playful environments meant to form a mindset for navigating an uncertain future without prescribing what anyone learns.
  7. Honest attribution as a practice, not a principle. He names an AI as sole author when that is the truth, uses a CRediT-AI statement, and says openly where publishing systems cannot yet handle this.

7. The posts in this batch that best reveal how he thinks#

  1. 2026-07-16 orphan-risks-frontier-ai-maynard. The fullest statement in the batch of risk as threat to value, orphan risks, and navigating a risk landscape. It shows the “how are they made” reframing, the refusal of villains, the augment-not-replace position, the falsifiable tests, and a first-person narrative academic voice. (Read with the provenance caveat above.)
  2. 2026-07-19 publish-or-perish-ai-vs-human-vs-human. Self-experiment in public, genuine uncertainty held open, the reader’s experience as the value to protect, and the worry that AI is retraining human judgement.
  3. 2026-08-02 what-we-can-learn-with-ai-by-not-trying-to-learn. Play without purpose as a way to thrive, learning through delight, and a risk landscape made into something you can play. He reads his own project on several levels and gently undercuts himself.
  4. 2026-09-15 will-ai-really-kill-us-all. Demands a mechanism, audits his own 2018 forecast, and holds the middle between doom and dismissal. States that talking about risk is not fear-mongering.
  5. 2026-08-23 pre-registered-play-open-april-25. Rigour applied to play. Accountability he imposes on himself. Willingness to set aside his reputation in order to experiment.
  6. 2026-09-20 reasoning-llms-just-want-to-have-fun. The most explicit statement of method in the batch (“much of my work uses play, creativity, and serendipity”) and of joy as a scholarly value.