Late Lessons, Jensen Huang and AI
Markdown version

B30 perspective notes: 2026-05-15 to 2026-07-10 (8 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. I read all eight posts in full, including the AI-written sections, to check for any passages in his own voice inside them. There are no Modem Futura posts in this batch. Quotes are exact, including his typos (“teh”, “it’s rigor”, “thats only frame”, “and valuable because”) and curly punctuation. Markdown italics are dropped.

Evidence rules applied

Context. Mid-2026. On 17 May he announces a sabbatical running to August 2027. Pope Leo XIV publishes Magnifica Humanitas. Anthropic releases the “Mythos-class” Claude Fable 5 (the 10 June post says “Yesterday”); a US export-control directive suspends access on 12 June, and it returns on 1 July. Four of the eight posts form a series of escalating hands-on experiments with Fable 5, written from a London theatre trip, an airport, and a “one-week access window”.

1. How he thinks here#

He tests a received trope instead of repeating it or arguing with it (2026-05-15)#

The movies post opens with the cliché in its strongest form: “Everyone knows that” AI is bad news in sci-fi. The next line turns it: “Except that, deeper analysis indicates that the connection between AI and imagined futures in movies is more complex than this.” The motive is a long-standing curiosity about evidence, not a position to defend: “I’ve long been interested in how strongly the evidence supports assumptions of AI-driven dystopias in movies.”

The trigger was Anthropic’s claim that dystopian sci-fi taught its models to act “evil”. He does not argue with the claim. He goes and looks at the thing the claim assumes. He searched the literature, found nothing that addressed the question directly, “And so I sat down with Claude Code and started digging.” The result is a 169-film corpus from 1927 to 2026, covering 31 countries and 21 languages.

Three features of the move are typical of him: - He breaks the binary. He goes past “the usual triad” of dystopia, utopia and protopia and adds five more future states: continuation, inheritance, supersession, heterogeneity and agonism. The question becomes which futures AI is tied to, not whether it is good or bad. - He is drawn to what doesn’t fit. “This was a surprise to me”; “Another surprising outcome”; “The interesting lines, of course, are the counterintuitive ones”. The finding he cares about is the “complex” portrayal that is “all over the map”, where there is “no bright line between artificial intelligence and a particular future state.” - He bounds the claim at once. The categorisation is “subjective”; the study “admittedly doesn’t rise to the level of a publication-quality study yet”; the aggregate hides country trends. He names what the data cannot show: “A blockbuster Hollywood dystopian AI movie will likely have far more influence and impact on society than a small indie movie.” The biggest question, how portrayal shapes the relationship between AI, society and the future, is left as “a question for another day”.

This is Films from the Future’s instinct (films as a way into how we imagine technology) turned into an open, checkable dataset.

He reads across centuries by structure (2026-05-21)#

He does not treat Magnifica Humanitas as a stand-alone news event. He sets it beside two earlier encyclicals and maps all three onto his own schema for technology transitions: “what we do, where we live, and who we are”. Rerum Novarum (1891) dealt with work under the industrial revolution, whose “insights and ways of thinking” still apply “as AI ushers in a new era of automation”. Laudato Si’ (2015) dealt with the coupling between technology and the planet. The new encyclical is expected to fill the third domain.

Two moves follow from the structure: - He sees what others missed. The Vatican framed the document against Rerum Novarum, so the link to Laudato Si’ “has largely been overlooked”. For him it is “a vital piece of the puzzle”. - The structure locates what is new. The earlier documents do not reach “the one domain where AI is shaking things up in ways that no other technology has come close to”. The historical analogy is used to show where the analogy breaks.

He holds both sides, often in one sentence#

He rarely lets a judgement settle on one side: - AI “both presents an environmental threat and a potential pathway to developing novel solutions to persistent environmental challenges” (2026-05-21). - The technology “both challenges and opens up new ways of revealing who we are” (2026-05-21). After listing cognitive threats, he turns: “And yet there is another side to this”. - Fable “threatens to pull the rug from under every effort over the past 3 plus years to accommodate ChatGPT 3-level capabilities, and to open up new learning possibilities that we’ve barely grappled with” (2026-06-10). - On the one-shot paper: “please read with caution. But also, read the paper with the seriousness it deserves” (2026-06-10). - On frontier capability: there is a risk “of falling for the illusion that these models are more capable than they actually are”, “And yet, it would be foolish to discount emerging capabilities” (2026-06-12).

These tensions are not hedges. Each one is where he thinks the real question sits.

He builds to find out, and designs each experiment from the last one (2026-06-10 to 2026-07-10)#

The Fable posts are a sequence of experiments, each prompted by what the previous one left open: 1. One-shot prompt (2026-06-10). He deliberately does what he would normally never do. It is “something I would usually never do as one-prompting research and papers tends to lead to outputs that are superficially OK and substantially poor. But this is what made it interesting.” Failure is built in as a result: “It felt like a good first-test that, even if it failed, would be instructive.” 2. Full agentic run (2026-06-12). “I couldn’t resist the temptation to see what it could do given greater latitude, and a lot more tokens!” He sets the goal and stays hands-off on method, then watches: “Watching Fable work based on what I asked of it in Claude Code was fascinating.” 3. Partnership in his own field (2026-07-04). The experiments left him with “more questions than I started with”. So he chooses a test bed where he can judge the result: “in an area where I would have a clear sense of where it was successful, and where it wasn’t.” This is experimental design, a physicist choosing a calibrated instrument. His own ten years of risk-innovation work becomes the reference standard. 4. Play (2026-07-10). In a lull, with tokens to burn, he asks Fable for a game based on his life’s work.

He reports the costs of the experiments as part of the results. He “ironically maxed it out” on a $200-a-month plan and bought more credits. The paper took “nearly two days” and six drafts.

He watches for surprise and reports it#

Surprise is where he finds the insight: - “This is not what I expected.” Fable’s agentic preprint was “deeply reflective, with meta-layers” (2026-06-12). - “And here I have a bit of a problem, but not one you might expect.” The problem was not bad work but work too good for him to assess quickly (2026-06-12). - Fable applied his framework “in a way that hadn’t previously occurred to me” (2026-07-04). - The game offered “a completely unexpected and delightful perspective on navigating advanced technology transitions” (2026-07-10).

He reframes what the question is#

Three reframes carry the batch’s thinking about AI: - From outputs to formation (2026-06-10). “This is no longer a technology that emulates the outputs of educational and learning processes, but extends this to the formation of those outputs.” The question moves from cheating on assignments to what learning is. - From “smarter” to faster emulation of knowledge-making (2026-06-12). “Not because AI is in some way “smarter” or “better” than us.” Rather, it is “getting so good at emulating the processes through which new knowledge is constructed and tested” that humans cannot keep up. - From autonomy to augmentation (2026-07-04). Attention belongs “not so much in autonomous research and development (although I have no doubt that this is coming), but in massively-augmented research and development.”

Each reframe moves attention away from headline capability and towards what changes in human practice.

Wordplay as a way of reframing (2026-05-17)#

The title “The nonsense I write” is self-deprecating and a real argument. His worry is not that his writing lacks worth (“I know my stuff”) but that readers see no sense in it. So it is nonsense, “literally making no sense”. The pun moves value from the writer to the reader: writing has value when readers can see why it matters to them.

Curiosity, play and serendipity drive the work#

The engine of the batch is appetite. He “couldn’t resist the temptation” (2026-06-12). He posts from vacation because the release “felt too significant to ignore” (2026-06-10). On the game: “I did what any self-respecting academic would do”, “who cares when you can have this much fun with a frontier AI model”, “it still brings a smile to my face every time I play!”, and in a later footnote, “I simply couldn’t leave this alone” (2026-07-10).

The play has purpose. It is a test (“Wanting to test its capabilities in coding a simple web-based game”) and a translation. He values the game because it shows how AI can “help translate experiences, ideas, research, and a lot more into something quite unexpected and serendipitously delightful.” He takes “toy” seriously: “a concept they take very seriously”. He lets serendipity in. Fable chose the soap bubble from the last pages of Future Rising: “It was an interesting choice, but one I let it run with.” In 2020 he chose that same bubble as his object for the future: “something full of wonder and promise, but at the same time, in need of care if it’s to survive and thrive” (2020-10-22 what-if-the-future-was-an-object-fe4eac545fa3).

2. What matters to him#

3. Risk as a way of thinking#

Novel technology demands a new mindset, and he says so repeatedly#

This batch contains several plain statements that AI does not fit the categories we have: - Not a tool. His one serious objection to the encyclical: “treating a technology that has the ability to fundamentally alter how we think, act, and even believe — and in ways that surpass our comprehension — as just a tool, is potentially dangerous” (2026-05-21). The inherited category is itself the hazard. - Assumptions about being human are being reopened. “the cutting edge of AI development is beginning to force a reckoning with long-held assumptions around what it is to be human” (2026-05-21). The encyclical preserves the old understanding; he suspects this “stops short of what I suspect is needed here”. He still credits its pragmatism: it “treads a pragmatically useful path as it extends that thinking without breaking it”. - Education has to think again. The release of Mythos-class models “forces us to change how we think about teh intersection of AI and education” (2026-06-10). Three years of accommodating ChatGPT may be undercut.

Threat to value, implicitly, in several places#

He does not use the phrase “threat to value” in his own prose here, but the logic runs through the batch: - The deepest threat to higher education is to value built on scarcity: “a world built on the assumption that intelligence, expertise, and new knowledge, and valuable because they are scarce” (2026-06-12). The release questions “even what the value of a university education is” (2026-06-10). - The threat to “who we are” is a threat to things people value that are not physical: reasoning, self-understanding, trust. He places his own research here: the “cognitive trojan horse” (AI bypassing “our cognitive defense mechanisms”) and “constitutive resonance” (“a two-way coupling where both human and artificial participants are changed in the process”). He expects these to be “just the tip of a growing area of research around how AI potentially threatens who we are” (2026-05-21). - He turns the same lens on his own writing (2026-05-17). Value lies in whether it makes sense to readers, not in its intrinsic worth.

Risk innovation named as a framing (2026-07-04)#

He describes a decade of work as “my work over the past ten years on how the framing of risk innovation applies to AI frontier models”. The word is framing, not method. The [mixed] paper uses “what I’ve previously referred to as orphan risks” to find “currently sidelined risks” in the gap between companies’ internal safety frameworks and their regulatory compliance documents. It argues constructively that there are “effective and productive ways for companies to close this gap”. This is where risk works most like an operational tool in the batch. Even so, the point is to see risks that existing categories leave without a home.

Navigating is his default verb for technology transitions: “successfully navigating advanced AI transitions” (2026-05-21, describing his 2025 argument); the encyclical as a blueprint “for not only thinking about the intersection between society, technology and the future, but for actively navigating it” (2026-05-21); Fable “will increasingly challenge how we navigate the intersection of AI, expertise, and knowledge generation” (2026-06-12); the game’s “whole trajectory reflects my work on human flourishing and navigating complex advanced technology transitions” (2026-07-10). Landscape appears in the same spirit: “the landscape around AI and is changing” (2026-05-15), and the higher-education “landscape” in the 2026-06-10 title.

Risk made playable (2026-07-10)#

HYPERBUBBLE is the most literal example of risk concepts working as mental models rather than procedures. In his own account of the game: - Orphan risks are things “you adopt (and which — bizarrely — become your pets)”. - Named risks come from “the often-overlooked risks that trip up emerging technologies”. - “the moral-panic fires” come from his work on risk perception and engagement. - The player literally navigates a fragile bubble through a transition.

He values it because it “doesn’t feel in any way instructional (I hope!)”. The frameworks are experienced from the inside, not applied. The mechanics are Fable’s, but his endorsement (“Fable did a pretty good job of capturing my work”) shows that he recognises his risk thinking in a form that is exploratory, playful and non-didactic.

Risk and benefit held together#

The movies analysis tracks risk, benefit, neutral and “complex” portrayals of AI rather than a single danger axis. He finds the “complex” category the most interesting and the counterintuitive links (dystopia with beneficial AI, continuation with risky AI) the most revealing (2026-05-15). Risk is one pole in a landscape, not the whole story.

Humility against false precision#

4. Scholarship and public writing#

5. His role as he sees it#

6. What is distinctive#

  1. His own body of work is the calibration instrument. Few commentators can test frontier AI against a domain they built themselves. He does exactly that (2026-07-04, 2026-07-10) and reports both the gains and how much of his own judgement and editing the result needed.
  2. He puts the evaluation gap at the centre. The unusual worry is not that AI is wrong but that competent AI work outruns human capacity to assess it, “faster than we are currently capable of validating and even understanding” (2026-06-12). This links to his cognitive trojan horse work: fluent language “potentially capable of slipping by our critical reasoning”.
  3. Formation rather than outputs, for education (2026-06-10). Much of the AI-and-education debate is about assessment integrity. He reframes the challenge as AI entering the process by which learning forms.
  4. He tests a cultural trope with open data and leaves the hard question open (2026-05-15). He neither repeats the “AI movies are dystopic” trope nor replaces it with a counter-slogan. He widens the categories, publishes the data, and points to influence as the next question.
  5. He reads theology as a structural map of technology transitions and critiques it from sympathy (2026-05-21). Placing Rerum Novarum, Laudato Si’ and Magnifica Humanitas on a what-we-do / where-we-live / who-we-are schema, and then objecting to “just a tool”, is an unusual combination of respect and challenge.
  6. Serious play. A risk scholar turning his frameworks into an arcade game where orphan risks become pets, and calling it delightful while noting it “probably won’t do much for my academic standing”, is rare. Play works here as a test, as translation and as a way of loosening fixed frames (2026-07-10).
  7. Prose as an embodied act of care (2026-07-04). His argument about AI writing is that readers can feel care and effort, and that a disembodied system cannot know what reading feels like. The argument comes from a reader’s experience, not from a list of AI “tells”.
  8. The public scholar examines public scholarship in public (2026-05-17). He puts the costs, doubts and personal toll of the role on the page.

7. The posts that best reveal how he thinks#

  1. 2026-06-12 a-quick-update-on-using-claude-fable-5. Curiosity he can’t resist, an escalated experiment, surprise reported honestly, humility about the limits of his own expertise, the “smarter” versus emulation reframe, and the evaluation-gap risk. It holds illusion and capability together.
  2. 2026-07-04 just-how-good-is-anthropics-fable-as-a-research-assistant. Experimental design using his own field as the standard, carefully calibrated claims, risk innovation as a “framing”, care as the measure of good writing, and augmentation over autonomy.
  3. 2026-07-10 i-asked-anthropics-fable-5-to-create-a-video-game-inspired-by-my-work. Play, delight and serendipity as method. His risk concepts appear as mechanics you experience rather than tools you apply.
  4. 2026-05-21 magnifica-humanitas-and-being-human. His clearest statement of values in the batch, the three-domain structure, the novelty of “who we are”, and a respectful but firm critique of the “tool” framing.
  5. 2026-05-17 the-nonsense-i-write. How he sees his role: public writing as integral to scholarship, public good over prestige, value defined by readers, and candid self-doubt.
  6. 2026-05-15 ai-movies-may-be-less-dystopian-than-we-think. Testing a trope by building an open dataset, widening binaries, delight in the counterintuitive, and bounding his own claims.