Late Lessons, Jensen Huang and AI
Markdown version

B29 perspective notes: 2026-01-31 to 2026-05-10 (18 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. I read all eighteen texts in full. One is a Modem Futura episode post (2026-03-15), read for his framing lines only; the episode itself was not listened to. Quotes are exact, including his typos, apart from markdown italics, which are dropped.

Evidence rules applied

Context. February to May 2026. Moltbook, a “social network for AI agents”, goes viral at the end of January. “Harness engineering” becomes a buzzword in February. He has been using the desktop version of Claude Code “for a few weeks” by late March. Anthropic moves from Opus 4.5 through 4.6 to 4.7, publishes an April postmortem on Claude Code, and withholds its Mythos Preview model. He gives a keynote in Copenhagen and sees The AI Doc at CPH:DOX. He builds spoileralert.wtf, an AI-first rebuild of FFTF. ASU’s AI course product Atomic draws press criticism in late April. Richard Dawkins suggests Claude may be conscious. He makes his first Risk Bites video in about four years. Two preprints sit behind the batch: the harness paper, and a second paper on how human and AI change each other (its file name refers to “constitutive resonance”). A third, earlier piece on AI as a “cognitive Trojan Horse” (epistemic vigilance) is cited but not in the batch.

1. How he thinks here#

He starts from something small that happened to him, says it is small, and follows it until it isn’t#

Almost every substantive post begins with an encounter, and he often labels it as trivial before showing where it leads.

The movement is always outward: from his hat to “all those other occasions where the reasoning and information seemed sound”; from his own flattened profile to “anyone whose identity doesn’t fit a neat and plug-and-play category”; from one incredulous person to “hundreds of millions — probably billions” of users. The small case is a probe, not an anecdote to decorate an argument he already had.

He lets the question change under him#

He is open that the post he ends up writing is often not the one he set out to write.

Serendipity here works as method. He follows what catches him (“It was the “this is the most traditional method” that caught my attention”) and lets it reset the question.

He finds out by doing, and keeps the experiment honest#

This batch is full of small experiments, run by himself, on himself or on his own work.

He also chooses the position he experiments from. “I must confess I’m a bit of a novice user compared to some of my grad students. But I’ve intentionally kept things that way as I’m interested in seeing what users with little time or patience for engaging with technical wizardry can achieve with easily accessible AI platforms.” (2026-03-29). He deliberately stands where ordinary users stand.

He tests analogies until they break#

The Moltbook post (2026-01-31) shows this most clearly. He starts by admitting that the available frames are failing: “most observers are struggling to find appropriate analogies, metaphors, frameworks, or even language, to describe what’s happening”, and “even I am struggling to grapple with how to even describe what we are seeing, never mind how to understand it.”

He then tries analogies one after another, from complexity science and biology: Conway’s Game of Life and cellular automata, viruses, DNA strands and prions. Each shows “highly complex behavior emerges out of seeming simplicity”. Then he finds the point where the analogy fails, which is where the insight lies: “analogies with biological viruses are both helpful and deeply unhelpful. Helpful in that a virus is not technically alive, but behaves as if it is. And deeply unhelpful because a virus doesn’t instinctively know how to use every cognitive trick in the book to make us believe it’s alive.” Out of that break he makes a new image: “Are we creating self-assembling and evolving agentic AI “organoids” that aren’t alive, and yet can wreak havoc as if they are?”

He does the same with a conventional safety frame. He reaches for laboratory biosafety (“the digital equivalent of biosafety level 4 containment”), then shows in fn 4 that it cannot hold the subtler risk of bots learning to ““hack” their human observers”: “And here, they are already beyond being contained.”

He treats words as things that shape the future#

Stories and films as instruments, not illustrations#

Questions left as questions#

“Ten Questions about AI and Higher Education” (2026-04-11) is questions “that I don’t have good answers to, but I think we should be taking seriously”. He deliberately withholds the explanations: “I thought I’d simply list them as conversation-starters and see where they go.” He also marks them as a reframing. They are “not the usual suspects” (cheating, AI-proofing assignments), which “feel increasingly out of touch”. Two of the questions are about obligation rather than technique: “What do we owe our students in an age of AI?” and “What does it mean to model mature AI use?”

He holds both ends, then places himself between them, with reasons#

He implicates himself#

He does not write about being fooled as something that happens to other people.

2. What matters to him#

What makes a person who they are: idiosyncrasy, weirdness, diversity#

The LinkedInification post (2026-03-08) is his clearest statement in this batch of what he wants to protect. The loss is “compressing the amazing, wonderful richness of real people into sea of turgid grayness”, and “squeezing the sheer diversity of human identity into a few narrowly defined and, if I’m being honest, rather conventional categories.” He widens the concern at once to those most at risk of being flattened: “original thinkers, people with alternative lifestyles, anyone who lives on the edge of convention”. What is at stake is “the eccentricities, weirdness, and glorious diversity of personalities, perspectives and ideas that fuels human creativity, innovation, and meaning.”

He lives this in small ways. The Easter egg he hid for AIs was about “my obsession with towels”. His skills list has fourteen items, not ten, and he makes a point of it: “an idiosyncratically human list (or maybe just an idiosyncratically “Maynard” list)” (2026-04-14). The same value sits under the do-not-do post’s deepest worry: AI puts at stake “in some cases the very things that make us who we are” (2026-05-10), and the Risk Bites video is about not getting “so sucked in by AI that you forget who you are”.

Words, voice and the texture of writing#

“words, and how they are used, and the meaning and stories they weave, are important to me.” (2026-04-26). What he misses in the updated model is texture: the old one could “reach into my soul as a writer”, and the new one cannot escape “all those turns of phrases, the rhetorical moves, the stylistic patterns, that make some LLMs OK technical writers and appalling communicators.” He wonders if he is “developing an allergy to AI-speak”. He chooses the human product even when it may cost him: on YouTube he “intentionally went with the far less slick non-AI content. We’ll see if this kills the video!” (2026-05-10, fn 11). His header image is captioned “NOT created using AI!”

Students, and education as a relationship#

Honesty about risk, and trust inside institutions#

He wants risk talked about “openly, frankly, and in very simple terms” and “not as an add-on to promoting the benefits of AI” (2026-05-10). He cares about the gap “between what is legally allowed, and what is good practice for an enterprise and the people who work for it”, and about the goodwill that makes organisations work: “trust and good will within an organization have a massive impact on its ability to operate and achieve its goals” (2026-05-03, fn 5).

Nuance, fun and opening up possibilities#

His framing of the podcast (2026-03-15) states the values compactly. The problem is “the preponderance of loud voices telling you what to think — whether they’re pushing visions of a tech utopia or impending apocalypse”, because “simple ideas ideas repeated often work — especially when they seem to reinforce what your gut tells you is true.” The counter is nuance, since “thriving in a complex future is all about nuance”. He also wants content “that opens up possibilities rather that closes them down, that makes you smile — or even laugh out loud, that is not AI generated”. Title: “The future has never been this much fun!”

What frustrates him#

What delights him#

3. Risk as a way of thinking#

The explicit statement: risk innovation as navigating threats to value in a complex landscape#

The clearest self-description comes almost as an aside (2026-05-03, fn 4): “my work some years ago on risk innovation was motivated in part by the challenges of introducing complex technologies into an equally complex stakeholder landscape by providing organizations with simple tools for identifying and navigating potential threats to value.”

The post shows the idea at work. The damage from ASU Atomic, as he reads it, is not a conventional hazard. It is to faculty trust, to goodwill, and to the university’s own AI ambitions, with coverage that “could potentially undermine ASU’s use of AI”. He separates compliance from value: using course material was “perfectly allowable” under the terms faculty work under, “But, of course, there’s often a gap between what is legally allowed, and what is good practice”. He also finds a less visible threat: a mismatch between Atomic’s assumed transmission model and the pedagogy of the teachers whose clips it reuses, and missing context, provenance and feedback. Here the principles are tools (“useful as a tool for aligning AI use with institutional ambitions, while avoiding unnecessary mis-steps”), with the condition that makes tools matter: “But only, of course, if they are actually used.”

Threat to value, applied to AI: what is at risk is identity, voice, trust and relationship#

Across the batch the “risks” he picks out are threats to things people value that conventional frameworks tend not to count. That is the logic of orphan risks, although he does not use the term here.

A new kind of technology needs a new frame#

Several posts argue, in different registers, that inherited mental models will mislead us about AI.

Risk science built on, not discarded#

The foundations here are the science of risk perception and risk communication.

The “tools” are rules of thumb: plain, personal, portable (“print off and paste by their computer”), and offered without dogma (“even if you don’t use this particular list, at least develop one that is useful to you”). There are no numbers anywhere in the batch. Even Moltbook’s scale is reported with doubt (“Not sure whether to believe this”).

Humility against false certainty#

Tools or mental models?#

In this batch, risk thinking works mainly as a lens that brings neglected values into view: identity, voice, trust, relationship and students’ futures. It also works as a stance for two-sided uncertainty, where both reckless use and refusal are risky. When he offers tools (rules of thumb, design principles, “I can” skills), they are deliberately simple and adaptable, and he treats them as useless unless people actually use them. The harness post is the purest case of risk as mindset. There the hazard is the frame itself, because it could leave the most important effects invisible.

4. Scholarship and public writing#

Preprints and posts feed each other#

Experimenting in public, with the materials open#

He publishes the raw materials: the full 223-page AI-generated degree proposal, his exact prompt (“bad grammar and all — and I’ve only just spotted “curse-specific learning objectives!””), the site’s files on GitHub, and the llms.txt. He hands readers the experiment: “feel free to point your AI to http://andrewmaynard.net and ask it about my obsession with towels!” He re-checks before posting (“Just before pressing publish I checked the figures”, 2026-01-31) and updates posts afterwards (the NotebookLM update on 2026-04-02).

Evidence and expertise: claimed when it counts, limited when it doesn’t#

He invokes his expertise directly when a judgement rests on it: “take it from me as someone who does this for a living, this is impressive”, backed by “well over a decade of teaching, developing courses and programs, and academic leadership” (2026-03-29). Also “from my perspective as an educator working at the edge of emerging tech and the future” (2026-04-14), and “As someone who’s studied, published on, and written about risk for most of my professional career” (2026-05-10). He flags weak evidence just as plainly: “if sources are to be believed (and already it’s hard to separate fact from fiction here)”; “Not sure whether to believe this”. When Claude gave him a method, he asked for a source: “Can you provide me with a link to a website describing the beeswax method?”

He also treats a new term’s history as something to trace. The harness post dates each step of the term’s spread through practitioners’ blogs and company posts before drawing on Verbeek, Clark and Chalmers, Tobias Rees and Aristotle. That is a scholar’s genealogy of a buzzword only weeks old.

Transdisciplinary by reflex#

In one batch he draws on complexity science (cellular automata), molecular biology and biosafety (prions, organoids, BSL-4), nanotechnology (“gray goo”), philosophy of technology, risk communication, pedagogy and program design, film (Fantasia, The AI Doc), speculative fiction, and web architecture (llms.txt). His degree prompt specifies “a venture between engineering and a business school that also intersects with arts and humanities”.

Accessibility as a scholarly commitment#

Transparency about his own use of AI#

He discloses as a matter of practice: the AI Use Statement on 2026-04-14 (“apart from that, what you see is what a human produced!”); the Gemini image that took “more time than it took to conceive, research, draft, and write the whole piece!”; Claude as grammar checker (2026-04-26, fn 2); the split of labour on spoileralert.wtf (“the feel, functionality and purpose of the site, as well as the type of content, all come from me”).

Self-conscious craft#

He writes about the writing. Placing the video early in 2026-05-10 is “an excruciating fingernails-on-a-blackboard moment for me as it completely messes up the flow. But sometimes you just have to suck it up as a writer” (fn 3). He breaks attention-economy norms on purpose and says so. On the podcast post: “No intro, no explanation, no exposition.”

5. His role as he sees it#

A scholar who uses the thing, from the user’s side#

He positions himself as someone who studies and communicates risk, teaches, and writes, and who works with the tools daily. He takes the ordinary user’s vantage point on purpose (the “novice user” choice). His stake is personal as well as professional: he has a “relationship with the LLMs I use”, built on “how it feels to use them, and how using them makes me feel”.

Insider critic of his own institution, and of his own committee’s work#

He sat on the ethics committee that wrote ASU’s AI principles and opens by admitting the problem may lie with them: “may not have been as useful as we’d hoped” (2026-05-03). A week later he says plainly that open risk conversations are “near-impossible” at his university (2026-05-10). In 2026-03-29 he criticises research-university teaching (“many faculty are thrust in front of student with no training on how to teach”) while counting himself in: “our own traditions and egos”. He is willing to provoke his own colleagues, “I’m sure some readers will disagree with me here — on principle if for no other reason”, when he thinks students’ interests are at stake.

With readers: conversation partners and co-experimenters#

He asks for comments and means it (“please feel free to weigh in in the comments”; “let me know how you get on”; “If you have others that aren’t here … please do add them”). He anticipates objections with humour rather than defensiveness: “I can imagine you rolling your eyes and mumbling something like “get over yourself,” or something much stronger! And you’d be right to.” Also: “forget the money bit, as this guarantees a slew of people proving me wrong and demanding payment!” He gives readers things to try (the towel test, the spoileralert prompt) and rules to adapt: “please make use of it (or don’t) as you will”.

With those he criticises: fair first, then critical#

What he refuses to do#

Changes of mind and held tensions#

6. What is distinctive#

7. The posts that best reveal how he thinks#

  1. 2026-05-10 do-not-do-this-with-ai. His role as a risk communicator in full: safety message first, grounded in risk science; the threat to “the very things that make us who we are”; navigating as the route to benefits; personal, adaptable rules; refusing both attention-economy norms and the Luddite label; open criticism of his own institution; self-implication.
  2. 2026-02-08 beeswax-hallucinations-and-ai-inventions. The whole method in miniature: a trivial encounter, being fooled while writing about being fooled, a hands-on experiment, curiosity about invention, and delight in a “storied” object.
  3. 2026-02-22 what-we-miss-when-we-talk-about-ai-harnesses. A novel technology needs a new frame. Metaphors are shown as mindsets that can hide the most consequential effects, and he moves from command and control to relationship. Preprint and post work together.
  4. 2026-04-11 ten-questions-about-ai-and-higher. Film as an instrument pushed past cliché, questions left as questions, humility about capabilities “that transcends our understanding”, and the two-sided risk of the Mickey inversion.
  5. 2026-03-08 ai-linkedinification. Playful self-experiment (the towels), idiosyncrasy and diversity as the value at stake, a coined word, and a post that grows more serious than he intended.
  6. 2026-04-26 why-im-falling-out-of-love-with-claude. Relationship without anthropomorphism, “character constancy” as an unnoticed risk, words as something he cares about, and a change of mind in a footnote.

Also revealing: 2026-05-03 (risk innovation defined in his own words as navigating threats to value, applied to his own university with fairness), 2026-01-31 (testing analogies until they break when the language runs out), 2026-03-29 (experimenting as a deliberate novice, and “what we owe our students”) and 2026-04-02 (experimental publishing, accessibility and the joy of making).