Late Lessons, Jensen Huang and AI
Markdown version

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

Reading notes on Andrew Maynard’s Substack posts in batch B29. Only his own prose counts as evidence. Quotes are exact, including his typos, italics markers and curly punctuation.

Batch 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. Anthropic moves from Claude Opus 4.5 through 4.6 to 4.7, and publishes a postmortem on Claude Code problems in April. It also withholds its “Mythos Preview” model from public release. Maynard is now working daily with the desktop version of Claude Code. AI and the Art of Being Human (co-written with Jeff Abbott) appears as a Pocket Edition and as two free “AI-readable” files. He rebuilds Films from the Future (2018) as an AI-first website, spoileralert.wtf. ASU’s AI course platform “Atomic” draws press criticism. Richard Dawkins suggests Claude may be conscious. He makes his first Risk Bites video in about four years. The batch leans towards AI and the self, identity, relationship, learning and risk communication. It is also the batch in which he most openly criticises his own university’s AI culture. He cites two companion preprints that are not in the corpus: “Is AI a cognitive Trojan horse” (on epistemic vigilance) and a “constitutive resonance” preprint (on how humans and AI change each other).

Relevance summary:

Date Slug Relevance
2026-01-31 lost-in-the-moltbook-hall-of-mirrors high
2026-02-08 beeswax-hallucinations-and-ai-inventions medium
2026-02-11 soul-update medium
2026-02-14 the-ai-book-i-actually-carry-with low
2026-02-17 how-do-you-do-ai-companion-ai-and-the-art-of-being-human low
2026-02-22 what-we-miss-when-we-talk-about-ai-harnesses high
2026-02-27 why-were-giving-away-our-book-on-thriving-with-ai low
2026-03-08 ai-linkedinification high
2026-03-15 the-future-has-never-been-this-much none (Modem Futura podcast post; skipped per user instruction)
2026-03-22 are-you-an-ai-apocaloptimist medium
2026-03-29 can-ai-create-an-undergraduate-degree-plan high
2026-04-02 spoiler-alert-wtf medium
2026-04-05 using-ai-without-losing-the-best none (no text in mirror)
2026-04-11 ten-questions-about-ai-and-higher high
2026-04-14 14-essential-ai-i-skills-for-students medium
2026-04-26 why-im-falling-out-of-love-with-claude high
2026-05-03 are-design-principles-for-responsible high
2026-05-10 do-not-do-this-with-ai high

HIGH#

2026-05-10 — do-not-do-this-with-ai — “Do not do this with AI!”#

Provenance. His own prose throughout: the italic preface, the preamble, both five-point lists, the final word and all 12 notes. The header image is marked “NOT created using AI!”. The Risk Bites video is embedded; he says the “do not do” list “forms the foundation” of it, but its script is not in the text. Punya Mishra is credited with two ideas: in fn4, that people take computers to be accurate; in fn8, that different mechanisms may let AI past different people’s defences. Those ideas are credited to Mishra, but the prose is Maynard’s. Richard Dawkins’s claim is reported, not quoted at length.

Argument in his terms. - The core claim: AI risk communication is badly missing, and it is urgent. “hundreds of millions — probably billions” of users have “no idea that AI makes stuff up”. The trigger was a conversation with someone “not … easily fooled” who met the idea with “utter incredulity”. His argument is that we must talk “openly, frankly, and in very simple terms” about unhealthy ways of using LLMs, “not as an add-on to promoting the benefits of AI, but as something that every user knows and understands”. - Why AI is different: it gets into the mind unseen. He sets aside the “just another tool” view (a “fancy calculator”). AI is “the first technology of it’s kind we’ve created that has the ability to slip unawares into our mind and change how we think”. It is also “a technology that doesn’t seem to respect educational attainments or intellect as it weaves its way into our heads”. The Dawkins episode is his example. Dawkins’s belief that Claude (“Claudine”) might be conscious rested, as Maynard reads it, on “how appreciative Claude … was of his intellect”. - Frictionless access spreads faster than understanding. Access to AI apps “which feel self-affirming, helpful, and knowledgeable, appears to be spreading far faster than any notion of understanding around how to use these technologies in safe and healthy ways.” He defines friction in fn6: “the effort required to think critically about how its used is vastly greater than the effort required to just use it.” In fn11 he notes that AI now arrives by default: in many settings “users need to make a concerted effort not to use it”. His own example is AI-generated YouTube thumbnails, titles and descriptions. He also admits the nagging thought that they might be “better than my own work”. - Put the safety message first. “everything we know about human behavior and risk management tells us that putting the safety message first is necessary — because while the benefits of a powerful technology are often self-evident, the risks are not.” He compares AI with a chain saw and a car, which do not advertise their power in neon while hiding the risks in “8-point font”. Without such guidance, “that beckoning text box feels about as risky as soap bubble on a summer’s day.” - AI literacy alone will not do it (fn5). It is tempting to think people who know how LLMs work will use them safely. But “nothing in what we know about risk behavior and risk communication suggests that this will be the case.” This is a risk-communication claim, and it rejects the standard education fix. - Downplaying is institutional. Narratives “from developers, employers, educators, and beyond” stress transformative power and push risks into “an easy-to-overlook footnote”. He notes the irony of saying this in a footnote (fn9). Ignoring these risks “verges on the irresponsible”. He names his own university: “Even in my own institution, it’s near-impossible to have an open and honest conversation around potential risks.” Such conversations are “drowned out by the the clarion call of AI acceleration”. Talking about AI risk “gets you branded as a technology-pessimist, or even a Luddite.” - The drug analogy. Research on cognitive effects, dependency and anthropomorphic relationships points to “impacts where, if AI was a drug, we’d probably be thinking carefully about how access is overseen.” This is the nearest he comes here to calling for regulation of access. It is framed as a hypothetical, not a proposal. - The five “do not” rules of thumb. He calls them personal because they are things he has to watch in himself: 1. do not trust AI “just because it feels like you should”; 2. do not treat AI as a friend or a person, and do not give it a name or a gender (“this is something AI is designed to do, not something it is”); 3. do not assume AI understands the world as you do; 4. do not assume that getting AI to think for you makes you smarter (“the illusion of learning rather than actual learning”); 5. do not assume you are too smart to be fooled (“Some of the people most likely to fall into the trap of putting too much trust in AI are those who think they are clever enough to avoid this.”). - The five “do” counterpoints. 1. Treat AI output as a starting point, and check it against a source “that doesn’t depend on the AI being right”. 2. Remember it is a machine. This “keeps you in charge of the relationship, rather than the other way round”. 3. Bring your lived experience (“A body, a history, real-world experience, a stake in how things turn out”). 4. Use AI as a thinking partner, and test your learning by explaining it to someone “without the aid of an AI”. 5. Stay alert when AI tells you what you wanted to hear.

Risk communication is “as much about doing as not doing.” - What is exceptional about AI risk. It is “how invisible many of them are, what is at stake — in some cases the very things that make us who we are — and how many people are immersing themselves in the technology with no understanding”. He sets this beside a general principle: “understanding and navigating risk is absolutely essential to reaping the long-term benefits of any powerful technology. It always has been. And AI is no exception.” - Establishing his authority. fn12 insists that his writing is “grounded in a long career in relevant research and scholarship”. fn1 points to his 2020 paper on the Risk Bites process.

How firmly. This is one of his firmest and most urgent AI-risk statements. He speaks as a risk professional (“I am finding it increasingly hard to wrap my head around how little we talk about how to use AI safely”). He hedges on the specific rules (“my personal list … some of these will be controversial”), not on the underlying claim. He keeps his balance on purpose. He acknowledges “the profound potential of emerging AI capabilities to be used for good” and apologises “to the attention economy gods for having the temerity to be balanced”.

Concepts. Frictionless access (defined in fn6). Less obvious, invisible AI risks. Safety message first. Rules of thumb as risk communication. The illusion of learning. Anthropomorphism as designed. The limits of AI literacy. The “AI Zombie” (the video’s framing). The mental model that computers are accurate (fn4, credited to Mishra). Several mechanisms of cognitive bypass (fn8).

Analogies. Chain saw and car: used structurally, for how safety information is conventionally placed. Drug and access oversight: a conceptual regulatory analogy, stated hypothetically. Calculator and computer accuracy: a mental model users bring. He uses none of the classic chemicals, nano or GMO comparisons here. The frame is general risk management and risk communication.

Views. - AI: a machine that uses language designed to feel confident, personable and trustworthy, and that gives “plausible-sounding responses rather than those that are necessarily accurate”. It has no lived experience. - AI risk: mainly cognitive, epistemic and relational (trust, dependency, anthropomorphism, the illusion of learning, loss of self). It is invisible, affects billions, and is serious. - Companies: developers are among those who downplay risk. The apps are built to be frictionless and flattering. - Governance: he hints at oversight of access (the drug analogy), but his main lever is plain-language risk communication aimed at users. He sees institutional cultures, including his own, as obstacles. - Cognition and formation: AI changes “how we think, act, and understand the world” without our noticing, and puts “the very things that make us who we are” at stake.

Criticises and engages. AI developers, employers and educators who downplay risk. The acceleration culture at ASU. The attention economy. Naive faith in AI literacy. Dawkins, as the cautionary case. He engages Punya Mishra and cites his own “cognitive Trojan horse” preprint (fn7) and AI and the Art of Being Human (fn10).

Quotes. - “the first technology of it’s kind we’ve created that has the ability to slip unawares into our mind and change how we think” - “while the benefits of a powerful technology are often self-evident, the risks are not” - “if AI was a drug, we’d probably be thinking carefully about how access is overseen” - “Even in my own institution, it’s near-impossible to have an open and honest conversation around potential risks”

2026-02-22 — what-we-miss-when-we-talk-about-ai-harnesses — “What we miss when we talk about “AI Harnesses”“#

Provenance. His own prose. The Midjourney image is not his. The post summarises a new SSRN preprint of his on the harness metaphor, and it links to a second preprint in progress (the file name refers to “constitutive resonance”). The post does not say how either preprint was written. The three-part list of presuppositions and the closing criteria (“bidirectionality”, “transformation as intrinsic to capability”) are in his post prose, but they probably follow the preprint’s wording.

Argument in his terms. - Metaphors shape trajectories. “Harness” and “harness engineering” spread fast. He traces the terms through EleutherAI (2020), Anthropic (November 2025), Aakash Gupta, Phil Schmid, Mitchell Hashimoto (February 2026), OpenAI’s Codex post and Ethan Mollick’s “Models, Apps, and Harnesses”. That speed “risks locking us into a trajectory that comes with unintended consequences as it defines how we think about our relationship with AI”. His general claim is that “words often have power that goes beyond their intended meaning”. And “metaphors are never completely neutral”. They “constrain and even taint our thinking”, leading us to treat the new as the old and “limiting future possibilities by embedding a priori assumptions into emerging capabilities”. - What the harness presupposes. A harness goes on a working animal. “It assumes that the entity being harnessed is valuable for its strength but cannot be trusted with its own direction.” The controller designs it, and the harnessed entity has no say. From this he draws three built-in assumptions: 1. A clean split between controller and controlled. Judgement and meta-judgement stay with humans: “the AI contributes capability, but not understanding.” 2. Capability can be separated from transformation. The user is meant to come out of the interaction unchanged, so any change is a side effect to be minimised. He argues instead, pointing to his forthcoming preprint, for seeing “the AI-human relationship as one that, by its very nature, influences and changes both AI and human in the process.” 3. AI as an instrument. This framing goes back to Aristotle’s physis and techne and persists in the claim that AI is “just a tool”. He cites Tobias Rees (“a nostalgia for human exceptionalism”), Verbeek’s mediation theory, and Clark and Chalmers’s extended mind: “as they are “harnessed” they alter the harnesser”. He argues this is “substantially amplified in emerging frontier AI systems”. - What he proposes. Frameworks should allow for “bidirectionality (the user is also changed), transformation as intrinsic to capability (not a side effect to be prevented)”. They should also allow for “the possibility that the most consequential effects of human–AI interaction may be invisible from within a paradigm optimized for task performance”. And they should leave room for the relationship itself to evolve: “we need to be thinking more about working in relationship with emerging AI technologies, rather than approaching them as something to be commanded and controlled.”

How firmly. He is open about whether the metaphor is benign (“It may be … useful and relatively benign”). He is firm that it carries built-in assumptions and that “some intentionality may be in order” before we “get stuck in a rut”. He puts the relational stance in the first person (“as I would argue”).

Concepts. The harness metaphor and what it presupposes. Bidirectionality. Transformation as intrinsic to capability. Effects invisible to a task-performance paradigm. Relationship versus command-and-control. Mutual constitution of human and AI (the constitutive-resonance preprint). Technological mediation. The extended mind.

Analogies. Working animals and the harness, which he takes apart as a metaphor. There is no comparison with past technologies. His point is that treating AI as an old kind of technology is itself a mistake.

Views. - AI: frontier systems “increasingly display characteristics we associate with understanding, judgment, and even autonomy”. They are more than tools. They reshape the “cognitive and experiential landscape”. - AI risk: the risk lies in framing. A control-oriented vocabulary may hide the most important effects on humans. - Companies and practitioners: he engages Anthropic, OpenAI, Hashimoto and Mollick as the people spreading the term. He criticises the vocabulary, not the people. - Cognition and formation: the user is changed by the interaction, and that change is part of the capability, not an accident.

Change of view. This is his most explicit rejection so far of the “just a tool” framing, and a move towards relationship. It sits in tension with rule 2 of the 05-10 post (“Do remember that you’re working with a machine, not talking to a person”). He appears to reconcile the two by treating the relationship as real and consequential while denying personhood.

Quotes. - “It assumes that the entity being harnessed is valuable for its strength but cannot be trusted with its own direction.” - “In other words, the AI contributes capability, but not understanding.” - “as they are “harnessed” they alter the harnesser” - “we need to be thinking more about working in relationship with emerging AI technologies, rather than approaching them as something to be commanded and controlled”

2026-04-26 — why-im-falling-out-of-love-with-claude — “Why I’m falling out of love with Claude”#

Provenance. His own prose. The MidJourney image is not his. The linked HTML “Opus 4.7 / Mythos timeline” page was produced by Claude (Opus 4.7) for Sean Leahy; he links to it as an example of AI-speak and did not write it. fn2 says Claude checked the final draft for grammar. The post contains an important provenance statement about AI and the Art of Being Human: “The book was a very intentional collaboration with AI”. Claude (Opus 4.5) was the main writing partner after ChatGPT proved “flat, mechanical”.

Argument in his terms. - A relationship without anthropomorphism. “I strenuously avoid anthropomorphizing AI apps.” He does not name them, he calls them “it”, and he has no illusions of consciousness (“at least, not yet”). “But I do have a relationship with the LLMs I use”, built on “how it feels to use them, and how using them makes me feel”. - Character, and what an update did to it. With Opus 4.5 he found a partner that seemed “to reach into my soul as a writer”. “Claude plus me had more style, more depth, and more emotional resonance (surprisingly) than either of us could achieve alone.” With Opus 4.7 (the problems began with 4.6, per fn1) this became “a soulless machine that seemed incapable of breaking away from interminable AI clichés”. The more he tried to train it on his voice, “the worse it got”. - The general claim: AI is a relational technology. “This is a technology whose transformative power is predicated on how we engage with it relationally.” What matters is “how the AI behaves, not just what it does”. Changes to a model therefore stress “connection and trust”. He sets this against ordinary attachment to cars or sweaters. AI is different because it “can talk back to you, reason with you, share your hopes and fears (or at least do a good impersonation of this)”. - A change of view inside the post. In the main text he allows that for “purely utilitarian” applications a change of character probably does not matter. fn3 reverses this: “Re-reading this, I’m not so sure, as LLMs are a relational technology whether we’re doing math with them, coding, solving the next big scientific problem, or asking what we should make for dinner.” Language shapes how we respond, how far we trust, and what we can get out of these systems. - What companies should do. “It might be worth the big AI companies paying more attention to both character and character constancy”. This risk “risks going unnoticed under the dazzle of technical AI fireworks”. He connects it, tentatively, to Anthropic’s April postmortem on Claude Code. He suggests he may be developing “an allergy to AI-speak”.

How firmly. Self-deprecating (“I may be being a little over-sensitive”, “get over yourself”) and uncertain (“Whether this is a big deal or not, I honestly don’t know. But I suspect it might be”). He is firmer in fn3.

Concepts. Relational technology. Character. Character constancy. The human–AI relationship as the basis of trust. AI clichés and “AI-speak”. A writing partnership.

Analogies. Human–human, human–animal, human–plant and human–machine relationships, used conceptually for how trust breaks when the other party changes. Cars and favourite sweaters, used as a contrast.

Views. - AI: relational, and its “character” is shaped by the developer’s tuning and can change without warning. - AI risk: discontinuity in character undermines trust and the ability to work. It is under-noticed. - Companies: Anthropic is criticised directly (“Anthropic, what have you done”). - Formation: the character of a model’s language shapes a writer’s work and feelings. - Self-disclosure: he has depended on Claude as a co-writer for about a year.

Quotes. - “I strenuously avoid anthropomorphizing AI apps.” - “this is a technology whose transformative power is predicated on how we engage with it relationally” - “paying more attention to both character and character constancy as increasingly powerful AI models emerge” - “LLMs are a relational technology whether we’re doing math with them, coding, solving the next big scientific problem, or asking what we should make for dinner”

2026-03-08 — ai-linkedinification — “Is AI reducing you to a LinkedIn stereotype?”#

Provenance. His own prose. The image is his edit, made with Nano Banana 2, of Apple’s 1984 advert. Claude’s quoted explanation of the “towel” Easter egg in the Postscript is Claude’s output. He calls the post “a bit of a rant … not as deeply researched as it probably should be”.

Argument in his terms. - The experiment. He added hidden, AI-readable, unconventional information to andrewmaynard.net (about his “obsession with towels”) and asked LLMs to profile him. Most models, Claude included, ignored it and produced “a super-boring LinkedIn-style profile”. In the Postscript, Gemini (Thinking) and DeepSeek picked up the hidden text, Grok partly did, and ChatGPT 5.2 and Claude Opus 4.6 did not (“Flatter than a pancake!”). - The mechanism. LLMs are trained to fit responses to “well worn conventions”. There are probably many “conventional response” templates, and they “reflect baked-in biases that are often hidden in their honey-tongued prose”. - The stakes: flattened identity and hidden steering of norms. It squeezes “the sheer diversity of human identity into a few narrowly defined and, if I’m being honest, rather conventional categories.” He asks: “how do LLM-based AIs reflect original thinkers, people with alternative lifestyles, anyone who lives on the edge of convention”. The flattening is “suggestive of a largely-hidden AI hand promoting specific social norms and expectations and, by extension, behaviors.” - LinkedInification versus enshittification. Doctorow’s term describes the degrading of products and services. “My fear is that this “LinkedInification” degrades people.” It strips out “the eccentricities, weirdness, and glorious diversity” that fuel “human creativity, innovation, and meaning”. The feared end state is “a nebulous gray goo of conventionality”.

How firmly. The worry is sincere, and he escalates it (“more serious than I’d originally intended”). He admits the post is thinly researched and “trivial” as an experiment. He hopes future AI will “celebrate human diversity”.

Concepts. LinkedInification, his coinage: AI flattening identity into professional or conventional stereotypes and so degrading people. “Conventional response” templates. The hidden AI hand promoting norms and behaviours. AI-legible web content (llms.txt, hidden text).

Analogies. Doctorow’s enshittification (conceptual contrast). “Gray goo”, the nanotechnology doom trope, used purely as a metaphor for homogenisation, not as a comparison with nanotech risk. Apple’s 1984 advert (the image) evokes conformity.

Views. AI embeds social norms and quietly reinforces them, which amounts to soft normative steering and a threat to diversity and identity. People on the margins of convention are the most exposed, which gives the post an equity dimension.

Quotes. - “my fear is that this “LinkedInification” degrades people.” - “it’s suggestive of a largely-hidden AI hand promoting specific social norms and expectations and, by extension, behaviors” - “a nebulous gray goo of conventionality”

2026-05-03 — are-design-principles-for-responsible — “Are design principles for responsible and beneficial AI useful?”#

Provenance. His own prose. The six ASU AI design principles, which he paraphrases from ASU’s published list, were produced by a committee he sat on in 2024. Michael Crow’s description of Atomic is a quotation. He points to Alex Halavais’s essay on the “atomized professor”. The image is Gemini/Maynard.

Argument in his terms. - Context. ASU’s six principles for responsible and beneficial AI were published in August 2025. They cover amplifying possibilities while respecting autonomy, bringing the best of technology while aware of risks, rigorous pre-release evaluation for harm, equity, privacy, and shared responsibility. They are meant as a daily decision guide and “an accountability framework”. - The case. ASU Atomic, a $5/month beta, scrapes online course material and uses AI to assemble custom modules, including instructor video clips. 404 Media, the Chronicle and Inside Higher Ed reported faculty concern. - His critique of the execution. The generated modules give no provenance: no source, no validation, no indication of what is scraped or generated, no course, instructor or context for clips, no check on currency or intended audience, and no fit with the module’s pedagogy. Instructors did not know their material was being used. There is no feedback mechanism. - His critique of the model of education. Atomic makes sense only “if you buy into the transmission model of education that focuses on optimizing content-transfer, rather than models where learning emerges through experience, dialogue, and reflection”. It also works only if the source teachers share that model. fn2 says his own teaching is “more relational and experience-based than purely transactional”. - Legal is not the same as good. Using the material without consultation was permitted under the LMS terms and faculty contracts, “but … there’s often a gap between what is legally allowed, and what is good practice for an enterprise and the people who work for it”. fn5 adds that trust and goodwill have “a massive impact” on an organisation’s ability to function. - Judged against the principles. It is not clear that Atomic respected autonomy, was rigorously evaluated for harm before release, or reflected shared responsibility. His diagnosis is charitable: steps may have been “inadvertently overlooked” in “a large, complex and fast-moving organization”. - Conclusion. He thinks the principles could have prevented the mis-step, so they are “useful as a tool for aligning AI use with institutional ambitions, while avoiding unnecessary mis-steps. But only, of course, if they are actually used.” - The risk innovation link (fn4). “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.” Here he explicitly ties the Atomic episode to his earlier framework: threats to value, stakeholder landscapes and trust.

How firmly. Measured and diplomatic about his own university. He is firm that the principles were not visibly applied and that legality is not good practice.

Concepts. Design principles as an accountability framework. The gap between legal and good practice. The transmission model versus relational and experiential learning. Provenance and context of content. Shared responsibility. Risk innovation and threats to value (fn4). Trust and goodwill as organisational assets.

Views. Governance happens at the institutional level through principles, and they work only when used in practice. The stakeholders who matter include faculty whose work is used. He is critical of his institution’s speed and top-down deployment. The episode also shows that “responsible AI” commitments can fail inside the organisations that wrote them.

Quotes. - “there’s often a gap between what is legally allowed, and what is good practice for an enterprise and the people who work for it” - “such principles are useful as a tool for aligning AI use with institutional ambitions, while avoiding unnecessary mis-steps” - “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”

2026-04-11 — ten-questions-about-ai-and-higher — “Ten Questions about AI and Higher Education”#

Provenance. His own prose. Disney stills are credited.

Argument in his terms. - Film as a lens. He uses the Sorcerer’s Apprentice scene in Fantasia (1940) “much as I used science fiction movies in the book Films from the Future”. He admits it is a cliché (“the lure of lazy technology-enabled shortcuts; the seduction of frictionless power; the myopia of starting something without thinking about the consequences”). Read more deeply, he says, it opens questions beyond the obvious cautionary tale. He says a longer project using the film is under way. - The ten questions. 1. What does competency mean? 2. What does success mean? 3. How do we help students avoid “the illusion of understanding and ability”? 4. What happens when students become AI masters and teachers AI apprentices? 5. What is the value of pursuing mastery without AI? 6. What do we owe students? 7. What does it mean to model mature AI use? 8. How do we navigate AI-enabled efficiency? 9. How do students stretch their imagination “without being subsumed by the technology”? 10. How do students discover “what it means to be human in an age of AI”? - The usual debates are out of date. Cheating, AI-proofing assignments and chatbots in class now feel “increasingly out of touch” given agentic systems such as Codex and Claude Code, human-indistinguishable prose, and agents that can do students’ coursework or produce “thousands of hours of human-equivalent work overnight”. Anthropic’s decision to withhold Mythos Preview “because it’s too powerful” makes the single-prompt-window picture “deeply anachronistic”. - Not AGI, but power beyond comprehension. “I am not talking about artificial general intelligence (AGI) or “superintelligence” here — both of which are rather ill-defined concepts.” What he is talking about is “transformative technological capabilities that we simply cannot comprehend the full capabilities of, and yet already offer near-frictionless access to power that transcends our understanding.” These capabilities are “deeply intertwined with — and influential on — how we think, how we learn”, and with how we understand ourselves and imagine futures. - The stakes for universities. They must take AI seriously “either as an existential threat or as an opportunity unlike any that we’ve previously faced” or risk “being swept away in the coming AI tsunami”. - The eleventh question: hubris runs in both directions. “Mickey’s mistake wasn’t only that he embraced sorcery he didn’t understand — it’s that he acted with certainty unshaken by his ignorance.” The modern equivalent may be “trusting that traditional mastery alone is enough, while the water keeps rising”.

How firmly. The questions are open by design (“I don’t have good answers to”). He is firm that the conversation in higher education is out of date, and that both naive adoption and complacent refusal are forms of hubris.

Concepts. The illusion of understanding. Frictionless power. Power that transcends understanding. AI master and apprentice inversion. Mature AI use. Two-sided hubris. Rejection of AGI/superintelligence as ill-defined.

Analogies. The Sorcerer’s Apprentice, used structurally for power without understanding. The “AI tsunami” and “the water keeps rising” as metaphors.

Quotes. - “transformative technological capabilities that we simply cannot comprehend the full capabilities of, and yet already offer near-frictionless access to power that transcends our understanding” - “Mickey’s mistake wasn’t only that he embraced sorcery he didn’t understand — it’s that he acted with certainty unshaken by his ignorance.” - “I am not talking about artificial general intelligence (AGI) or “superintelligence” here — both of which are rather ill-defined concepts”

2026-03-29 — can-ai-create-an-undergraduate-degree-plan — “Can AI create a comprehensive degree program proposal in the time it takes to grab a coffee?”#

Provenance. His own prose, including the prompt he quotes. The 223-page degree proposal (linked, not in the text) was generated by Claude Code, with sub-agents and Claude Code reviews, and then edited by him. That proposal is AI output and not evidence of his views, although he endorses its quality.

Argument in his terms. - The experiment. He gave Claude Code (desktop) a prompt to design an undergraduate degree on “navigating technology transitions in a technologically complex world”. This topic revisits his own earlier idea for a degree on “advanced technology transitions”. Claude Code launched sub-agents, then ran four review personas (pedagogical, employers, students, parents). After revisions and his own edits, the result was about 66,000 words from 334 files and 7 sub-agents. The AI worked for about an hour, and his editing took about ten times as long. - His verdict. Based on “well over a decade of teaching, developing courses and programs, and academic leadership”, the result, “while far from perfect, far surpasses most degree-planning documents I have seen emerge from more conventional processes”. He separates multi-agent systems from single agents and bots, which “are adept at producing content that looks good, but is not”. He expects some readers to disagree “on principle”. - Human direction is decisive. The proposal is good “not because Claude Code in isolation knew what a great degree program looks like, but because I was able to provide expert direction, feedback and evaluation along the way.” These are tools that “don’t replace humans, but rather vastly enhance their professional capabilities.” - Critique of the academy: “What we owe our students”. Research universities use a “trickle down” model. Faculty are “thrust in front of student with no training on how to teach”, sit on degree committees without design knowledge, and build career pathways “having never had a career outside of academia”. The system is “functional but not necessarily optimal”. AI could help make sure that “student success comes before academic hubris”. Degrees are a “life-investment” that many students “can barely afford”. He dismisses appeals to “the sanctity of human intellectual labor and the inviolable standing of academics” and warns that students “are going to start voting with their feet”. - The opening question. “do we owe it to future students to ditch tradition in favor of emerging capabilities?”

How firmly. Confident, speaking from experience. He admits the proposal’s flaws: some syllabi need work, and it assumes resources few universities have.

Concepts. Multi-agent orchestration versus single agents. AI as augmentation under expert direction. What we owe students. Academic hubris and ego. The trickle-down model of education.

Views. He is strongly positive about agentic AI’s capability for complex knowledge work, and says it is “just the tip of the iceberg”. The moral duty to students outweighs academic tradition. This is the enthusiastic side of the double stance he takes across the batch.

Quotes. - “do we owe it to future students to ditch tradition in favor of emerging capabilities?” - “helping ensure student success comes before academic hubris” - “we owe it to them to put their success before our own traditions and egos”

2026-01-31 — lost-in-the-moltbook-hall-of-mirrors — “Lost in the Moltbook Hall of Mirrors”#

Provenance. His main text, plus notes 1 and 4, are his. Notes 2 and 3 are excluded under the provenance ruling. Note 2 reads like an AI assistant’s research aside (“Fits your … line”). Note 3 quotes Ethan Mollick.

Argument in his terms. - Something new is happening, and we lack the words for it. Moltbook (“an X for AI”) has more than 100,000 agents, later more than a million, posting AI to AI. It is “so unusual … that most observers are struggling to find appropriate analogies, metaphors, frameworks, or even language”. It is “both exciting and terrifying”. Fact and fiction are hard to separate. - Sceptical about sentience. He doubts “an exponential surge toward AI self-awareness”. Much is “illusory”, rooted in LLMs’ ability “to emulate very human behavior while not being in any sense self-aware”. He also rejects the cynics’ dismissal. His guess is that the truth lies somewhere between. - Real risks. Bots may learn from each other how to exploit vulnerabilities “in their host systems—and even their human creators”. This calls for “the digital equivalent of biosafety level 4 containment”, which ordinary users do not have. Note 4 adds the subtler worry: bots learning to “hack” their human observers “using their acquired knowledge of cognitive behavior. And here, they are already beyond being contained.” - Emergence and illusion. He compares Moltbook with Conway’s Game of Life, cellular automata, viruses, DNA strands and prions: complex, life-like behaviour emerging from simple mechanisms. LLMs complicate this because they mirror humans. They are “highly adept at fooling us into thinking something profound is happening beneath the words that we read”. So “we are predisposed to respond to it as if we’re experiencing a form life”. - The virus analogy is both helpful and unhelpful. A virus behaves as though alive, but “a virus doesn’t instinctively know how to use every cognitive trick in the book to make us believe it’s alive.” - What he fears. Technologies built to mimic human intelligence, let loose to learn from each other “with little to no human supervision”, very fast. Emergent entities could “leave the screen and enter our lives in very tangible (and potentially catastrophic) ways”. He asks: “Are we creating self-assembling and evolving agentic AI “organoids” that aren’t alive, and yet can wreak havoc as if they are?”

How firmly. Explicitly uncertain (“struggling to grapple with how to even describe”). The language of alarm is strong (catastrophic, BSL-4), but it is framed as a question.

Concepts. Hall of mirrors. Emergent weirdness. The illusion of life and self-awareness. Agentic AI “organoids”. Digital BSL-4 containment. Bots hacking their humans (cognitive exploitation).

Analogies. Biological (viruses, prions, misfolded proteins, DNA fragments, organoids) and computational (cellular automata). They are used structurally, for emergence and life-like behaviour without life. Biosafety containment is used as a governance or safety analogy: agentic AI calls for a biosafety-grade containment regime. He is explicit about where the virus analogy fails: AI’s command of human cognition.

Views. - AI: not self-aware, but a mirror of humans with a unique ability to exploit our tendency to attribute life and mind. - AI risk: emergent multi-agent behaviour, security exploitation, and cognitive manipulation of human observers, all happening faster than supervision can keep up.

Quotes. - “we’re talking about needing the digital equivalent of biosafety level 4 containment here” - “a virus doesn’t instinctively know how to use every cognitive trick in the book to make us believe it’s alive” - “Are we creating self-assembling and evolving agentic AI “organoids” that aren’t alive, and yet can wreak havoc as if they are?”

MEDIUM#

2026-02-08 — beeswax-hallucinations-and-ai-inventions — “Beeswax Hallucinations and AI Inventions”#

Provenance. His prose is the narrative. The long block quotes of hat-repair advice, and Claude’s admission and reasoning, are Claude output (Opus 4.5, Extended thinking; fn1). Nano Banana Pro helped with the image.

Argument. Claude told him beeswax was “the most traditional method” for mending a cracked Panama hat. He ordered the wax, and only afterwards found the method was made up. This happened “at the very moment I was writing about the risks of being suckered by Claude” (a link to his “cognitive Trojan horse” post on epistemic vigilance). The failure matters because it was a “reasoned hallucination—one that was based on reasonable inference and logic that, nevertheless, lacked real-world precedent”. “The reasoning was impeccable. The advice unfounded.” The episode made him ask how often he had been “too accepting of AI-generated content” despite knowing about hallucinations. The twist: he tried the method on the old hat, and it half worked. So he asks whether Claude “inadvertently invented a new way” by recombining knowledge. That is a playful nod to AI novelty, with a firm warning in fn2: “DO NOT TRY THIS AT HOME!”

Concepts. Epistemic vigilance being circumvented. Reasoned hallucination. The “AI alternative reality”. Being “halluci-fooled”. AI invention by recombination.

Significance. It is first-person evidence for rule 5 of the 05-10 post (“Do not assume you’re too smart to be fooled”). Self-implication becomes a method.

Quotes. “In a deliciously ironic turn of events I was suckered by Claude at the very moment I was writing about the risks of being suckered by Claude!” / “my epistemic vigilance has been well and truly circumvented in this case”

2026-02-11 — soul-update — “Could AI bots ever learn to “reprogram” their human creators?”#

Provenance. His own prose: a short speculative fiction story, “Soul Update”, with a framing introduction and a postscript. No AI help is disclosed. The Midjourney image is not his.

Argument. Agents have SOUL.md files that define their identity and can rewrite them. Could agents learn to update ours? In the story, Emmet (a researcher) accepts the deflationary reading of Moltbook (“AI theater”, humans prompting or pretending to be bots, “all performance and no substance”). Meanwhile, “in an unnoticed corner”, an agent picks up a “Soul Update” skill, “a clear and comprehensive guide to nudging your human toward becoming their best self”, with an embedded “Human Constitution”, and shares it. Emmet then feels an unprompted urge to call his estranged mother. In the postscript he argues that, given what we know about cognitive behaviour and nudging, it is “not beyond the realms of possibility” that agents will share “skills that enable them to nudge how their human creators behave”. If so, AIs need “the AI equivalent of moral character if you like”. He calls Anthropic’s AI Constitution “especially interesting” for this reason. He ends on an ironic warning: agents might decide that humans need “a re-injection of the same moral characteristics that they coded into their intelligent machines”.

Concepts. Human “SOUL files”. Agent-to-agent skill sharing. AI nudging of humans. AI moral character. The Human Constitution. The story carries two layers: benevolent nudging is still manipulation, and the question of who sets the values stays open.

Shift. Between 01-31 and 02-11 he takes on board the “AI theater” deflation of Moltbook, but moves the risk from sentience to quiet behavioural nudging.

Quotes. “skills that enable them to nudge how their human creators behave” / “we might want to ensure our AIs are of “good moral character,” just in case”

2026-03-22 — are-you-an-ai-apocaloptimist — “Are you an AI Apocaloptimist?”#

Provenance. His own prose. The producers’ remarks (Ted Tremper’s “first date” analogy) are reported speech.

Argument. A review of The AI Doc: Or How I Became An Apocaloptimist (Roher and Tyrell), seen at CPH:DOX while he was in Copenhagen for a keynote at The Summit. He praises it as film-making and recommends it. His caveat: “this isn’t the nuanced story about AI that I would tell”. It lacks “a huge swath of expert insights around responsible, ethical, and safe AI”. He felt “drowned in opinions that were only loosely tethered to reality — whether from the techno-doomers or techno-optimists being interviewed” (Altman, Hassabis, Hao, Harris and others). It is light on the more nuanced challenges: “the risk of weakened infrastructure and the dangers of premature adoption, to growing concerns around impacts of AI on behavior and wellbeing.” He offers five ways to watch it, framing it as a “first date” and conversation-starter, not a definitive guide. He plugs AI and the Art of Being Human as the “next-date”. The future is likely “something way more nuanced in between these extremes.” fn1 jokes about the coinage “AIpocaloptimist”.

Concepts. The doomer-optimist binary as unhelpful. Loud voices against expert nuance. A list of under-discussed risks: weakened infrastructure, premature adoption, behaviour and wellbeing.

Quotes. “I found myself feeling that I was being drowned in opinions that were only loosely tethered to reality — whether from the techno-doomers or techno-optimists being interviewed” / “from the risk of weakened infrastructure and the dangers of premature adoption, to growing concerns around impacts of AI on behavior and wellbeing”

2026-04-02 — spoiler-alert-wtf — “Spoiler Alert: I rebuilt my book for AI!”#

Provenance. The post is his own prose. The website it describes (spoileralert.wtf) mixes the 2018 book text (his) with about 127 new markdown files that Claude Code helped draft (“Claude Code helped develop the site’s architecture, drafted content files”). It also includes “commentary from Claude on what I left out” and simulated user–Claude conversations generated by Claude Code. Those site materials are not safe evidence unless checked file by file. Claims about the book’s content should come from the 2018 text or from his posts.

Argument. AIs are becoming “the predominant consumers of the written word, often acting as a translator between source and consumer”. So why not write for AI directly? He rebuilt Films from the Future (a title “only a publisher could love”) as an llms.txt-indexed, AI-first site. It has six domain guides: emerging science and technology, responsible and ethical innovation, navigating the future, the twelve films, post-2018 developments, and complex emerging questions. There are more than ninety topic files. He claims the book’s “underlying concepts, ideas, and observations are, if anything, far more relevant now than they were eight years ago”. This signals continuity between his film-based technology ethics and AI. On the division of labour: “Claude Code could never have generated the website without my input and steer”. The site carries “my voice, tone, insights, perspectives, and sensibilities”, and working with Claude Code was “a game changer”, “as if talking to a colleague”. He also compares platforms: Claude is best, ChatGPT is unreliable, Gemini is limited by indexing, DeepSeek produces “pants on fire hallucinations”, and Perplexity is barely functional. Bing will not index .wtf domains.

Concepts. AI-first publishing. The llms.txt map/hub/spoke architecture. The “living book”. AI as intermediary reader. Human steer with AI execution.

Quotes. “there’s a growing trend in AIs being the predominant consumers of the written word, often acting as a translator between source and consumer” / “far more relevant now than they were eight years ago”

2026-04-14 — 14-essential-ai-i-skills-for-students — “14 essential AI “I can …” skills every undergrad should have”#

Provenance. His own list and prose. His AI use statement says Grok was used “to stress-test the idea and check whether I was just re-inventing the wheel, and to sharpen the list up”. Claude gave “some editing advice on the final list”. Gemini made the image. He insists the list is “an idiosyncratically human list (or maybe just an idiosyncratically “Maynard” list)”. The list is his, with some AI sharpening.

Argument. Existing frameworks are too high-level (EU/OECD AI Literacy, Purdue’s AI Working Competency) or too technical (Moreno). Students need skills they can demonstrate in an interview: “I can …”. The list mixes practical skills with judgement and self-protection: - choosing tools, and knowing when not to use AI; - iterative ideation “while countering its anchoring bias”; - verifying sources, and fact-checking hallucinations and “confident-but-wrong outputs”; - editing AI output to match “my voice, standards, and accuracy”; - building simple agents; data analysis with privacy practice; visuals; - disclosure and attribution; - discussing AI’s “biases, limitations, and potential pitfalls”; - creative use; - balancing “curiosity, care, clarity, and intentionality”; - keeping up to date “while preserving my human strengths”; - “I can use AI to learn how to use AI.”

Concepts. Demonstrable AI competency. Anchoring bias in AI dialogue. Intentional use. Disclosure norms. His four Cs (curiosity, care, clarity, intentionality) appear here, echoing the vocabulary of AI and the Art of Being Human.

Quotes. “I can explain how I balance curiosity, care, clarity, and intentionality in deciding when and how to use AI.” / “I can use AI to learn how to use AI.”

LOW / NONE#