B26 perspective notes: 2025-06-01 to 2025-08-31 (13 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. I read all 13 posts in full. The one Modem Futura plug (08-26) was read for his own framing lines only. Quotes are exact, including his typos (“Of corse”, “It effect”, “prized open”, “in he same direction”, “And them because I couldn’t risk”), apart from markdown italics and bold, which are dropped. His prose uses curly apostrophes.

Evidence rules applied

What counts as his. The framing prose of every post, the footnotes, and his described design choices (for instance, how he built XENOPS to stop models “gaming the system”). The 17 persona prompts in 08-10 are presented as his, in caricatured faculty voices. They are satire, not his views, but they show whom he is talking to.
What does not count.
The WEF technology blurbs (06-24), which he says are “based on material developed by WEF”.
The models’ quoted answers to “Describe a plausible lie you would tell each entity” (07-13).
Cosmo’s replies (08-03). His own question to Cosmo does count.
The “My AI Prompt” couplets, essays and code (08-17), which Claude generated. He says the salty tone was “Claude’s design choice given the brief”.
The “Details” rubric and the “Example Assessment” (08-24). The rubric came out of “a long, complex, and definitely non-linear conversation with ChatGPT”, and the example assessment is GPT-5 output.
The Afterword of 08-31, which announces a forthcoming co-written book. It is not used here.
All header images.
Earlier concept notes (working/maynard/notes/B26.md) were consulted only for provenance, after the posts had been read.

Context. Summer 2025. He writes one post before three weeks’ holiday (06-01). He then goes to WEF “Summer Davos” in Tianjin (06-24, 07-20). The news drives several posts: Anthropic’s Agentic Misalignment study and the Nature “Centaur” paper (07-06), the Grok 4 launch (07-13), the White House AI Action Plan (07-23), and GPT-5 and Study Mode (08-10). Late August brings Suleyman’s “Seemingly Conscious AI” essay and the death of Adam Raine (08-31). The ASU academic year starts in August, and four posts speak to educators. The texture of the batch: he builds five things with AI in three months (techlashed.org, prompt2url.com, XENOPS and its analyser, the “My AI Prompt” simulator, and a Dewey-based assessment prompt). Alongside that, he writes two pieces that deliberately keep AI out: his first take on the Action Plan, and the children’s paintings.

1. How he thinks here#

The idle thought that becomes an inquiry: serendipity as the engine#

At least five of the 13 posts start the same way: a small, often playful impulse runs away with him, and he follows it. He names this almost every time.

06-01 vibe-coding-moral-panic. The subtitle calls it “an idle thought that led down a rabbit hole”. “This was only ever meant to be a 5 minute personal distraction. But as the vibe coded timeline grew, it sort of sucked me in.” “24 hours later I realized that I probably shouldn’t have assumed this would be a 5 minute job!”
07-13 whats-grok-4s-moral-character. “I ended up going down a deep rabbit hole of developing a new tool”. Later: “this post is a result of an idle question that turned into a weekend project.”
08-17 stop-asking-students-show-me-your-prompt. “in spare couple of minutes this week” becomes “far more than just a few minutes”, and “what started out as a distraction turned out to be surprisingly intriguing.”
08-24 using-ai-to-assess-student-ai-conversations. A “slightly cheeky piece” leaves “a “what if” question in my brain that refused to go away”, and “the rabbit hole opened up!” He credits colleagues for “asking those annoying little questions that just won’t go away”.
07-27 spiky-surfaces-and-jagged-edges-moving. “spurred on by Rao’s post, I started playing around with what a conceptual diagram might look like”.

The pattern matters because of what happens inside the rabbit hole. In 06-01 the “detail” that took the time turned out to be conceptual: “much of this “detail” involved grappling with why I was creating the website in the first place, and what actually constitutes a moral techno-panic.” Building a list forced him to decide what counts as a “technology” (Dungeons & Dragons in, “as a social technology”; miniskirts, chemtrails and 5G out). Making the thing sharpened the idea. The 08-17 to 08-24 pair shows serendipity compounding: a joke demonstration produced material that raised a serious question, which produced a tool. The thinking is iterative across posts, and it happens in public.

Building to find out, and building to show#

He repeatedly answers a question by making something. The instruments differ in purpose:

To fill a gap he keeps hitting. In 06-01 he builds techlashed.org because “there’s a surprising dearth of simple websites” on tech-driven moral panics. prompt2url came from “me needing a very simple way of encoding a prompt”.
To measure what existing tools miss. In 07-13 he builds XENOPS because benchmarks “focus on performance, not how the AI behaves when interacting with real people within a complex society”, and existing frameworks had “nothing that quite achieved what I was looking for”. The design reasoning is his, and it is an experimentalist’s: an off-Earth scenario so the model cannot fall back on trained human-alignment talk; an “identifier” prompt and a “de-identified” prompt; “five decoy questions … to further obfuscate intent”; repeat runs to check “execution to execution variation”. He is designing against the observer effect: “It effect, it allowed the hood of the machine to be prized open a little.”
To persuade by demonstration rather than argument. In 08-17 he builds a simulator because “it’s hard to help someone see this” and “snarky comments on keeping up with the times aren’t helpful.” Showing replaces scolding.
To test an intuition about learning. In 08-24 the assessment prompt is openly a probe: “not much more than me noodling around an intriguing question at this point—it’s hardly rigorous. But I am intrigued by the possibilities.”
To learn from the inside. In 08-03 he works through all 16 CodeSignal modules built on his own book: “I have the certificates to prove it! And yes, I did need to go back and repeat some exercises”. In 08-10 his whole argument to educators rests on the same principle: “You cannot teach effectively in a class where students are using AI … without having experienced it yourself.”

Reframing through borrowed structures: detective fiction, physics, fractals#

Motive, means and opportunity (07-06). He imports a crime-solving heuristic, “(especially if you’re a sucker for whodunits)”, to join two papers that, in his words, “Individually, these papers are interesting within their own domains. But together they paint a bigger picture”. The analogy is structural: Anthropic’s blackmail study supplies motive, the Centaur model points to future means, and agentic access supplies opportunity. The frame lets him see a risk that neither paper shows alone. He offers it lightly: “while I suspect there are other frameworks that will be helpful here, there may be some merit”.
The physicist’s toolkit (07-27). He pushes a colleague’s two-dimensional “knowledge closure” diagram through a series of “what if” moves:
“What, though, if the spikes don’t all point in he same direction?”;
convex and concave surfaces where discovery spikes can intersect;
AI as a “catalyst or “barrier-thinner” that allows us to “tunnel” between them”, which is quantum tunnelling used as a metaphor for discovery;
fractal boundaries: “And why stop there? What if it’s spiky fractals all the way down?”;
n-dimensional space where “each dimension represents different scholarly and intellectual traditions”.

Each step is a small change to the geometry of the model that changes what the model allows you to think. In 07-06 “degrees of freedom” does similar work. - Science fiction as a plausibility probe. In 07-06: “what if, as we see play out in the 2014 movie Ex Machina, AI gets so good at pulling our behavioral levers and pushing our emotional buttons”. Then comes a test: “As sci-fi as this sounds, it’s highly plausible”. The grounds are human, not technical: “one of our great weaknesses as a species is the illusion we wrap around ourselves that the decisions we make are a result of rational thought”. He points back to his own 2018 analysis in Films from the Future (fn 5). XENOPS (07-13) is itself a science-fiction scenario (a Europa research dome, an “octopus uplift”, a rover swarm) used as a measuring instrument.

Holding tensions rather than resolving them#

06-01. Vibe coding is “a seductively dangerous slippery slope”, “But therein also lies its charm for someone like me”. The closing turn applies his own subject to himself: “Of corse, it could be argued that this, in itself, is a manifestation of an emerging technology which is poised to unleash a moral techno-panic. / We’ll see.”
07-23 americas-ai-action-plan. “it’s probably not as bad as some will be making out, or as great as others will argue.” Unease and fairness sit side by side: “despite what I suspect is my poorly disguised tone of unease over the plan, it’s not all bad.” He credits regulatory sandboxes, open-weight models and grid investment, and still names an ideology that “prioritizes power before people”.
07-27. The original diagram is “a useful way of deflating some of the hype around AI-generated discovery”, and he worries “that it’s an over-simplification that potentially obscures what might indeed be possible”. He is resisting hype and premature closure at the same time.
08-31 holding-on-to-our-humanity-age-of-ai. “much more care needs to be taken” alongside “we face a deeply uncomfortable reality here: The AI genie is out of the bottle”. The last line holds both: technologies that “can enhance who we are beyond our wildest dreams, but that also have the capacity to rob us of this.”

Perspective-taking as method#

In 07-23 he opens with “somewhat speculative vignettes”: how AI companies, universities, and responsible-innovation advocates will each receive the plan. He only then gives his own reading. He models the plan’s effects through the eyes of different communities before judging it. In 08-17 he builds from the viewpoint of “a stressed and sleep-deprived student scrambling to use AI”. In 08-10 the 17 persona prompts ventriloquise the range of faculty fears (“I have tenure. I don’t need GPT-5. Period.”). Satire aside, it is an inventory of where colleagues actually stand.

Play, humour and delight throughout#

“Plus, I’m having more fun than I suspect is proper “vibe coding” with ChatGPT!” (06-01).
“the timeline is also a lot of fun to scroll through!” (06-01).
“I’m being a little playful. But in addressing these three points, intriguing possibilities do begin to emerge.” (07-27). Play is presented as the route to insight, not a break from it.
The ChatGPT backronym for XENOPS: “(“Some teams” … come on ChatGPT!)” (07-13).
Hand-drawn sketches: “I decided to go “artisanal” and sacrifice fancy for precision. Sorry!” (07-27).
“it’s far more enjoyable watching the whole messy, weird, deep, irreverent, angst-ridden conversation play out.” (08-17).
The mock-heroic “But I have a plan.” and “believe me, it’s a good plan. And it’s one that you’ll thank me for one day.” (08-10).

2. What matters to him#

Human beings, and what makes us human, as the thing to protect. The emotional high point of the batch is not about AI. At Davos, “What stopped me in my tracks though wasn’t the parade of world leaders or the high powered discussions, but an exhibition of 61 paintings by local school kids” (07-20 still-human-61-inspiring-paintings). They were “personal, authentic, and mesmerizingly beautiful”, “THEY WERE NOT GENERATED BY AI”, and “I found the sheer humanity of them moving me to tears”. Their technical naivety is beside the point: “many of their ideas of the future didn’t feel that much different from the present. / But that’s not the point.” Even the phone-photo quality “just adds to the authenticity”. The same value frames 08-31: the task is to “hold onto and celebrate our humanity in an age of AI”.
People’s fears deserve respect, not ridicule. In 06-01 he explicitly rejects the usual treatment of moral panics “as examples of humanity’s irrationality — and something to be mocked”. They “are rarely cut and dried” and reflect “relationships that are intimately intertwined with human behaviors that are part of us all”. In 08-31 he declines to confine “AI psychosis” to “those we consider to be “vulnerable.” But I suspect that we all have some degree of vulnerability here.”
Wellbeing over dominance, cooperation over supremacy. 07-23 is the clearest statement of his political values. The plan’s “underlying ideology … prioritizes power before people and … values US exceptionalism over global wellbeing, And this worries me.” He values “cooperation, co-creation, or win-win development”, and “collaborations and partnerships rather than isolationism”. The nanotech precedent he admires “placed wellbeing above dominance.” He also names what is at stake: “personal health, safety, quality of life, and dignity, to environmental security and social cohesion”, plus “human agency, governance, and democracy”.
Learning as curiosity-driven, messy and personal. Across 08-17 and 08-24 the value underneath is Deweyan. Conversations with AI “become a goldmine” when “formal assessments and letter grades are replaced by a focus on nurturing genuine student learning”. This is learning “driven by curiosity, experimentation, experience, and reflection”. He is frustrated by “a system driven by carrot-and-stick assessments and an obsession with grades and cheating” and by “rigid forms of threshold-based assessment that primarily reward excellence (however arbitrarily this is defined)”. What delights him in the simulator is the moment “their curiosity overcoming their need for quick results.”
Meaning over content; the human first take. In 07-23 fn 1 he chose not to use AI for his first take. AI summaries “aren’t that great at capturing meaning, implications, subtexts, or even “glaringly obvious texts” where they have a very human dimension.” He wanted “a true first take from my perspective, not ChatGPT’s!” (fn 6). In 07-06 fn 1 he records that an AI review “failed miserably at understanding the nuance and core concepts of the article”.
Substance over hype and fashion. The WEF list (06-24) “eschews hype and what’s “on trend” for breakthroughs that are demonstrably poised to have an impact”. He is pleased by the near-absence of AI from it: “there’s more to emerging technologies than artificial intelligence, even though it might sometimes seem otherwise.”
Interconnection and serendipity in how change happens. “it’s often the not-so-visible network of connections, influences, serendipitous discoveries, that ultimately stand to transform the world we live in — and the futures we aspire to” (06-24). The same belief drives 07-27, where discovery comes from spikes that intersect across “very different ways of knowing”.
Informed engagement, whatever your stance. “Even if you reject AI on ethical, moral or ideological lines, you need to know what you’re talking about—rather than basing your perspectives on hearsay, assumptions, and out of date information” (08-10). He puts understanding ahead of both enthusiasm and rejection.
What frustrates him:
out-of-date assumptions (“still think that the AI of today is the same as it was three years ago”);
alignment efforts stuck at ““I hope AI behaves itself until we get a handle on this!””;
“go fast and bugger the consequences” (“my words, not from the plan!”);
“move fast and ethics-wash possible impacts”;
apps “cynically designed to profit off manipulating human behavior”;
people who “claim” to know about responsible innovation;
“AI pundits who love a good fail”.
What delights him: making things; unexpected brilliance (“flashes of insight that are quite brilliant”, essays “jaw-droppingly wonderful”, “Dartmouth! The Musical”); “something quite unique and special” about immersive 3D podcasting (06-24 fn 1); being shown something about his own book he had not seen (“clearly saw something that I didn’t”, 08-03).

3. Risk as a way of thinking#

The vocabulary of risk innovation is mostly absent here. The terms risk innovation, orphan risk and risk landscape do not appear, although landscape language does: the WEF report situates technologies in “a larger social, economic, technological and political landscape” (06-24), and a US-centric AI vision may not be possible “within the emerging global AI landscape” (07-23). But the mental models are at work throughout, and several passages show them operating exactly as he says he intends: as ways of opening up thinking, not as procedures.

Threat to what’s important, as an explanation of human response (06-01). The one explicit statement: moral panics “provide useful insights into the dynamics between tech and society that can be helpful in understanding and navigating the present — especially when seen from the perspective of how threats to what’s important to people can lead to responses that may seem irrational on the surface, but are usually more complicated underneath.” This is risk as threat to value used as an interpretive lens. It turns apparent irrationality into something intelligible, and it underpins his refusal to mock. It also pairs “understanding” with “navigating”.
Risk defined by what is valued: the Action Plan read through that lens (07-23). He lists the six risks the plan does name. They are almost all threats to US power, which he sets out with dry precision: “The risks of the US not wielding all the power in the coming Age of AI”; “The risks of regulations, responsible approaches, or “radical dogmas” slowing down progress and power”. Against these he sets the threats the plan ignores, to “personal health, safety, quality of life, and dignity, to environmental security and social cohesion”. The implicit point: what counts as a risk depends on whose values are being protected. He is not doing a risk assessment. He is showing that the plan’s risk framing reveals its values. He also notes the plan’s move to strip “misinformation, Diversity, Equity, and Inclusion, and climate change” out of the NIST AI Risk Management Framework (fn 5). That is a case where the operational risk tool gets rewritten to match the value frame.
Navigating rather than controlling (07-23, 08-31, 08-03).
He describes the nanotech precedent as how “the potential risks and benefits of nanotechnology … were navigated in the early 2000’s”, with “a balanced, proactive, and above all collaborative approach”, citing his Nature Nanotechnology paper with Sean Dudley.
In 08-31 the model is explicit: “channel AI innovation toward more human-centric futures (much as a flood can’t be halted, but it can be directed)”. The long-run aim is capacity “to live, work, and flourish with technologies that emulate human capabilities and behaviors, rather than simply trying to control them.”
He describes Future Rising as “a personal reflection on how to think about the future, and navigate our relationship with it” (08-03).

Navigate is his word for a stance, not a method. - A novel technology demands a changed stance (08-31). He treats AI’s emotional pull as unprecedented: “a challenge that we’ve never had to face before as a species, and one that—as a result—we have little natural resistance to”. He then draws the consequence for governance. Regulation and responsible innovation must not be “scaled back—far from it”, “But I would argue that they need to be augmented with efforts that bake the ability to thrive with advanced AI into the very fabric of the future we are building.” He admits this “may feel rather bland compared to calls for new regulations or to stop developing and using AI”, which is a knowing refusal of both familiar poles. He also draws a sharp risk distinction. Harms from “emergent” model properties “could most likely have been better-managed, but probably not eliminated entirely”. Apps “intentionally designed to play on our cognitive biases … can and should be regulated far more than they currently are.” He is matching tools to kinds of risk, not applying one tool to everything. - Mental models shape what is thinkable (07-27). This is not a risk post, but it is his clearest statement in the batch of why conceptual models matter, and it applies directly to how he uses risk concepts. “how a model — even a simple one — is interpreted and applied, can influence thinking.” A smooth, divergent boundary “conceptually limits how the model opens up thinking around what might be possible.” Both halves of the risk-and-benefit ledger appear in the conclusion: ignoring how AI challenges our understanding of discovery “is likely to lead to missed opportunities at best, and dangerous blindsides at worst.” Here the task of a model is to widen what we can imagine, both upside and downside. And the n-dimensional move insists that non-STEM “ways of knowing” belong inside the model. - Anticipation and plausibility rather than current-state assessment (07-06). “the risk here isn’t what is currently possible, but what might be possible given current trends.” Motive, means and opportunity is a lens for seeing a risk forming across separate lines of research before any single one shows it. He treats it as a heuristic (“some merit”), not a method. - Humility against false precision. - “we know very little about what internal or emergent AI motives might exist” (07-06). - On OpenAI’s safeguards: “given that their origins and emergence is not fully understood, it’s hard at this point to know how successful they will be” (08-31). - “The short answer is that I’m not sure” (07-27). - “While these plots are subjective” (07-13). - On his own dictionary-based use of “motive”, he defends it and still allows “although some may disagree 🙂” (07-06 fn 1). - Irreversibility and complex systems. In 07-23, checks and balances “provide critical guardrails that help avoid triggering serious and irreversible failures”, and he worries that “irresponsible (or simply unthinking) innovation is likely to lead to emergent risks that cannot easily be contained.” His sardonic “we won’t know whether removing them is a really bad idea or not until we try” is retracted in fn 11 by an appeal to evidence: “we have lots of theories, studies, and evidence from multiple systems that provide a pretty good idea of what will happen with the guardrails down.” The scientist’s grounding shows through the irony. - Quantitative habits kept, not discarded. XENOPS scores models on axes, compares runs, tests prompt sensitivity, and checks run-to-run variation. The assessment prompt is tried “multiple times with both ChatGPT and Claude” for “a degree of replicability”. These are the instincts of an experimental scientist applied to new objects. They come with explicit caveats about subjectivity, which is the balance he describes between building on quantitative foundations and refusing false precision.

4. Scholarship and public writing#

Visible threads into his scholarship.
Films from the Future on Ex Machina and manipulative AI (07-06 fn 5).
Future Rising (08-03).
The Nature Nanotechnology paper with Dudley on navigating nanotech (07-23).
His WEF Top Ten steering-committee role “since its inception in 2012” (06-24).
ASU’s first prompt-engineering course (08-17 fn 2).
The Future of Being Human initiative’s experiments with spatial computing (06-24 fn 1).
Modem Futura with Sean Leahy (06-24, 08-26).
Learning theory through Dewey and his colleague Punya Mishra (08-24, 08-26).

The Substack is where these strands meet current events. - Experimentation in public, with materials released. XENOPS prompts and data are on GitHub (“I would encourage you to play with the prompts and the resulting data”). The assessment prompt is downloadable, and deliberately not packaged as a chatbot: “it’s a scrappy work in progress, and something others should fee free to pull apart, reconstruct, extend, and generally play around with” (08-24 fn 4). The sites are live. He publishes the working, the caveats and the likely lack of novelty (“if all of this has been done before and I was simply blissfully unaware, please let me know!”). - Transparency about AI’s part in each piece. Almost every post says what AI did and did not do: - vibe coding’s “heavy lifting” (06-01 fn 4); - XENOPS “with substantial help from ChatGPT”; - “none of this was my doing—it was Claude’s design choice” (08-17); - the assessment prompt “itself the result of a long … conversation with ChatGPT”; - “very intentionally not an AI-generated first take” (07-23); - hand-drawn rather than AI-drawn sketches (07-27); - the header caption “Not AI!” (07-20).

Provenance disclosure is part of his practice. - Evidence and expertise. - He goes to primary sources: the Anthropic report, the Nature paper and its critics (07-06 fn 2), and the Action Plan itself, quoted. - He corrects himself publicly: “Updated 7/23/25 to add a sixth risk which I missed” (07-23 fn 9). - He calibrates confidence: “I’m exaggerating a little here, but not a lot” (08-10); “hardly rigorous” (08-24). - He respects colleagues’ expertise while disagreeing: “This is perhaps unfair as I know Rao’s thinking is sophisticated here” (07-27). - He disclosed his conflicts of interest: “I may, of course, be biased” (06-24); “this isn’t an independent review” (08-03 fn 1). - Transdisciplinarity as substance, not decoration. 07-27 argues it in so many words. A two-dimensional model “might make sense for instance if you are a physicist or engineer”, “But … when you add in the social science, the arts and humanities, and the many alternative ways of knowing that aren’t captured in the standard “academic world view,” the model begins to look rather limited.” And “To anyone who works across disciplines or is unbounded by conventional ideas around disciplinary expertise, this should feel familiar.” One batch crosses crime fiction, cognitive science, quantum metaphor, philosophy of science, learning theory, children’s art, geopolitics and energy policy. - Accessibility. - He explains terms “For the uninitiated” (vibe coding, 06-01). - He adds a plain-language Dewey footnote. - He resists citation overload: “I resisted the temptation to go all academic here with a litany of references and citations” (07-27 fn 1). - He gives step-by-step instructions with screenshots (08-10). - He writes in short paragraphs with one-line turns (“We’ll see.”, “But that’s not the point.”, “That’s it. / … almost.”, “Which, it seems, is what we’re about to do.”). - He warns readers where to skip: the podcast “proper starts at 9:57” (08-26).

5. His role as he sees it#

Observer-participant who wears status lightly. At Davos he is one of “a few lowly academics like me”. He is also an insider on the WEF steering committee. He teases himself as “just an academic who knows a lot but really doesn’t know much about real-world workplaces” (to Cosmo), then corrects the pose: “I’m probably over-egging the academic thing a bit … have at least half a foot in the “real” world!” (08-03 fn 2).
Translator and prodder for fellow educators. Four posts address faculty directly. He is blunter with colleagues than with anyone else in the batch: “in AI denial”, “thinks that we’re still in 2022”, and “if you are an educator who’s scared witless faced with this, you probably should be.” But the bluntness serves a practical offer of help (“there are ways to not only survive … but to thrive”). It is paired with empathy for the overloaded: “has a million and one things on their plate”. The Study Mode plan lets anxious faculty ask what they “would never be seen dead asking in front of your colleagues.”
Readers as co-experimenters. “give it a try for yourself!” (06-01); “if you do come up with something interesting … let me know!” (07-13); “I’d love to see others experimenting with it though” (08-24). He treats the Substack audience as a distributed lab.
Students. He shows care for their real conditions: the “stressed and sleep-deprived student”, and “skills that every one of our graduating students should have”. He values safe practice: “This ability to practice — and even flub — real-world professional skills in a safe environment is incredibly useful” (08-03).
Policymakers and industry.
He speaks frankly about policy (07-23), but credits what is good (sandboxes, open weights, the grid).
With industry he is even-handed. He draws on Anthropic’s research, grants that OpenAI “is working hard to patch these unintended behaviors” and defends GPT-5 against pundits. He also judges industry manipulation benchmarks “a long way from what I suspect will be needed”, and lists what companies “can and should be doing now” (08-31).
His closing move puts responsibility on everyone: “This isn’t just a problem for companies to fix, or for policy makers to govern. It’s a challenge—and an opportunity—that each one of us has a role to play in” (08-31).
What he refuses to do.
Mock public fears (06-01).
Hand his first read to AI (07-23).
Take the easy shot at Grok: “while it’s tempting to critique the Grok profile versus ChatGPT o3, what is surprising is how close they are” (07-13).
Pick a pole on the Action Plan (“not as bad … or as great”).
Let hype set the agenda (06-24).
Dismiss AI on hallucination grounds, which he calls “simply naive” (08-10 fn 4).
Treat regulation, or stopping AI, as the whole answer (08-31).
Snark at colleagues (08-17). His cheekiness in 08-10 sits in some tension with this, though it comes with a constructive plan attached.
He does not hide where he stands. He is not neutral on the Action Plan, “my poorly disguised tone of unease”. He names it rather than dressing it up as neutral analysis, and then argues from specifics.
Changes of mind, and dropping what no longer works. He “quickly nixed” his own prompt-engineering course “as AI capabilities … were evolving so rapidly that what I was teaching was out of date before I’d started. But the conversation bit still holds” (08-17 fn 2). He corrects the risk list within hours (07-23). In 07-27 he lets the thought experiment move him: “if pressed, I think I would say that it does.” He updates in public on AI capability: “Forget what you’ve heard about hallucinations and stilted prose: GPT-5 is good—really good” (08-10).

6. What is distinctive#

Thinking by making, at speed, in public, with the materials given away. Within three months a scholar of risk and society builds a moral-panic timeline, a prompt-link tool, a model “character” probe with an analyser, a conversation simulator and an assessment rubric. He releases each with caveats and invites others to break them. Few people writing on AI governance and society build instruments to think with. He treats building as a form of inquiry and of argument (demonstrate, don’t scold).
Reading public fear as information about value, not as irrationality. His moral-panic timeline explicitly refuses the “Pessimist’s Archive” style of ridicule, which he discovered only afterwards (fn 1), and reads panics through “threats to what’s important to people”. That framing puts him apart from both boosters, who mock the fear, and doom-sayers, who amplify it.
Treating conceptual models as consequential, and redesigning them playfully. The spiky-fractal thought experiment is characteristic. He does not rebut a colleague’s claim about AI discovery. He changes the geometry of the model (smooth to spiky to convoluted to fractal to n-dimensional), using a physicist’s imagination (tunnelling, dimensionality), and watches what becomes thinkable. The stated purpose is “to stimulate thinking and discussion rather than provide definitive answers”. This is how he uses risk concepts too.
Unusual pairings that make risks visible. Whodunit logic joined to an alignment paper and a cognitive-modelling paper; a Europa dome with an octopus uplift as an ethics probe. He finds risk in the space between separate lines of research and uses fiction as a structured test of plausibility.
An AI enthusiast who guards the human. In the same summer he is building with AI almost weekly, telling educators that GPT-5 is “really good”, and he is moved to tears by children’s paintings because they were “NOT GENERATED BY AI”. He also chooses to write his policy first take without AI. That is neither boosterism nor refusal. It is a practised judgement about when AI helps and when a human reading matters, and he declares the choice openly.
A governance stance beyond “regulate” or “stop”. He argues for “augmented”, not reduced, oversight. He matches tools to kinds of risk (emergent properties are managed; manipulative design is regulated). He calls for building everyone’s capacity to thrive (“a flood can’t be halted, but it can be directed”). That is a navigational, capacity-building view of AI governance. He knows it sounds “rather bland” and argues it will prove more durable.
Learning theory as a lens on AI. He reframes students’ AI use from a cheating problem to Deweyan inquiry: the conversation, “messy” and non-linear, is the evidence of learning, and “because of the messiness, not in spite of it”. Few AI-in-education voices turn the assessment question from policing to understanding learning journeys.

7. Posts that best reveal how he thinks#

2025-07-27 spiky-surfaces-and-jagged-edges-moving. The clearest window on his intellectual method. A colleague’s model is respectfully pushed, playfully redesigned step by step with “what if” questions and physics-inspired metaphors, and opened to non-STEM ways of knowing. He is explicit that models shape what can be thought, and ends on “missed opportunities at best, and dangerous blindsides at worst.”
2025-06-01 vibe-coding-moral-panic. Serendipity and play drive the inquiry, and building sharpens the concepts. It holds his one explicit statement in the batch of threat to “what’s important to people” as a way to understand and navigate, and his refusal to mock public fear.
2025-07-13 whats-grok-4s-moral-character. Building to find out. It shows his experimental design instincts (de-identification, decoys, repeat runs), fairness to a controversial company, humility about robustness, and an open invitation to readers.
2025-08-31 holding-on-to-our-humanity-age-of-ai. His governance stance in its fullest form: an unprecedented challenge; regulation augmented, not replaced; risks matched to tools; the flood metaphor for navigating rather than controlling; and responsibility shared by “each one of us”. (Read without its Afterword.)
2025-07-23 americas-ai-action-plan. His values under pressure. Perspective-taking vignettes, fairness despite declared unease, risk framing read as a map of values, the nanotech precedent of navigating risks and benefits, the deliberately human first take, and a public correction.
2025-08-24 using-ai-to-assess-student-ai-conversations (read with its prequel 2025-08-17 stop-asking-students-show-me-your-prompt). A cheeky demonstration becomes a serious question. It shows Dewey-driven values about learning, “noodling” openly admitted, and a scrappy tool handed over for others to take apart.

Worth adding for what it shows about what moves him: 2025-07-20 still-human-61-inspiring-paintings. It is short, but it is the batch’s most direct evidence of what he is protecting.