B21 perspective notes: 2025-01-14 to 2025-02-25 (14 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. All fourteen posts were read in full. The six Modem Futura posts (2025-01-14, 01-21, 01-28, 02-11, 02-18, 02-25) were read for his own framing lines only; the audio was not heard. Quotes are exact, including his typos. FFTF page numbers are the PDF page markers in working/maynard/book/, which match the printed pages.

Evidence rules applied

2025-01-30 ai-at-a-crossroads. The title, subtitle and entire article body were written “entirely by ChatGPT”, and ChatGPT chose the image. None of that is evidence. His italic preface, his Postscript, “A word on the process” and his two prompts are his own words. The first prompt contains a self-description written so that ChatGPT could match his style. It is a statement of intent, not a neutral account, and is treated that way.
2025-02-16 the-artisanal-intellectual-in-the-age-of-ai. The central article was drafted by Deep Research and then line-edited by him. It refers to “Maynard” in the third person, and his sentences cannot be separated from the AI’s. It is therefore not used as evidence of his thinking. Only his introduction and closing Notes are used. The attached “Missing Perspectives” paper is AI-generated and unchecked (he says so).
2025-02-04 openai-deep-research-ai-scholarship. The abstract of the attached paper and Deep Research’s replies are AI output. His side of the transcript is his own prose, and it is among the most revealing text in the batch: he is telling a research agent how to think about his own field. Mark Daley’s line about staying “in the game” is Daley’s.
2025-02-09 can-ai-write-your-phd-dissertation. The dissertation, its abstract and the “Originality Score” block are AI output. The driving question and its three sub-questions are his own.
2025-01-28 the-500-billion-ai-gamble. The title was generated by DeepSeek, as his subtitle says. The body is his.
2025-02-02 toe-tapping-ai-pizza-joy. The song, its lyrics and Jacob Seipert’s explanation are not his. Only his framing and footnotes are used.
Modem Futura timestamp labels (2025-02-11, 02-18, 02-25) appear in his posts but may be shared show notes with co-host Sean Leahy. They are used only as signs of what he chooses to put in front of readers, not as arguments.
2025-01-14 are-your-cars-autonomous-features-unsafe. The ADAS expertise is Brunno Moretti’s. Maynard’s contribution is what struck him.

Context. Six weeks at the start of 2025: the $500B Stargate announcement, the rescinding of Biden’s AI Executive Order, DeepSeek R1, the International AI Safety Report and the Vatican’s Antiqua et Nova in the same week, OpenAI’s Deep Research, the Arc Institute’s Evo 2 genome model, and the 50th anniversary of Asilomar. He is teaching “Pizza and a Slice of Future” at ASU and recording Modem Futura weekly; the 2025-02-25 episode is the 20th. He is paying for OpenAI’s o1 pro tier and using it hard. The batch catches him at the most hands-on, experimental point of his AI writing so far. Over three weeks he builds with AI (the games), commissions work from it (a framing paper, a dissertation), co-writes with it (the artisanal intellectual article) and hands it his byline (the Vatican/safety-report piece). Each post tests a different relationship between a person and the machine, and each stakes something of his own.

1. How he thinks here#

He finds out by building, and he escalates the test#

Almost every substantial post in the batch is an experiment he designed, ran and reported.

2025-01-26 i-asked-chatgpt-to-create-three-video-games. The post is organised around a question: “Could I get it to create fully functioning code for three very different web-based video games, with no direct coding input from myself?” He sets “three progressively challenging tasks”: a clone, then a game inspired by another, then an original game. He fixes his own role as a constraint (describe, ask for changes, report bugs), states what the test is not (“not to show how effective and reliable generative AI is in a professional coding environment”), and publishes the playable results.
2025-02-09 can-ai-write-your-phd-dissertation. “I was intrigued to know how much further I could push it.” The first two attempts were “a train wreck”. He redesigned the method (roadmap first, then eight chapter prompts carrying forward the foundations) and documented it.
2025-02-16 the-artisanal-intellectual-in-the-age-of-ai. The next step is collaboration, because “AI-only “intellectual labor” is — at least at the moment — less interesting to me than what a person and a powerful reasoning/research AI might be able to achieve together.” His last move was to feed the finished article back to Deep Research “and ask it what I had missed.”

Failure is reported in the same voice as success. The original game “failed miserably”; the dissertation attempts were a “train wreck”; the day after his Deep Research post he added that some runs were “just mediocre — and with plenty of false references” (2025-02-04 update).

He tests the machine on his own ground#

He does not test AI on benchmark tasks. He tests it on the things he knows best, so that he can judge the output.

The framing paper (2025-02-04) is on “navigating advanced technology transitions”, his own emerging field, chosen because it “is not a widely recognized area of scholarship”. The point was “to see how much Deep Research was able to replicate my own thinking — or to challenge it.”
The artisanal intellectual (2025-02-16) is a concept he had just coined, so the AI had little to copy.
The Crossroads prompt (2025-01-30) asks ChatGPT to write in his style.

He is candid when he may have contaminated his own test: “I may have hinted at my own ideas too much in the prompt (you can check this by comparing the conversation with Deep Research and the subsequent paper below).” He names the flaw and gives readers the means to check it.

Serendipity sets the agenda, and he says so#

The dissertation question “started its life in the shower”. “Somewhat randomly I came up with the question” (2025-02-09).
The artisanal intellectual “was a bit of a throwaway at the time”, a footnote, before it grew through panels and a podcast into “a foundational paper I will be using” (2025-02-16).
The Evo 2 release “scuppered that plan” to write about permissionless innovation (2025-02-23, fn 2).
Misreading McLuhan’s title on air (“I fall into a literary trap head first!”) opened “such a serendipitous and revelatory conversation” (2025-02-18).
The podcast’s value is its “often-serendipitous conversations that go in unexpected and intriguing directions” (2025-01-21).
He asks the machine for surprise: “I want to be surprised rather than push you in a certain direction”, and wants “ideas and insights that are new to me, even though I know the field very well” (2025-02-04 prompts).
Windmill, the game he co-designed, “is a game of discovery and serendipity” (2025-01-26). He built serendipity into the artefact itself.

His own statement of method: playful, speculative, grounded#

A footnote to a podcast note (2025-01-21 is-this-the-year-of-agentic-ai) describes where his ideas come from:

“the best conversations are those that are playful and speculative, while being grounded in a solid and serious foundation of understanding.”

He came away with “my brain sparkling with new ideas and insights”, and says the podcast lets “others get to listen in” to that kind of academic conversation: “there’s something special to me in this.” This is the closest thing in the batch to a statement of method, and the Evo 2 post follows it exactly.

“Imagine…”, then a plausibility check#

2025-02-23 evo-2-dna-ai. Five paragraphs in a row begin “Imagine”: precision gene editing, new molecular machines, bio–machine interfaces, DNA as a functional material, and hybrid systems designed by coding DNA and atoms together. Then he steps back: “Admittedly this is all beginning to feel a little sci-fi”. Then the test: “Speculative as this is, it’s not beyond plausibility”. The justification is an observed rate of change: just over two years from early ChatGPT to systems that “simulate reasoning, carry out complex research, and accelerate discovery”. The speculation is labelled as speculation, and its warrant is stated.

Analogy by structure, with the break point marked#

DNA as language (2025-02-23). DNA “is just another type of language that connects a sequence of symbols with functional outcomes.” That shared structure is why a next-token model can work on genomes.
Reclaiming a dismissive term. Evo 2 is “a DNA-based stochastic parrot — except that it’s ability to “parrot” biology far exceeds anything humans are capable of on their own.” He takes the critics’ phrase and shows that it describes a capability, not a limitation.
Where the analogy breaks. The model has no training on how genes relate to phenotypes, “much like an advanced text-based AI model can generate beautiful text that emulates understanding, but which isn’t actually grounded in understanding.”
Asilomar 1975 (2025-02-23) is a structural precedent (scientists taking “collective responsibility”) faced “on what now looks like a much smaller scale”.
BASIC as a gateway (2025-01-26, fn 1). Quick HTML games might lead novices to GitHub the way BASIC led his generation to professional platforms.

He reframes by changing who, or what, the question is about#

From professionals to novices (2025-01-26). Most debate about AI coding is “between people who are already adept at coding.” “But what about coding novices?” Changing the population changes the answer.
From product to process (2025-02-04). After concluding that it is “hard to imagine serious scholarship without tools like this surviving”, he adds: “Unless, that is, the future value of the scholarship lies primarily with the process of creation rather than the relevance and impact of what is created.” The question moves from whether the AI is good enough to where the value of scholarship lies.
From a low ranking to a limit of method (2025-01-19). The WEF’s low ranking of AI is reframed twice: AI is “an enabling technology” whose harms show up under other headings, and aggregated expert opinion has a built-in blind spot (section 3).

He holds tensions open#

What a PhD is for (2025-02-09). He sets out both answers: “more about the journey than the new knowledge it produces”, or “pushing the bounds of what is known”. He shows what follows from each without choosing.
The writer’s craft (2025-01-30). “As a writer, using generative AI to create copy scares me profoundly.” He published the AI piece anyway, gave four reasons, and left the discomfort showing.
Deep Research (2025-02-04). “Brilliantly frustrating about sums it up!” He “felt bereft” when the session broke, and says the paper was “flawed and unfixable”.
Handing the unease to the reader. “And if you find that a little disconcerting, you probably should” (2025-02-04). He neither soothes the reader nor sounds an alarm.

Generativity is the test he applies#

“Generative” is his most frequent evaluative word in this batch. Accuracy matters to him (section 2), but the higher bar is whether something produces new thinking.

To ChatGPT: “I’m looking for new ideas, not simply synthesis of existing ones. Be generative” (2025-01-30 prompt).
To Deep Research: “it will need to be generative”; “I want you to be intellectually generative here” (2025-02-04 prompts).
His verdict: “not as generative and insightful as a leading transdisciplinary and polymathic scholar would most likely be” (2025-02-04).
On the collaboration: “genuinely generative”; “just how synergistic and generative this collaborative process was” (2025-02-16).

2. What matters to him#

Joy, delight and the amateur’s pleasure#

2025-01-26. He asked ChatGPT for an original game “that focused on bringing joy to the user rather than competition”. His follow-up prompt asked for “five new features that will bring joy to users”. Joy is written into the specification.
“The second is that this is not professional-level coding, but I don’t care.” What mattered was re-living “the pleasure I got out of writing simple code as a teenager”, and “for all that seeming banality of using generative AI to create simple games, that’s important.”
Windmill is “the game that I had — and still have — the most fun with.”
2025-02-02. A student’s AI song about his class: “I couldn’t stop grinning as I listened to the song — and I challenge you not to do the same”.
Podcast labels (2025-02-18, 02-25): “Technological playgrounds vs playpens”, “The value of being an amateur rather than a professional”, “Are professionals constrained to playpens of the imagination?”, “The need for space to play”, “Do we have a creativity problem in research?”, and “the regenerative power of simply doing nothing”. Taken together they show play, amateurism and unstructured time treated as serious conditions for creativity, and professional training as a possible constraint on it.

Opening capability to anyone#

Conversational coding “begins to redraw the line between what someone can imagine, and how that’s recreated on a digital platform” (2025-01-26).
He chose plain HTML because the barrier to entry is “near-negligible” and the code is “viewable, downloadable, and buildable-on by others”. He invites readers to remix: “anyone download the code, give it to ChatGPT, and come up with even more creative versions!”
He puts himself “in the shoes of someone who’s never created a web page before” and closes: “I’d encourage anyone who’s interested to start experimenting”.

The person in the work#

“part of my craft is the very human piece of me I bring to my writing — my expertise, the connections I make, the idiosyncrasies that make my writing an expression of who I am, and even the flaws that litter my work” (2025-01-30 postscript).
What the machine lacked: “the subtle nuances and modulations of expression that a good human writer would insert” (2025-02-04).
“I do not like bullet points or short sections – prose that develop ideas are preferred” (2025-02-04 prompt). He thinks in developed prose.

Being able to check claims#

“My rule on formatting and citations is that they need to be consistent and the reader needs enough information to find the source and verify it” (2025-02-04 prompt).
Missing in-text citations are “a big deal as there’s no way to validate claims in the text” (2025-02-04). “Where they didn’t seem correct I’ve flagged this in the paper.”
He went through the artisanal draft “line by line and link by link”, tracing every reference to its primary source (2025-02-16). The unchecked follow-on paper carries a “caveat emptor”.
He was unimpressed that Deep Research relied on easy websites: “it even used Goodreads at one point!” (2025-02-09).

Breadth across disciplines#

Deep Research’s advantage over most PhD students is that it gets past “the restrictive training and perspective that comes with disciplinary boundaries” (2025-02-04).
His prompt lists philosophy, STS, responsible innovation, policy and governance, sociology, behavioural science and futures studies, “and scholars that cross over these areas”.
“a leading transdisciplinary and polymathic scholar … But such scholars are rare” (2025-02-04).
He read a Vatican document and a scientific safety report side by side because of the “generative insights that can be teased out of approaching them together” (2025-01-30 postscript).

Human flourishing and what it means to be human#

In his own framing, the transitions paper “should consider the nature of value within a human context, even down to what it means to be human” (2025-02-04 prompt).
His dissertation sub-question asks about “emergent concepts and ideas around human survival and flourishing” (2025-02-09).
A student’s joke song still prompts: “the theme of what it means to be human in the future feels especially relevant at the moment” (2025-02-02).

His students#

He apologises to his current student “for throwing this potential wrench in their well-laid plans” (2025-02-09, fn 1). He expects his graduated students to read the formatting ordeal “probably with a touch of schadenfreude”. He credits Jacob’s song as “totally unprompted!” (2025-02-02).

What frustrates him#

An industry culture where “the name of the AI game is increasingly to go fast and break things in the hope that someone else will clean up the mess” (2025-02-23).
Responsible AI “going out of fashion at lightening speed” (2025-02-23).
Mainstream experts who “simply do not grasp how disruptive the technology may turn out to be” (2025-01-19).
A research tool that would “bend over backward to incorporate suggested areas and leads from our initial discussion rather than “thinking” for itself” (2025-02-04).
Every AI redraft being a rewrite: “If you see something that you like that just needs a few tweaks to move it from good to great, good luck!” (2025-02-04, fn 2).

What delights him#

The Windmill game; Deep Research’s progress report (“I was loving this”); the tool finding and citing his day-old podcast “with no prompting” (2025-02-16); a student’s song; a misread book title; and a long-held curiosity: “I’ve long been intrigued with the parallels between coding using DNA and coding using digital ones and zeroes” (2025-02-23, fn 1).

3. Risk as a way of thinking#

The batch has no explicit “risk innovation” vocabulary. His risk thinking shows in five places, and in each it widens what counts as a risk rather than supplying a procedure.

Novel risks and the limits of mainstream expert judgement#

2025-01-19 wef-global-risks-2025 is the batch’s clearest argument that new kinds of risk need a different way of seeing.

He knows the instrument from inside. He has contributed to the WEF expert survey “for almost as long as it’s been around” and is “especially interested in how technological risks are framed and ranked.” He reads the rankings year on year (AI 32nd on the two-year outlook and 6th on the ten-year; 29th and 6th in 2024), and treats the numbers as data about perception shaped by method, not as measures of risk.
The mechanism. Survey opinion tends “to regress to the mean in terms of understanding and perception.” He credits what that protects against: it “helps avoid a dangerously high focus on speculative risks at the expense of more immediate and plausible concerns.” And he names what it costs: “it does have a tendency to devalue risks that are poorly understood by a broad base of mainstream experts.”
The consequence. For frontier technologies and advanced AI, disruptions will be ““I told you so” moments for people immersed in the various fields, but blindsides to many other professionals and experts”.
Where he stands. Critics will read the low ranking as a sign “that saner minds are seeing through the speculation.” He gives that view a fair hearing, then disagrees, with a hedge (“I suspect”).

The same idea appears three weeks later in his dissertation sub-question: “What can be learned from thinking at the edge of the distribution rather than in the mainstream” (2025-02-09). A podcast label asks the underlying question directly: “When does tech history stop repeating itself?” (2025-02-18).

This is quantitative risk literacy used to question the method, not to reject it. He shows why an averaging method, however sensible, will lag behind a technology that does not fit earlier patterns.

Navigating transitions: a frame that asks for new thinking, held with humility#

His prompts to Deep Research (2025-02-04) are the fullest statement in the batch of how he frames his own field.

The premise. “the idea of advanced technology transitions is that we need new models, frameworks etc to understand how to navigate transformative, convergent and synergistic technologies.” The paper should be “developing new ideas on the assumption that conventional ways of thinking, understanding the world, and making decisions, may not work in the future.”
A landscape of possibility. Technologies are “perturbing the local, national and global landscape in terms of shutting down and opening up possibilities.” Risk and benefit are described as one terrain, with possibilities closing as well as opening. There are also “hidden trends that are equally as important”.
Novelty is a question, not an assumption. The paper “must look at the historical context and what we can learn from it” but “also needs to tackle what is unique about this point in human history and whether this justifies new thinking.” His main criticism of the result was that it said too little about “what makes this point in human history unique”. His dissertation sub-questions have the same structure: “where are the commonalities and points of departure from previous points in history, and what can be learned from the past?” (2025-02-09).
Humility as an instruction. “be humble in your writing – recognize that bold and radical ideas are worth considering, but may need testing or even be wrong.” And: “No recommendations at this point – remember the humility bit. I don’t think you’ll have enough for recommendations. But I do think you should be able to explore possible next steps and possible consequences of following certain pathways.”

This last instruction is the clearest evidence in the batch that his concepts are meant as mental models, not operating procedures. In a poorly understood domain he rules out recommendations as overreach, and asks instead for pathways and their consequences. Exploring possible routes rather than prescribing one is what navigating, as opposed to managing, means in practice. It fits FFTF’s pairing of humility with plausibility: “humility alone isn’t enough. There also has to be some measure of plausibility around how we think about the future risks and benefits of new technologies” (FFTF p.168).

Beyond the known hazard#

2025-02-23 evo-2-dna-ai shows managing and navigating side by side.

He credits the managing moves. The team were “cognizant of the potential unintended consequences”. “Rather smartly” they left pathogenic virus genomes out of the training data and red-teamed the model.
He says they are not enough. “But I suspect that the domain of unexpected consequences from this emerging technology to human and environmental health and wellbeing go way beyond harmful viruses.”
His remedy is a capability, not a control. He wants teams to bring in experts “with a deep knowledge and cross-disciplinary understanding of how to successfully navigate highly disruptive and deeply complex technology transitions”. The goal is that the capability “can be steered toward positive outcomes.”
Precedent and difference. Asilomar 1975 is the model of collective responsibility; the 1975 challenge now looks “much smaller”. He took part in the 2017 Asilomar AI meeting himself (FFTF p.169).
The permissionless turn. He notes responsible AI giving way to “permissionless innovation”. In FFTF he traced that lure to curiosity, “our innate curiosity, our desire to know, and understand, and create”, and admitted it in himself, recalling a PhD all-nighter where he “risked damaging millions of dollars of equipment by bending the rules” (FFTF p.161). My interpretation: the same drive shows in 2025-01-26, fn 4, where the pull of coding “where time simply disappears” nearly brought “everything crashing down”. He knows the pull of building from the inside, which is why his criticism targets a culture (“go fast and break things”) and not curious people.
Orphan-like risk (my interpretation; he does not use the term). “in the hope that someone else will clean up the mess” describes risk handed on to people who neither created it nor own it.

Threat to value: the scholarship experiments#

The Deep Research posts are a threat-to-value analysis in everything but name. What he sees at risk is not a hazard but where the value of human intellectual work lies.

The value of scholarship may move to “the process of creation rather than the relevance and impact of what is created” (2025-02-04).
Human-only research may be relegated “to a class of artisanal intellectualism where the primary purpose is the provenance and process, not the product” (2025-02-09, fn 2).
The tool is “forcing me to rethink what it means to be someone who makes a living by thinking” (2025-02-16).
“there’s unique value to what I write because it’s written by a person” (2025-01-30).
A podcast label asks the question outright: “What do we stand to loose with advanced AI?” (2025-02-11).
On the benefit side, the values are joy, access and generativity (section 2). A conventional risk frame would not register any of them.

Hidden and unpredictable risk#

ADAS (2025-01-14). What struck him was “the hidden dangers of the sensor-based safety features”: safety systems that are not recalibrated after crashes become hazards. The risk comes from the risk-reducing technology itself and falls between manufacturers, repairers and drivers. (The orphan-risk reading is mine.)
Stargate week (2025-01-28). The challenge is “steering accelerating AI capabilities toward benefitting society, rather than leading to a whole flock of black swan events”. The verb is steering, and the risk is the unforeseeable.

How these ideas function#

They widen what is seen: from the pathogen to the whole domain of unexpected consequences, from a ranked list to “a deeply complex and interconnected risk landscape” (2025-01-19), and from AI harms to the value of human intellectual work. He respects operational tools where they exist (Evo 2’s data exclusion and red-teaming) but treats them as insufficient for a technology of a new kind. He also refuses to turn the frames into prescriptions before the understanding is there (“No recommendations at this point”).

Balance note. In the three scholarship posts his attention to risk is light compared with his enthusiasm. Verification is the main hazard he names, and the “disconcerting” line flags unease without developing it. The Evo 2 post carries most of the batch’s explicit risk reasoning.

4. Scholarship and public writing#

The Substack as an open lab notebook#

He publishes the whole apparatus.

2025-02-04: the full prompt conversation, the resulting paper, and a same-day update correcting his first impression.
2025-02-09: every chapter prompt and all three foundation files.
2025-02-16: the AI’s original draft, his edited version and a document showing the differences: “Just in case anyone’s interested, you can see how I added to and changed the original draft”.
2025-01-30: his unedited prompts, typos included (“in getting the unadulterated prompts you can see where I’m a little numerically challenged!”).

Transparency serves two purposes. It lets readers test his judgements, and it makes the process itself the contribution. The games post does the same by releasing playable, view-source code.

Public writing incubates scholarship#

The artisanal intellectual moved from a footnote (2025-02-09) to panel discussions and a podcast (2025-02-11) to “a foundational paper I will be using as I continue to explore the idea” (2025-02-16). Its meaning shifted along the way (section 5). The concept was formed in public.
Navigating advanced technology transitions is visibly his research programme. He had used o1-pro to draft “a series of seven papers on foundational thinking around navigating advanced technology transitions”, then set them aside because pursuing them “would be a waste of my time — such is the rate of AI progress” (2025-02-04). The Substack is where the programme is being worked out, and where he tests whether AI can advance it.
The WEF analysis (2025-01-19) extends his two long 2024 posts on the report’s history, which he calls “important context”. It is a running public analysis of one data source.

Translating, then taking his own step#

2025-02-23 explains base pairs, context windows, genome scale and phenotypes in plain terms before making his own argument (beyond pathogens, Asilomar, transition expertise). 2025-01-19 gives the numbers first and his interpretation second.

How he treats evidence and expertise#

He respects expertise and says where it runs out. Immersed specialists see what “a broad base of mainstream experts” miss (2025-01-19).
He learns in public from guests: the ADAS conversation “was an eye opener for me” (2025-01-14).
He credits other experimenters by name (LaMoth, Siikaniemi, Gill, his colleague Riz Virk) and points to Mark Daley’s essay (2025-01-26).
He applies his citation rule to AI output and reports the results, including failures.
He anticipates the argument that AI scholarship is “a mere smoke and mirrors illusion” and answers it with the artefact: read it yourself.

Transdisciplinarity as the default#

A Vatican document read alongside a scientific safety report (2025-01-30).
Heidegger, McLuhan, Postman and Socrates in the podcast conversations (2025-02-11, 02-18).
His nanoscience roots in the Evo 2 post: DNA origami, DNA data storage and computing, “nanoscale machines to be integrated with biological systems”.
His own coding history as evidence: BASIC, then Fortran 77, then “the unforgiving syntax of Mathematica” (2025-01-26, fn 2).

Register: his own description#

In prompts meant to reproduce his work, he describes his register twice.

“Imagine the tone of a book that has been written by a leading scholar that is accessible to everyday readers, but is also scholarly. There are many scholars who write in this way for a public audience, so you should be in good company” (2025-02-04).
“expert yet approachable and een a little playful at times” (2025-01-30).

The posts match. He explains HTML, CSS and JavaScript in a parenthesis, apologises for a pun (“get under the hood (sorry!)”), jokes that Deep Research is Douglas Adams’s Deep Thought, reports that “My wife thinks that I overworked it!”, and confesses to using “the ASU Dissertation Wizard” and then escaping “the dissertation formatting police”.

5. His role as he sees it#

His own account#

Written for ChatGPT to imitate, so it is a statement of intent (2025-01-30):

“I strongly aim for my writing to be inclusive and to invite thinking and discussion while offering new insights – I do not write pieces that are polarizing, that are preachy, that push an ideology or an agenda, but rather these are articles that encourage people to join me in thinking deeply about technology, society and the future. That said, I do place human wellbeing and flourishing at the heart of my work.”

It names his refusals (polarising, preaching, pushing an agenda), his posture (“join me”), and his one declared commitment (flourishing). The batch largely bears it out.

A fellow experimenter, not an oracle#

He invites readers to try things themselves: “You may be surprised with what you create with your AI coding sidekick!” (2025-01-26).
He asks for engagement on informed terms: “please do use the comments, but do read the paper and the process first before providing opinions” (2025-02-04).
“Let me know what you think of the experiment and the ideas” (2025-02-16).

With sceptics#

“Of course, there will be skeptics”. He asks them to read the dissertation “with an open mind”, and allows that “You may still conclude that there’s nothing to be seen here. But you may also be surprised — and jolted into thinking differently” (2025-02-09). This is an invitation, not an insistence. With the WEF critics, he first states their reading (“saner minds”) and only then disagrees.

An insider who puts his own profession in question#

“For the first time in my experience as a pretty well established and respected academic, it feels like AI is capable of extending what researchers, academics and scholars can achieve beyond anything we’ve seen before” (2025-02-16).
“if I was doing my PhD now without the help of AI, I would be deeply worried” (2025-02-09).
He casts himself, with some self-irony, as the artisanal intellectual in his own experiment: “the craft of the “artisanal intellectual” (i.e. me in this context)” (2025-02-16).

He applies his questions about disruption to his own work first.

Industry#

He is a paying, hands-on user who reports strengths and flaws plainly, and compares OpenAI with DeepSeek on the same task. He credits the Evo 2 researchers by name for their precautions while criticising the wider culture of speed. He records the policy moment (Stargate, the rescinded Executive Order) without taking a partisan line.

What he refuses to do#

Over-prescribe: “No recommendations at this point”.
Overclaim for AI: “To be very clear, this is not a PhD dissertation — AI isn’t there yet.” Independent AI scholarship “would require independent intent and understanding on the AI’s part” (2025-02-09, fn 2).
Wave away AI, or catastrophise about it. He acknowledges the “smoke and mirrors hype” debate, and says unease is the right response (“you probably should”) without escalating it.
Hide his contradictions. He breaks his own rule on AI-written copy in public and explains why.

Changes of mind, made visible#

“I’ve resisted diving into the coding abilities of platforms like ChatGPT for a while now” (2025-01-26); then he did.
“I’ve resisted doing this for so long” (2025-01-30); then he broke his rule.
“the generative AI of today is most definitely not the generative AI of 2022” (2025-01-26). This explicitly revises his view of what these systems are, away from “smooth but vacuous text”.
The seven o1-pro papers were overtaken within a day by Deep Research (2025-02-04), and his verdict on Deep Research was itself qualified within a day (the Feb 4 update).
The artisanal intellectual shifts meaning within a week: first, human-only research valued for “provenance and process, not the product” (2025-02-09); then “someone who thinks without using AI” (podcast label, 2025-02-11); then the craftsperson inside an AI-augmented process, checking and building on the machine’s draft (2025-02-16). He does not say the meaning has moved.

Tensions a careful reader might note#

Peak enthusiasm. “hard to imagine serious scholarship without tools like this surviving”, “a wonderful synthetic research partner”, “felt bereft”, and a comparison with a National Academies-style analysis all rest on a few runs over a few days. He adds counterweights (the update, “flawed”, “not a PhD dissertation”, the intent-and-understanding limit), but the tone is more boosterish than elsewhere in his writing.
The craft tension. He values the idiosyncrasy and even the flaws of human writing, yet publishes an AI-written article and asks the machine to imitate his style. He names this himself, which is his usual way of handling a contradiction.
Access. A podcast label lists “The value of OpenAI’s $200 a month plan” (2025-02-11), but his own prose here does not ask who can afford the tools his experiments depend on. (In the hybrid artisanal article, the “new inequality” point cannot be attributed to him.)

6. What is distinctive#

Self-disrupting experiments, published with their apparatus. Most AI commentary either cites benchmarks or offers opinion. He tests reasoning models against his own expertise and his own profession (his field, his concept, the PhD, his byline), publishes the prompts, drafts, differences and failures, and invites readers to check him. The experiment is the argument (2025-02-04, 02-09, 02-16, 01-30).
Joy, discovery and amateur creativity as criteria for judging AI. While the debate was measuring AI by productivity or harm, he wrote joy into a design brief, built a game of “discovery and serendipity”, and said of professional standards “I don’t care”. Podcast labels on playgrounds versus playpens, and amateurs versus professionals, point the same way (2025-01-26, 02-02, 02-18, 02-25).
Asking where value lies, not whether AI is good enough. The artisanal intellectual turns “can AI do scholarship?” into “is the value of scholarship in the product, the process or the provenance?” In effect this is a threat-to-value analysis of intellectual work, and it leaves several futures open (2025-02-04, 02-09, 02-16).
A method-level critique of how expert consensus handles novel risk. Regression to the mean in expert surveys protects against speculation but “devalue[s] risks that are poorly understood by a broad base of mainstream experts”. He makes this point as a long-time contributor to the WEF survey, and it pairs with his interest in “thinking at the edge of the distribution” (2025-01-19, 02-09).
Past the obvious hazard. On Evo 2 he credits pathogen exclusion and red-teaming, then argues that the real domain of unexpected consequences is far wider and calls for people who can navigate transitions, not just more controls. He reads Asilomar’s 50th anniversary as a precedent on a smaller scale (2025-02-23).
Humility built into the method, even when delegating. He tells a research agent to be humble, to treat bold ideas as possibly “wrong”, and to make no recommendations because the understanding is not there yet. It is a rare instance of guarding against false precision written into the instructions for producing knowledge (2025-02-04).
Reading across ways of knowing. He paired a theological document with a scientific safety report specifically for the insight their combination might yield (2025-01-30 postscript), and his own framing of transitions reaches down to “what it means to be human”.

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

2025-02-04 openai-deep-research-ai-scholarship. His prompts are the fullest account in the batch of how he frames navigating advanced technology transitions: new thinking tested against history, a landscape of possibilities opening and closing, value down to “what it means to be human”, humility and no recommendations. The post shows him testing the tool on his own field and then reframing from product to process.
2025-01-26 i-asked-chatgpt-to-create-three-video-games. Play and joy as criteria, experiment design, the move to novices, open sharing, nostalgia, and a stated change of view about what generative AI is.
2025-02-23 evo-2-dna-ai. Analogy by structure with its limit marked, “imagine” followed by a plausibility test, managing versus navigating, Asilomar as precedent, and criticism of permissionless culture.
2025-01-19 wef-global-risks-2025. Quantitative literacy used to question an averaging method, novel risks lost in the mainstream, and a fair hearing for the view he disagrees with.
2025-02-09 can-ai-write-your-phd-dissertation, with 2025-02-16 the-artisanal-intellectual-in-the-age-of-ai (intro and Notes only). Serendipity (the shower, the throwaway footnote), both purposes of a PhD held open, the artisanal intellectual formed and revised in public, radical transparency, and asking the machine “what I had missed”.
2025-01-30 ai-at-a-crossroads (preface, postscript and prompts only). His own statement of role and refusals, his fear for the writer’s craft, and a rule broken in the open with reasons.