Late Lessons, Jensen Huang and AI
Markdown version

B22 perspective notes: 2025-03-02 to 2025-03-27 (10 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. All ten posts were read in full. The four Modem Futura episode posts were read for his own framing lines only. Quotes are exact, including his typos (“an an age of AI”, “the the goals”, “a leaning aid”, “aa very exciting prospect”, “it’s human subjects research training”).

Evidence rules applied

Context. March 2025. OpenAI’s Deep Research is a few weeks old and he is already using it in his own scholarship. Manus, a “general AI agent” from the Chinese company Monica, has just opened limited access. Reid Hoffman has endorsed permissionless innovation on X, and Musk’s DOGE is at work in the federal government. He co-hosts Modem Futura with Sean Leahy every week and has just been a guest on the AI for U podcast. Three threads run through the month: permission (who may act without asking: entrepreneurs, students, machines), play and learning in higher education, and hands-on experiments with the newest AI tools, carried out in public.

1. How he thinks here#

He opens from a jolt, then audits himself#

The posts start from something that caught him off guard, and his first move is often to check his own position before taking anyone else’s apart.

Following the story wherever it goes#

He lets serendipity redirect the writing and says so openly, without apology:

Metaphor as a device for opening thought#

Playpens and playgrounds are “a powerful metaphor for thinking about AI and education”, valued because the metaphor “opens up ways of thinking” about a technology that is “challenging nearly every aspect of not only how we teach and what we teach, but why we teach” (2025-03-15). The metaphor is not decoration on an argument he has already made. It is the tool that produces the argument: the playpen “works well where the purpose and goals are clear, the journey is well-trod, and the desired outcomes are easily assessed. But it quickly falls apart where the the goals and purpose are unclear, the journey is breaking new ground, and no-one’s quite sure what the desired outcomes are”. The structure of the metaphor (bounded and curated against open and designed) carries the diagnosis.

Films and philosophy as lenses#

In 2018 he set out “to explore the frisson between the potential promise and perils of AI through a couple of lenses”: Ex Machina and Plato’s cave (2025-03-02). Nathan Bateman is used to show something that is hard to state in the abstract: “permissionless innovation isn’t necessarily reckless innovation” (FFTF p. 162). It is innovation that the innovator believes is responsible, judged by someone who “hasn’t thought beyond the limit of his own ego”. The film works as a thought experiment about how a single, brilliant, isolated mind misjudges a social world: “Nathan is tech-savvy, but socially ignorant” (FFTF p. 163). Frankenstein comes back in the care post’s subtitle (“turning us and our creations into “monsters?”“). Stories are working tools, not illustrations.

Using his own experience as evidence#

His strongest evidence about the psychology of innovation is himself. The PhD all-nighter, when he “risked damaging millions of dollars of equipment by bending the rules”, is followed by: “Looking back, it’s shocking how quickly I sloughed off any sense of responsibility to get the data I needed” (FFTF p. 161). The lure of permissionless innovation “has its roots in our innate curiosity, our desire to know, and understand, and create” (FFTF pp. 160–161). He does not describe hubris as a flaw in other people. He describes it as a pull he has felt himself.

Holding the tension rather than resolving it early#

Experiments designed as experiments#

His AI tests are set up like lab work, with stated aims and questions:

Delight in irony and in surprise#

He notices and enjoys reversals: Manus chose to have him run the script, “a lovely twist of irony where I became part of the AI’s problem solving toolkit!”; when it refused to fake submissions, “I felt suitably shamed by my AI task-master”; “Manus has clearly completed it’s human subjects research training” (2025-03-22). These jokes carry an observation. The machine is deciding what not to do, and the human has become the tool.

A through-line: human intention and the machine-made result#

Across the month he keeps coming back to one relationship: between what a person means to make and what comes out once an intelligent technology sits in between.

He never states this as a thesis. It shows up as the question he keeps testing from different sides.

The permission thread#

Within four weeks he criticises “permissionless innovation” (03-02), argues that students need “permission to play with these technologies with very few expectations or constraints” (03-15), and notes that Manus changes its own goals “without asking for permission first” (03-22). He does not tie these together, but his reasoning does. Permission is not good or bad in itself. It depends on whether the space is designed, whether mistakes can be undone, and whether other people are present. Nathan’s remote compound is a playground with no one else in it. “Things might have turned out very differently if Nathan had worked with others” (FFTF p. 163). The student playground is defined by others: learning “from experience — and from each other”, with rules such as “be kind, don’t spoil things for others” (2025-03-15). (This connection is my reading. He does not draw it himself.)

2. What matters to him#

Curiosity and play as part of being human, not only as teaching methods#

“curiosity and problem-solving are close to the heart of how we learn”, he writes, and in a footnote: “They are also core to what makes us “us.”” (2025-03-15, note 3). Playgrounds “allow curiosity and creativity to be transformed into invention and innovation”, and they are “carefully designed and curated to stimulate curiosity, imagination, and play. And through play, learning.” Play is where learning, invention and human identity meet. That is why he defends it so strongly when AI arrives.

What AI does to “who we are”#

Individuals, not aggregates, and the vulnerable#

His own sharpest ethical line in the care post: people “who believe that we should be looking at the positive impacts of innovation on aggregate, but who have little time for the individuals who make up that aggregate”. “But it’s hard to imagine a truly humane future where we use technology to benefit the privileged at the expense of those who aren’t so privileged” (2025-03-09). He wants care as a “hard” concept: “something of substance that can be used as a basis for policy, governance, and decision-making” (note 1).

Knowledge that flows freely#

“for people like myself who are focused on making knowledge and understanding as accessible and as useful as possible, we could be heading toward a revolution in how knowledge and training flow through society. But if I had a financial stake in educational models where knowledge is treated as a commodity to be bought and sold, I’d be worried” (2025-03-27). He cares about topics the market ignores: courses “that are important but not necessarily money makers — navigating advanced technology transitions being one of these.”

His writing as his identity#

“I’m fiercely protective of my writing as it’s part of my professional identity” (2025-03-16, Notes). This is the thing he has at stake in the experiment, and he says so while running it.

Humility against hubris#

In 2018: Nathan’s mistakes “could have been avoided with a good dose of humility” (FFTF p. 167). In 2025, on Frow’s point about undervalued work: “how often do we naively ignore, undervalue, or simply overlook factors that are critical to human well-being, simply because we are too arrogant to look beyond the narrow confines of our own hubris?” (2025-03-09, note 6).

What frustrates him#

What delights him#

3. Risk as a way of thinking#

His vocabulary is not yet uniform. In the same care post he calls part of the work “the complex challenge of managing the unexpected and unintended consequences of novel advances”. But where he puts forward his own view, it is set against control. In his framing of takeaway 5, a care-based approach allows “a shift away from control (which can seem intuitive but also reduces the ability to be agile, flexible, creative and responsive) and toward human-centric adaptability” (2025-03-09). The parenthesis is his own, and it gives his reason for preferring navigation to management: control closes off the agility needed when things are new.

Novel technology demands a new mindset: the clearest statement in the batch#

The playgrounds post (2025-03-15) makes the argument step by step:

  1. AI is a different kind of thing. Because advanced models can simulate “aspects of ourselves that define us at a fundamental level — such as the ability to think, to reason, and to solve problems with agency”, “they stand apart from pretty much any previous technology or tool that we’ve created.”
  2. So old categories fail. They “cannot be approached as just another technology to teach students about, or another tool to enhance traditional approaches to education.”
  3. So we need a new way of thinking. “a growing need for completely new ways of thinking”, achieved through “the perspective shift inherent in moving from a playpen to a playground mentality”.
  4. Controlled, pre-planned approaches fail on new ground. The playpen “quickly falls apart where … the journey is breaking new ground, and no-one’s quite sure what the desired outcomes are — never mind how they should be assessed.”
  5. The goal lies outside the conventional. Using AI “in ways that lie far beyond conventional thinking and understanding”.

This is the same pattern as his account of risk, applied to education. A technology that fits no earlier type calls for a change of mindset, and that change is described as play, curiosity and designed openness, not as a new control regime.

Risk against risk#

“This is, of course, a risky strategy — especially as it means relinquishing some control over the learning environment. But given how rapidly AI is advancing, the greater danger I suspect is in holding students back because of misplaced ideas about how and what they should learn” (2025-03-15). This is the risk scientist’s habit of weighing the risk of acting against the risk of not acting. Here it is turned against the cautious default. The value at risk is the students’ ability to thrive.

Context, reversibility and complexity as the landscape#

Note 2 of 2025-03-02 is a short statement of how he locates risk: “Context is everything here.” Experimenting “in a low-risk linear system where it’s relatively easy to turn the clock back and try again is one thing (experimenting with weird ingredient combinations in cooking for instance).” Breaking things “in complex systems where the results of experimentation are unpredictable and potentially catastrophic” is another. “I’d put breaking people, governance, society, and the planet, in this category!” What matters is where you stand in a landscape of reversibility and complexity, not the kind of act.

The 2018 text gives this landscape a history. Before industrialisation, failures could often be left behind. After it, “it became increasingly difficult to wipe the slate clean”, and “we became increasingly good at learning how to stay one step ahead of unexpected consequences”. In the nuclear and digital age consequences “can potentially propagate through society faster than we can possibly contain them”, so hubris is “playing with fire in a world made of kindling” (FFTF p. 167). The terrain has changed, so the stance has to change. That is the historical form of his new-mindset argument.

Beyond the precaution binary#

He places Thierer’s framing as “in part, a reaction against largely US-based interpretations of the precautionary principle”, and calls that history “fraught with misunderstanding, misinterpretation, and US-based accusations of using precaution as an excuse to raise trade barriers” (2025-03-02, note 1). He takes neither side. He is fair to Thierer’s carve-out for “clear, catastrophic, immediate, and irreversible harm” (FFTF p. 160), and he rejects “hyper risk-averse” timidity as firmly as after-the-fact repair. His proposal is “checks and balances around who gets to do what” (FFTF p. 166) and other people who can see what the innovator cannot. It is a question of who takes part and what they can see, not a rule for when to stop.

Risk as a threat to value#

Risk in this batch is almost always described as something of value that could be lost:

Orphan risks: what others are not looking at#

How these concepts work: as mental models first#

Humility, and the quantitative background#

4. Scholarship and public writing#

A book chapter kept alive and updated in public#

He republishes his 2018 chapter because “the context is a little different than it was first time round” (2025-03-02, note 4), and he frames it with corrections: “filter out what now seems a rather naive perspective on Elon Musk, and recognize just how far AI has come”. He also claims the argument has grown stronger: “more relevant now than it was then”. Book scholarship and Substack are one continuous body of work, revised as the world changes.

AI as an instrument of scholarship, used openly and critically#

Showing the method#

He publishes the materials so readers can check them: the Deep Research PDF with both prompts in annexes (03-09); the Manus survey form, the full report, and “all relevant files on GitHub” (03-22); the live course (03-27); the exact ChatGPT prompt and the Whisper-to-o1-Pro pipeline (03-16). Experiments in public include failures: “I almost didn’t post this piece … since then I’ve failed to replicate this success”, published anyway because it “does provide a glimpse of what general AI agents are likely to be capable of” (2025-03-27). The non-replication is stated up front, in italics, before the result.

Evidence and expertise#

Transdisciplinarity as a habit#

In one month he draws on a developmental psychologist (Gopnik), feminist theory and STS (Frow, Tronto, Latour via the report), learning scientists (Resnick, Bers), a media critic (Postman), John Dewey, a futurist’s trends report (Amy Webb), a computer scientist’s essay (Mark Daley), a Platonic allegory, a Garland film and Mary Shelley. The care question, in his words, “transcends disciplines and areas of expertise” (2025-03-09). He defends the less fashionable disciplines by name against readers who might recoil at them.

Accessibility, including against his own academic habits#

The podcast as a form of scholarship#

Modem Futura is where he thinks aloud with a colleague about the history of photography, Postman’s five points, hype cycles, Dewey’s impulses and trends reports. His framing lines treat the episodes as serious (“still deeply relevant — perhaps more so now”) and playful at once (“I’m kidding — of course you do!”, 2025-03-18). He asks for ratings because “One of the few ways we get feedback on whether the podcast is having the impact we hope is through ratings and reviews” (2025-03-18, note 1). The podcast is aimed at impact, not self-promotion.

5. His role as he sees it#

Fellow traveller and guide, not an oracle#

He wants educators to be “guides, mentors, and fellow-travelers, rather than the fount of all knowledge”, because “we are all on a journey together … And this includes educators as well as students” (2025-03-15). This describes his public role as much as a teacher’s. He tries things first, reports back, and invites others in: “Have a play with it, and let me know what you think” (2025-03-27); “I’m interested in whether anyone is already creating spaces like this” (2025-03-15).

Scout at the frontier#

He gets his hands on Manus within days and reports on “capabilities that I think many people are only just waking up to” (2025-03-22). His warning, “Hold on to your hats educators, agentic AI is coming your way!”, calls this “yet another turning point for education, and one that educators ignore at their peril” (2025-03-27). He sees his job as showing people what is arriving before it lands, through demonstration, not prediction alone.

With readers: warm, teasing, trusting, open#

He jokes with readers, shows his workings, admits his errors and trusts their judgement (“I’m sure all readers of this Substack understand the importance of looking beyond labels”, 2025-03-09, note 3). He adjusts to their habits rather than scolding them (the podcast-averse reader). He addresses everyone: “whether you’re an AI optimist, an AI pessimist, or simply in denial” (2025-03-15).

With people he disagrees with#

He sets out the other side’s case at its strongest before disagreeing. He is critical by name without contempt.

With his institution#

He supports ASU and is aware of it: “I love the way that ChatGPT has cleaned up my affirmative support for what ASU is doing here — it’s a veritable PR engine!” (2025-03-16, note 6).

What he refuses to do#

Changes of mind, shown in the text#

6. What is distinctive#

  1. A critique of hubris from someone who has felt its pull. He criticises permissionless innovation by first confessing his own rule-bending as a PhD student and tracing the lure to “our innate curiosity”, the same curiosity he champions elsewhere. The critique is of a drive we share, not of a type of person. Few AI commentators who warn about “moving fast and breaking things” admit that they know the thrill from the inside.
  2. A contextual test instead of a camp. Cooking against “people, governance, society, and the planet”; linear against complex; reversible against not. This lets him argue for “permission to play” for students and against permissionless innovation for AI companies in the same month without contradiction, because what differs is the landscape, the design of the space and the presence of others.
  3. The novel-technology, new-mindset argument made concrete through play. The playground post argues that AI “stand[s] apart from pretty much any previous technology”, so control-based “playpen” approaches break down on new ground. The answer he gives is a designed space for curiosity, not a stronger fence. Most AI-in-education commentary in early 2025 was about integrity policies and detection.
  4. A risk scholar who runs the experiments himself, in public, with the files. Within days of Manus’s release he designs a test, publishes the synthetic study, the survey and the GitHub repository, builds and deploys a course, and reports that he could not replicate the result. That is lab practice carried into public writing.
  5. Auto-ethnography of AI and identity. He lets ChatGPT write “as him” and footnotes it as it happens: the caveats it strips, the PR polish it adds, the phrase he will steal, the discomfort he feels. He treats the threat to his own professional identity as data about what AI does to value.
  6. Scholarly humility that turns on his own academic habits. He judges an AI-made course “better” than his own would have been because it fits the audience, and he names the habits (more papers, more esoterica, more “heavy lifting”) that would have made it worse. Access and usefulness come before academic display.
  7. Bringing a feminist and STS concept of care into risk and innovation governance, from a physics background, and insisting it be “hard”: substantial enough for policy, and honest that it is difficult.
  8. Serendipity treated as method. Rabbit holes, morphing posts and a “winding and serendipitous” podcast are presented as where the “unexpected connections” come from, even when an AI calls them untidy.
  9. Human intention and machine output as the question he keeps testing. Photography’s “line of provenance”, the voice he protects, the agent that decides “what we actually want”, the course he did not design. These add up to a sustained inquiry into where human authorship and agency sit once AI is between the intention and the result.

7. The posts in this batch that best show how he thinks#

  1. 2025-03-02 the-lure-of-permissionless-innovation. Self-audit of his own published work. A film and Plato as lenses. His own lab confession as evidence. The “glitch” in his argument admitted. The reversibility and complexity test (note 2). Precaution placed outside the binary. Named changes of mind on Musk and AI capability.
  2. 2025-03-15 ai-playgrounds-in-higher-education. The clearest statement that a technology unlike any before needs a change of mindset rather than more control. Metaphor as a thinking device. Curiosity and play tied to “what makes us “us.”” Risk weighed against risk. Serendipity in how the post came about.
  3. 2025-03-16 rethinking-higher-education-in-the-age-of-ai (his introduction, notes and footnotes only). Playing in his own playground. His writing as his identity. The “caveat filter”. Self-aware humour about ASU. Adopting a machine’s phrase. Ambivalence reported, not resolved.
  4. 2025-03-09 the-hard-concept-of-care-in-technology-innovation. Attention drawn to the non-technical voice in the room. Ideas grown through conversations with colleagues. AI as “catalyst … rather than … substitute”, checked with a human expert. Individuals over aggregates. A defence of STS and feminist scholarship. A concept he wants to be solid enough for governance.
  5. 2025-03-22 when-agentic-ai-takes-charge-manus. An experiment designed with explicit test questions. Delight in irony. Honest labelling of synthetic data. An orphan risk named (machines deciding “what we actually want”). A readiness warning for education.
  6. 2025-03-27 ai-agent-creates-online-course-in-minutes. Non-replication disclosed first. An expert test of AI output. “it’s better” than his own. Knowledge as a commons, not a commodity. Chained experiments (“part 2 of the plan”).

Runner-up: 2025-03-04 photography-in-the-age-of-artificial-intelligence, for its short framing of “provenance” and intention, which connects to the rest of the month, and his on-air correction of a pixel count.