Late Lessons, Jensen Huang and AI
Markdown version

B07 — perspective notes (2023-01-31 to 2023-04-12, 8 posts)#

What this batch covers. Two ChatGPT experiments from the Medium period (January and March 2023), then the April 2023 move to Substack, in the weeks after GPT-4 arrived and the Future of Life Institute’s “pause” letter appeared. There are no Modem Futura posts in the batch.

Evidence rules applied. - Counted as evidence: his introductions and postscripts, his prompts in the ChatGPT transcripts, and all the prose in the April posts. - Excluded: every ChatGPT turn, including the self-evaluation drafts, the tenure lists and the mental-health “editorial”. - A caveat on the January prompts: he wrote them for an annual self-evaluation, a genre he says he dislikes. Even so, they are his own unguarded account of how he works, and they are unusually direct about method.

1. How he thinks here#

He opens from himself, and from a feeling. - 2023-01-31 can-chatgpt-take-the-pain-out-of-annual-academic-reviews starts with dread: “I hate writing my annual academic self-evaluation. Scratch that, I dread it.” - 2023-04-12 navigating-advanced-technology-transitions starts with a confession: “This is, by any measure, a procrastination post.” - 2023-04-04 welcome-to-the-future-of-being-human starts with a small embarrassment. He went looking for a link to his own writing on AI risk and found he hadn’t “written much at all that’s neatly citable. … I was shocked!”

In each case the way into a big question is a personal, often self-deprecating moment. The question isn’t posed from the podium.

Experimenting to find out, and treating that as a professional duty. The self-evaluation experiment is a joke with a serious point: “there was method in my madness … as I study the possible societal impacts of emerging technologies like this, I felt obliged to see just how far this might be stretched” (2023-01-31). The experimenting goes further than the task: - He runs a side experiment in politeness, being deliberately courteous to see whether it changes the tone. His answer: “I have no idea whether it did”. - He spins off a playful follow-up (“I thought I would experiment with channeling other writers”).

The tenure post (2023-03-29 maximizing-value-in-the-evolving-landscape-of-tenured-tenure-track-faculty) begins in pure curiosity: “Reflecting on this in an idle moment, I got to wondering”. He is open that the conversation surprised him: “What started as an idle experiment turned into a surprisingly nuanced conversation”. He also names his method: “layered and sequential prompts that allow for context-setting and ideas-development”.

He uses himself as the instrument. He watches his own reactions and treats them as data: - In January he notices that “even though I know there was no real personal connection here, it felt like there was”. He adds that “It also intrigues me and slightly worries me that I’m sitting here already thinking of ChatGPT as a colleague and a collaborator.” - Ten weeks later (2023-04-05 can-chatgpt-adversely-impact-mental) he asks ChatGPT for mental-health help himself. He watches the chatbot thank him for reaching out, “laying the seeds for something that feels like an affirming personal relationship”, and builds his risk argument from there.

The relational pull he first noticed in himself becomes the mechanism behind a societal risk. He doesn’t draw that link explicitly; the arc is my reading.

Reasoning from mechanism, not from counting. The mental-health argument moves from human nature to plausibility: - “we tend to be predisposed to form trusting relationships with people whom we feel “see” and understand us”; - “Language plays a large part in how we develop these relationships”; - so some of these relationships will go wrong, and “as soon as things become personal, it’s very hard to protect against every eventuality”.

There are no numbers, just a well-built causal story. He is careful about the triggering case: it is “near-impossible to tell whether there was a direct link”, but “what is known is disturbing enough to raise questions”.

Reframing is the first move. On the pause letter (2023-04-04 what-are-the-alternatives-to-calling) he doesn’t argue about AI capability at all. He argues about the frame: “A first step is moving away from an ethics framing to one around risk and socially responsible/beneficial innovation.” He places the value of expertise in asking the right questions, not in having answers: “we do have some sense of how to frame the questions that are necessary to move forward.”

Analogy by structure, drawn from the history of governance. He runs through recombinant DNA (bioethics), the Human Genome Project (ELSI), and nanotechnology and synthetic biology (agile, anticipatory and “soft law” governance). He uses them as stages in how governance thinking has evolved, not as comparisons of hazards (2023-04-04). In the Washington post, social media serves as a record of Congress’s institutional capacity, not a comparison of harms (2023-04-10 as-ai-goes-to-washington-whats-being).

Holding tensions instead of resolving them early. - Existential risk but no pause. He hasn’t signed the pause letter, “not because I don’t think there’s a risk of potentially existential proportions emerging here (I do)”, but because he doubts a pause will work. In the same breath he rejects LeCun’s dismissiveness (2023-04-04). - The past teaches, yet this is unprecedented. Past transitions hold lessons that people are “blissfully unaware of”, yet the present transitions are “unlike anything we’ve had to grapple with before” (2023-04-12). My reading is that what transfers is how to navigate (diversity of expertise, partnerships, agility), not what the hazards are. - Speed without haste. He wants “the agility to respond to new challenges with speed” (2023-04-12) and “we need this, like, yesterday!” (2023-04-04). Yet he distrusts those who “mobilize fast” with agendas (2023-04-10). The line he draws is between speed informed by many kinds of expertise and speed driven by a single agenda. - Benefit and harm together. Chatbot support “probably isn’t such a bad thing for many people” but may harm the vulnerable (2023-04-05).

Questions as structure. The tenure post opens with four questions. The Washington post is built on “three questions nagged at me”. Even the subtitles are questions (“are we prepared?”, 2023-04-05).

Writing as a way of thinking. The procrastination post is openly an attempt to “dissipate some of that fog” before a formal journal commentary (2023-04-12). The pause post is “a little rough, but given the speed with which things are developing here, it’s worth posting” (2023-04-04). Public writing comes before formal scholarship, not after it.

Where play, curiosity and serendipity drive the thinking. - Play: the self-evaluation experiment; the “channeling other writers” spin-off; running Midjourney headers as “a broader project on using generative AI to develop images around topics that inspire, challenge, and stimulate new thinking” (2023-04-04 welcome). - Curiosity: “Just out of interest” (a tenure prompt, 2023-03-29); “What especially intrigued me” (2023-01-31). - Serendipity: - a brainstorm at CIFAR gave him the name for both his ASU initiative and this Substack (“I was inspired by the process and the people involved in it!”, 2023-04-07 cifar-launches-bold-new-lines-of); - an idle moment produced the tenure conversation; - failing to find a citable link produced the Substack.

2. What matters to him#

Knowledge that is accessible, legible and useful. This is the most repeated value in the batch: - expertise that is “highly legible and accessible” and “accessible, legible, and meaningful” (2023-04-12); - in his own prompt, “a strong and strategic mission to make new knowledge and insights as accessible as possible to the widest possible audience” (2023-01-31).

He is frustrated that the pieces of the puzzle exist but are “buried in academic research, or within disconnected communities and initiatives” (2023-04-12).

Work that doesn’t fit categories. In his words, “research, scholarship and creative activities and innovation transcend conventional categories of activities”. He does “not approach service as something that is separate from teaching and scholarship, but something that is deeply intertwined with them”. He values “blurring the lines between teaching, scholarship and informal learning through communication and engagement” (all from his prompts, 2023-01-31).

People as individuals, not “humanity”. The future-of-being-human frame is “powerful” because it focuses on “how transformative technologies may impact each of us personally, rather than the rather generally handwaving around “humanity” that usually goes on” (2023-04-04 welcome). Underneath is a stake in agency: “how we can hold on to some semblance of being masters of our own destiny”.

The vulnerable and the unheard. - In the mental-health post his worry settles on “vulnerable individuals” and “a machine that has only the illusion of a reciprocal relationship” (2023-04-05). - In the Washington post it settles on the “quiet voices”: affected communities, developers “buried within layers of institutional structure”, “experts in fields that no-one has realized yet have something important to bring to the table”, and “ordinary people who sometimes have extraordinary insights” (2023-04-10).

Ideas that are unbounded and brave. He delights in CIFAR’s aim to “radically push the bounds of human knowledge and understanding, unconstrained by conventional disciplines and ideas”. He endorses stretching “our thinking and imagination beyond” human exceptionalism (2023-04-07).

What frustrates him: - Institutional inertia: tenure-reform committees producing “muddled and convoluted outputs that rarely lead to progress” (2023-03-29). - Reinventing the wheel with each technology wave (2023-04-12). - Loss of sophistication: AI swinging the conversation back to ethics, led by people who “haven’t necessarily been aware of the sophistication of conversations going on elsewhere” (2023-04-04). - Policy that relies on “old tricks to reign in new technologies” (2023-04-10). - Performative self-promotion: being expected to “puff up what I do for others to critique or, worse, patronizingly pat me on the back for” (2023-01-31).

What delights him. - Bold interdisciplinary programmes: “I’m incredibly excited (I may have said that already) … it’s not going to be dull!” (2023-04-07). - A tool that turns out better than expected (2023-03-29). - Collaborative work that “felt good” (2023-01-31).

3. Risk as a way of thinking#

The landscape and navigation language is already fully present in his own prose. It is used as a mental model more than as a method: - “a deeply complex risk landscape that we are, at this point, unprepared to navigate” (2023-04-04); - “the messy nexus of risks and benefits around emerging technologies” (2023-04-04), with risk and benefit tangled together from the start; - “I wanted to explore the landscape around chatbots and mental health” (2023-04-05); - “a broader landscape of governance options” (2023-04-10); - “navigate the technology transition between where we are now, and a future where AI transforms society for the better” (2023-04-10).

Forms of “navigate” occur about twenty times in the April posts. “Navigating advanced technology transitions” is the umbrella he names for his work (2023-04-04, 2023-04-12).

A novel technology demands a new mindset. This is the clearest early statement in the corpus of what he said in September 2026. His “first step is simply recognizing that we are faced with navigating technology transitions that are unlike anything we’ve had to grapple with before, and that success is going to depend on shaking up how we approach” the problem. We need “unconventional approaches to unconventional challenges” (2023-04-12).

The same logic appears elsewhere: - he asks whether lawmakers are “sufficiently creative in how they approach a challenge that is anything but conventional” (2023-04-10); - on urban futures: “the cities of the future will likely depart from what has been successful so far” (2023-04-07).

Humility against false precision. The risks are “near-impossible to fully map out, and even harder to navigate”. The one certainty is that “there are no silver bullets” (2023-04-04). The metaphor itself is telling: if a landscape cannot be mapped in full, it has to be navigated. It cannot be managed to a specification.

Risk innovation as he framed it then. Risk innovation appears in its early, more decision-support form: “aimed at supporting agile decisions around intangible risks” (2023-04-04). The word “intangible” points toward the later threat-to-value idea. The phrase “threat to value” itself does not appear in this batch.

The tenure post does show a value lens at work outside risk. It asks what “create[s] institutional and social value” and what universities “should avoid if they don’t want to stifle value creation” (2023-03-29). That is a structural echo of risk-as-threat-to-value; the connection is my interpretation.

Orphan-risk thinking, without the term. - In the mental-health post the public conversation is all about chatbots as therapists, while “there’s remarkably little being written about the potential dangers here” (2023-04-05). This is a risk that falls between existing framings. - In the Washington post the risk is to the process itself: an “insights vacuum” that loud voices fill, “introducing deep vulnerabilities into the process of identifying productive pathways forward” (2023-04-10).

Inaction as a risk. “The biggest risk is not taking action or, worse, assuming no action is needed” (2023-04-04). Risk runs in both directions, which fits the risk–benefit framing.

Mental models versus tools. The frames (landscape, navigation, advanced technology transitions, unconventional challenges) do the opening-up. He offers them tentatively: “And, maybe, framing this challenge through Advanced Technology Transitions will help us achieve this” (2023-04-12).

The tools come afterwards and in the plural: - soft law; - agile, anticipatory, outcomes-focused, co- and experimental regulation; - a world congress.

He points readers to practical sources such as the WEF agile regulation toolkit and Marchant et al. on soft law (2023-04-10). None of these is presented as the answer. The frame widens the repertoire, and hard regulation is “a very unwieldy double edged sword” (2023-04-04).

Foundations still visible. His prompts mention co-authoring “one book chapter on aerosol physics” in 2022 (2023-01-31). He judges ChatGPT’s mental-health ideas “through my own knowledge and understanding (including drawing on my background in public health)” (2023-04-05). The physical and health sciences remain the ground he builds on.

4. Scholarship and public writing#

Public writing is the working surface of the scholarship. - The Substack began because his core AI-risk thinking wasn’t “neatly citable” (2023-04-04 welcome). - The Advanced Technology Transitions post is the public warm-up for a “prominent journal” commentary (2023-04-12). - The pause post is a Twitter thread with Gary Marcus, lightly edited into an essay (2023-04-04).

One idea moves across tweet, blog, video (“some of the aspects are captured in this video”) and journal.

His own model of valid scholarship, in his prompts (2023-03-29). - He pushes back on ChatGPT’s first answer: “These are quite conventional and outmoded ways of approaching tenured professor evaluations”. He adds that “there are a growing number of ways to demonstrate rigorous scholarship beyond peer review publications”. - He then offers what is almost a self-portrait: “ground-breaking work on AI and the evolution of social norms” published “in blogs, on social media, and through non peer reviewed papers … because this disseminates information in a more responsive and relevant way”. His question is “how is the validity of this work assessed?”

Rigour is not given up. Responsiveness and relevance are added to it.

Experimenting in public, openly. He publishes whole AI transcripts, with a share link for the tenure one. He says where the “interesting stuff” is, and marks the output’s limits: “not much that’s new here”, missing “equitable communities and a collaborative culture that includes staff, students, and non-tenured faculty” (2023-03-29). He discloses the Midjourney prompts behind his images. Readers see the process, not only the result.

How he treats evidence and expertise. - Careful about causation in the Belgian case (2023-04-05). - Notes provenance: a psychiatry piece is “not a peer review publication” but still signals interest (2023-04-05). - Treats AI output as a starting point, filtered through expertise: “a useful starting point for what I hope will be much deeper conversations” (2023-04-05). - Admits the irony of turning “to the very source of the problem” (2023-04-05). - Values expertise, but broadly defined. In the governance posts it includes the arts, the humanities, “areas of understanding that don’t have neat disciplinary labels” (2023-04-12) and “creative and disciplinarily unbounded thinkers” (2023-04-04).

Transdisciplinarity as a working requirement, not a slogan. - The world congress should include “smart and collaborative thinkers from the humanities, arts and social sciences — diversity of thought and perspective will be essential here”, chosen for the ability to work “without ego getting in the way” (2023-04-04). - CIFAR’s Future Flourishing programme, with its philosophers, curators, conservators, artists and anthropologists, is held up as the kind of scale needed (2023-04-07).

Accessibility as craft. He writes with informal markers (“like, yesterday!”, “(I may have said that already)”, “sorry something else”). He uses genre-fluid labels (“blog/newsletter/column (I’m not sure what you’d call this)”; the subtitle “newsletter/blog/column/podcast (possibly)”). He describes his engagement as iterative learning: “constantly learning from what is effective in engaging with people and empowering them though [sic] making understanding accessible” (2023-01-31 prompt).

5. His role as he sees it#

A navigator drawing on experience, who stays humble. He speaks from “over 20 years” of work on the risks and benefits of advanced technologies (2023-04-04, 2023-04-12). He uses that experience to limit his claims (“the only thing I’m certain about is that there are no silver bullets”), not to pronounce. His proposals come with qualifiers: “one possible way forward (but by no means the only one)” (2023-04-04), and “maybe” (2023-04-12).

A connector, and a builder of coherence. He describes his method as “making connections across very different disciplines to gain insights, and working with people from very different backgrounds”, and as talking with “stakeholders and members of the public to better understand challenges and opportunities, as well as aspirations” (2023-01-31 prompts).

He sees a gap in thought leadership that is coherent, visible and accessible, one that can “join the dots in coherent and informed ways” (2023-04-12). Advanced Technology Transitions, the Future of Being Human initiative and the Substack are all ways of filling it.

With readers. He is invitational and companionable: “I’d love for you to subscribe and follow along” (2023-04-04 welcome). He is candid about procrastination, about shock, and about not knowing how his reviewers will react (“I guess I’ll find out sooner or later!”, 2023-01-31). He also shows care: “please do seek professional advice — and not just from ChatGPT” (2023-04-05).

With students. In his own words (2023-01-31 prompts), teaching aims at “learning and empowerment that is responsive to student needs, potential, and future opportunities”. He uses “unconventional and innovative approaches”: The Moviegoer’s Guide to the Future, and an Antarctica study-abroad course.

With policymakers. Constructive and critical at once. He asks whether they are “sufficiently creative”, supplies a reading list of governance tools, and ends in hope (“Hopefully though this is just a temporary glitch”, 2023-04-10).

With industry. He isn’t hostile. He wants “public private partnerships” (2023-04-04), is sceptical of the “Silicon Valley echo chamber”, and is sympathetic to “the developers within companies who are asking questions about the broader implications of their work” (2023-04-10).

With other experts. Gary Marcus is a conversation partner. He engages LeCun’s rhetoric without insulting him. He declines the pause letter respectfully. The AI-ethics community is “well-intentioned” but under-informed. He discloses his own position openly: “full disclosure, I was there at the 2017 Asilomar meeting” (2023-04-04).

His pattern: charitable about motives, sharp about gaps in knowledge. Advocates are “well-meaning” but “blissfully unaware” and have “a very rudimentary understanding” (2023-04-12).

What he refuses to do: - Join a camp. He neither signs the pause letter nor dismisses the fears (2023-04-04). He calls loud voices “an important part of the policy landscape” while declining to trust “people who have an agenda, who can mobilize fast, and who have the ear of decision makers, to get things right” (2023-04-10). - Sensationalise a death (2023-04-05). - Present a single remedy. - Accept a “talking shop” (2023-04-04). - Puff himself up. He asks ChatGPT to make sure he doesn’t “come across as arrogant and self-centered”, then, self-mockingly, asks it to “pull back a little on the but [sic] about humility” (2023-01-31).

Changes of mind and of practice. - No reversal of substance appears in this batch. - Concern deepens with experience. January’s “intrigues me and slightly worries me” about his own attachment to ChatGPT becomes April’s “This worries me” about other people’s. - A change of practice: he admits he drifted from informal blogging toward books and articles and “missed the informality and responsiveness of a blog”, and that his AI-risk thinking had not been written down in citable form (2023-04-04 welcome).

6. What is distinctive#

7. Posts that best reveal how he thinks#

  1. 2023-04-12 navigating-advanced-technology-transitions — a new mindset for unprecedented transitions (“unconventional approaches to unconventional challenges”); legibility; transdisciplinarity including unlabelled knowledge; public writing openly upstream of formal scholarship.
  2. 2023-04-04 what-are-the-alternatives-to-calling — holding the tension (existential risk, no pause); reframing from ethics to risk and responsible innovation; structural history of governance; “risk landscape … navigate”; “no silver bullets”; rough thinking posted because speed matters.
  3. 2023-04-05 can-chatgpt-adversely-impact-mental — self-experiment, mechanism-based plausibility reasoning, caution about causation, focus on the vulnerable, a neglected risk, care for readers.
  4. 2023-01-31 can-chatgpt-take-the-pain-out-of-annual-academic-reviews — play as inquiry; reflexive self-observation; his own unguarded account of integrated scholarship and accessibility (in the prompts).
  5. 2023-04-10 as-ai-goes-to-washington-whats-being — loud and quiet voices, the insights vacuum, creativity demanded of governance, regulation as one tool within a broader landscape.
  6. 2023-04-04 welcome-to-the-future-of-being-human (shorter, but revealing about role) — the personal-scale framing, agency, informality, readers as companions.