Late Lessons, Jensen Huang and AI
Markdown version

B24 perspective notes: 2025-04-06 to 2025-04-08 (2 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. I read both posts in full. For the o1-pro report embedded in the first post, I read the Executive Summary and main body and skimmed Annexes A and B, checking for any passages in his own voice. There were none. Quotes are exact, including his typos (“growing number of leaders”, “it’s length”) and curly punctuation. Markdown italics are dropped.

Evidence rules applied

Context. Early April 2025. AI 2027 (Kokotajlo et al.) has just come out. He is “heading into a workshop on AI and responsible innovation”. GPT-4o’s image generator has set off the “Studio Ghibli style” controversy.

1. How he thinks here#

He opens with his own honest first reaction, then steps back from it on purpose#

The responsible-innovation post (2025-04-06) is a record of a thought process in sequence. It is not a finished position.

  1. What was on his mind. “The scenario was published just as I was heading into a workshop on AI and responsible innovation this past week, and so the question of how we ensure artificial intelligence is developed and managed appropriately was on my mind.” The question reaches him through coincidence and timing, not through a research agenda.
  2. The gut reaction, stated plainly. “my first reaction on reading AI 2027 was to worry that, even if the projections represent an edge case, we might be facing a near term future where current efforts to develop artificial intelligence responsibly seem futile.”
  3. A hope, then both sides of the reception. “I hope they are not — and the scenario has already attracted considerable pushback for being too alarmist. Yet it’s also been cautiously welcomed by some big names in cutting edge AI as a salutary warning”.
  4. A deliberate cooling-off. “The first step I suspect is to take a deep breath and move back from speculation to firmer ground. AI 2027 is speculation — no more.”
  5. Taking it seriously anyway. “And yet it does force the question of how we might think about responsible innovation and AI, just on the off chance that there’s a sliver of truth here.”

This is not the choice between alarm and dismissal that most responses to AI 2027 made. He shows the reader that he felt worried, then works out loud to a calibrated position: speculative, yes, but worth thinking through in case it is partly right. The emotional response is kept and named, not hidden, and then it is disciplined.

He lays out the assumptions before arguing about the conclusion#

He does not argue with the scenario’s ending. He lists its four premises instead: coding-capable internal models, self-improving AI that turns developers into “AI managers”, hardware and energy keeping pace, and game-theoretic race behaviour (“an AI arms race is inevitable, no matter how bad an idea anyone thinks it is”). Then he rules on them in two sentences: “Each one of these has its flaws. Yet they are not unreasonable as a starting point for imagining edge case scenarios.” He calls the assumptions “all of which can be contested, but nevertheless are useful for exploring potential (if not necessarily likely) near term AI futures.”

So a scenario, for him, is a tool for exploring what is possible, not a prediction to be believed or refuted. He keeps “potential” and “likely” apart in so many words. He also pays attention to form. It is “(or, to be more accurate, scenarios, as this is something of a “choose your own ending” story)”. This matters because he reads the document as a branching exploration rather than a single forecast.

He moves the problem from institutions to human cognition#

The post’s key move comes in two steps. First comes the expected institutional point: RI processes “can take years”, while “a lag of even a month in the development cycle might mean the difference between abject failure and world domination.” Then he turns: “What worries me just as much though is that nothing about how we think, how we plan for the future, or how we develop approaches to ensuring better futures, is geared toward exponential advances that happen over months rather than years.”

This recasts a governance-speed problem as a problem of mindset and perception. The consequence he draws concerns seeing, not regulating: “we would most likely fail to recognize it — or would actively deny it — until it was too late. And all because we are really bad at wrapping our heads around rapid exponential growth.”

He uses a flawed film to carry a physicist’s thought experiment#

He reaches for Inferno, the subject of FFTF chapter 11 (see 2018-10-18 the-honest-broker-meets-dan-browns-inferno-ed637700b633). He says frankly that it is “a deeply flawed but nevertheless compelling illustration”. Through the film he brings in Al Bartlett’s 1978 beaker puzzle. The answer, “11:59 PM. One minute to midnight”, is a small shock of the kind a physicist would recognise, and it lands because it is counter-intuitive.

He also flags the model’s limits at once: “The illustration would never work in real life as resource constraints would slow or halt the exponential growth. But it is a good illustration of how hard it is for us as individuals or as a society to plan for exponential growth”. The analogy works by structure, as a picture of perception and planning lag. It is not a claim that AI will actually grow exponentially. He uses a known-imperfect model for what it shows, and says so.

He then turns the thought experiment into two “what if” questions: - “what happens when AI development is on an exponential growth path and we simply cannot accept or even see this until it’s metaphorically one minute to midnight?” - “what happens if we’re still planning for the world as it was at 11:00 PM when we get to the AI equivalent of 11:59 PM?”

He anticipates the reader’s dismissal and folds it into the argument#

“I suspect that, to many, this will feel like an intellectual exercise and no more. But this is precisely the point of the illustration — it always will feel like an intellectual exercise until it’s too late.” The expected reaction becomes evidence for his point. This is a reflexive move: the reader is asked to notice their own thinking as part of the phenomenon.

Building to find out: a deep dive with a reasoning model#

His response to his own question is not to write a position paper. It is to run an experiment in a working method: “a mode of working that I’ve been finding increasingly useful recently — engaging with OpenAI’s o1-pro model to develop nuanced and widely informed insights into complex questions.” He describes his part as active: “part of my process is actively engaging with o1-pro in the research and writing process, and evaluating and editing the final report where necessary”. He publishes the result in full, with caveats.

For context (outside the batch): two months earlier he had described o1-pro as “much faster than what I could have achieved my own (by a matter of months)”, while sensing “something lacking in the depth of simulated thinking and scholarship” (2025-02-04 openai-deep-research-ai-scholarship). So B24 continues that experiment in public. There is also a reflexive irony, which is my reading rather than something he states: he uses an accelerating AI tool to think about the problem of AI acceleration, and so produces 40 pages in a timeframe that proves the point about speed.

Play as a way of engaging, not just decoration (2025-04-08)#

In the Studio Ghibli post he does the thing under discussion. The header image caption reads: “Generated, appropriately, using OpenAI ChatGPT 4o — not using Studio Ghibli style!” He writes: “I couldn’t resist abandoning my usual use of Midjourney version 3 to use ChatGPT 4o to generate the image for this post!” He had the tool choose its own style from “a transcript of the podcast and a photo of Sean and myself”. Then he notices what came back: “I have a sneaky idea that ChatGPT didn’t appreciate the irony of the caption, given the image rendering 😊”. The video’s thumbnail is openly “generated using Studio Ghibli style”.

He explores the controversy partly by playing with the technology at its centre, in public, and he enjoys the irony. Here play is a way of knowing: you learn what the tool does with “style” by asking it to pick one.

2. What matters to him#

3. Risk as a way of thinking#

None of the named vocabulary appears in this batch: no risk innovation, risk landscape, threat to value, orphan risks, or “navigating”. His own prose says “developed and managed appropriately” and “govern AI”. Even so, the underlying stance is very close to his September 2026 self-account.

4. Scholarship and public writing#

5. His role as he sees it#

6. What is distinctive#

  1. He makes the AI-acceleration debate a question about human cognition. Most responses to AI 2027 argued about whether its timelines are plausible or which institutions are needed. His worry “just as much” is that our ways of thinking and planning cannot perceive change on this timescale, and that denial is the default. The question moves from “will it happen?” to “would we even see it?”
  2. He is willing to ask whether his own field’s tools are futile, calmly and in public. He is a responsible-innovation insider who puts RI’s fitness for purpose on the table without getting defensive and without catastrophising.
  3. Physics-style thought experiments delivered through popular film, with the flaws named. A “deeply flawed” Hollywood film is a legitimate route to a precise, counter-intuitive insight. The imperfection of the model is part of the lesson. This comes straight from FFTF, and he still reaches for it seven years later.
  4. Reasoning about possibility without false precision. Taking an edge case seriously “on the off chance”, while refusing to assign likelihood, is neither the probabilistic stance of the AI-safety world nor the dismissal of its critics.
  5. Enacting the subject. He uses a fast AI reasoning model to research AI acceleration, and ChatGPT 4o’s image generator to illustrate a post about its own style controversy. He tests the technology on itself and shows readers the result. He engages with AI by using it, reflexively and openly, not only by commenting on it.
  6. Readers get several levels of depth. Perplexity and ChatGPT summaries, the full report and the PDF are offered side by side. This treats the post as an experiment in how people read complex material in an AI age, not just as a text.

7. The posts that best reveal how he thinks#

This batch has only two posts, so fewer than three can be listed.

  1. 2025-04-06 responsible-innovation-and-ai-acceleration. This is the key post, although only about 7% of it is his own writing. That portion shows the full sequence: coincidence prompts the question, he admits his first reaction, he surfaces the assumptions, he reframes the problem as one of cognition, uses a film-borne thought experiment with its limits flagged, folds in the reader’s likely dismissal, cools off deliberately, experiments with a new method, publishes openly and curates for readers. His substantive views on what to do are not here. They are delegated to the o1-pro report, which he endorses only in general terms.
  2. 2025-04-08 openai-and-studio-ghibli-style. Minor, but it shows the play clearly: using the contested tool on itself, enjoying the irony, framing the issue as a question (“Theft or homage?”) that is “far from black and white”, and respecting the “hundreds of hours of skilled work” behind a human style.