Late Lessons, Jensen Huang and AI
Markdown version

B10 perspective notes: 2023-07-25 to 2023-09-01 (18 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. All eighteen posts were read in full, including the six Moviegoer’s Guide episode posts (their written introductions 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

Context. ChatGPT is eight months old. He has just finished marking over 2,000 student–ChatGPT conversations from his first prompt-engineering course, and a new semester is two weeks away. At the same time he is reading Films from the Future aloud, a chapter a week, for The Moviegoer’s Guide to the Future podcast. The two activities feed each other. Rereading his 2018 chapters (Ghost in the Shell, Ex Machina, Transcendence) keeps sending him back into 2023 questions about personhood, consciousness, manipulation and where ideas come from. Oppenheimer has just opened. So the batch shows him in two modes at once: reporting straight from the classroom, and re-examining his own five-year-old thinking in public.

1. How he thinks here#

He starts from something that happened to him, often with a confession#

Almost every substantial post opens with a personal encounter:

The confession does real work. It places him beside the reader as someone still finding out, not as someone handing down a verdict. It also often marks the point where he changed his mind. He did not want to like Oppenheimer and was won over. He “struggled” with Ghost in the Shell and says “I now have a deep respect” for it (2023-08-18 being-human-in-an-augmented-future).

He lets serendipity steer, and says so#

He puts himself in the dock as the test case#

In the provenance post he turns the charge usually made against LLMs on himself: “Did I do the equivalent of ChatGPT and regurgitate other people’s work as my own?” He had even checked with Mark Daley “just in case I’d inadvertently coopted something that didn’t belong to me.” He then uses his own case to show that human idea formation “is something of a black box — even to me.”

The move is to test a criticism of the machine against the human baseline, starting with himself, before judging the machine from outside. It is analogy by structure, applied to himself.

Analogy by structure, with the break points marked#

He reframes by turning the obvious problem around#

Plausibility tests come before engagement#

Films as scenario engines#

He asks rather than answers#

In 2023-08-02 fifteen-questions-about-generativeai the questions are the intervention: “I’m going to be awkward though and not provide any — not because I don’t have them, but because these really should be the responsibility of instructors and their schools and colleges to research, discuss, and agree on.”

He holds past and present together#

His curiosity lands on our frames, not only on the technology#

In the Limitless introduction (2023-08-04) he is intrigued “not because I’m enthusiastic about using such a technology myself, but because the social norms and expectations around chemicals-based augmentation seem so different to physical augmentation”. Then: “What intrigues me even more though is how an obsession with improving “intelligence” often reflects our biases and preconceptions around what intelligence actually is.” He says this matters “just as” much for artificial intelligence. The chapter behind it argues that “how we think about intelligence is remarkably colored by our sense of our own importance” (FFTF p.104).

2. What matters to him#

Access and fairness, especially for people who don’t fit the standard mould#

Anger appears rarely in this batch, and when it does, it is about injustice.

Minds switched on: learning how to think#

Dignity and personhood, extended outward#

His care for whoever is most exposed runs from students without coaching, through workers carrying risk, to possible machine minds. And he keeps asking what our treatment of them says about “how this in turn reflects and defines our own humanity” (2023-08-18).

Nuance over polarity#

Can versus should#

Taking students’ effort seriously#

“I’m invariably disappointed when my students make the effort to do stuff like this and it lands on deaf ears” (2023-08-28).

Honesty about where ideas come from#

Attribution is “part of the bedrock of academic scholarship, intellectual property, and ownership over ideas — and something that my academic writing reflects extensively.” (2023-08-21) He also corrects himself in public: “On a re-listen I realize that I verbally slipped and stated Neuralink was launched in 2007” (2023-08-18).

What frustrates him#

What delights him#

3. Risk as a way of thinking#

This batch uses little explicit risk vocabulary. His risk thinking shows mostly in how he redefines what counts as harm and what counts as value.

Explicit: “thinking differently” about novel risk#

2023-08-08 thinking-differently-about-ai-and-risk. The title is itself the claim. “At the heart of the talk is a discussion of how to frame novel risks associated with AI, and how the risk innovation framework we have developed here at ASU could be applied to good effect.” The key verb is frame: risk innovation is offered as a way of seeing risks that don’t fit existing categories. He calls it “perhaps more relevant than ever in an age of generative AI”. The slide file is titled “AI-Human-Condition”, which ties AI risk to the human condition. The talk itself was not examined.

A risk landscape full of people, which you navigate#

Threat to value, in practice if not by name#

Orphan-like risks (my interpretation; he does not use the term here)#

New technology needs a new mindset#

How these ideas function#

They work as mental models that open up possibilities. In the education posts, reframing the risk opens an agenda (AI literacy, equity of access, course redesign) that a hazard frame would close down with bans and detectors. He lists detectors only as a question (Q8). The one place anything is framed as a tool is risk innovation “applied to good effect”, and even there the emphasis is on how to frame novel risk. (Allenby’s complex-systems argument appears in the batch but is a guest post and is set aside.)

4. Scholarship and public writing#

The classroom as a research site, reported with caveats#

Translating research, then taking his own step#

In 2023-08-23 he explains the preprint’s three assumptions, its 14 indicators, and the risks of over- and under-attribution in plain language. Then he adds his own steps: the tension between “economic expediency” and moral responsibility; a broader idea of suffering; the irony that the rubric “could also be used as a roadmap for achieving this”; and the call for “more than just research”.

Re-testing his own work in public#

He reads FFTF aloud, reposts the 2021 talk, shares version 7 of the mind map, updates the book’s web resources, and uses the book as his course text. Public writing is where he tests earlier scholarship against the present. Revisiting Ghost in the Shell, “I can see very clearly the seeds of how my own thinking had developed” (2023-08-18).

The provenance post gives his own account of how the parts connect: ““artifacts” such as Films from the Future, my more recent book Future Rising, the ASU Future of Being Human initiative, this Substack, and even the CIFAR call for proposals, are merely the visible fruits of a messy and largely hidden network of influences”. In his own image, scholarship and public work grow from one mycelium.

Experimenting in public#

Transdisciplinarity as the default#

Accessibility and disclosure#

Evidence and expertise#

He respects expertise without deferring to it. The consciousness authors are “highly respected thinkers in the field”, but he tests their key assumption and goes past their conclusion. On teaching policy, he places the expertise and the responsibility with instructors and their institutions, not with himself (2023-08-02).

5. His role as he sees it#

Companion and convener, not preacher#

A teacher who treats students as agents#

In 2023-08-28, the class “encourages participants to develop their own ideas around what responsible innovation means”. Students’ views are quoted, including “have they not seen any movies?” The plan was collective (“And so we came up with a plan.”), and 54 students signed the letter. Two years later he stands up for them in public.

A good-humoured, engaged critic of industry#

An insider pushing his own institutions#

He argues for universities to adopt Enterprise ChatGPT, “at least, where they have the vision to adopt it”. He welcomes ASU Law’s decision, and says “every high school in the country should be ensuring” AI literacy.

How he treats other experts#

What he refuses to do#

Changes of mind, made visible#

Tensions a careful reader might note#

His own coda (2023-08-14) shows he knows his enthusiasm needs checking. That is part of the method.

6. What is distinctive#

  1. Flaws as features for learning (2023-08-14). In mid-2023, most commentary judged LLMs on accuracy. He judged them on what they do to a learner’s thinking, used 2,000 conversations as evidence, and argued that the flaws force critical engagement. This is a risk–benefit reframe grounded in what is valued, not a count of errors.
  2. A reflexive analogy on provenance (2023-08-21). Instead of arguing about copyright from the outside, he audits his own mind as the comparison case and finds “a black box”. He proposes the “ideas mycelium”, which moves the governance question from traceable attribution to the “shape, structure, and composition” of what goes in. He still holds that ownership and consent “are important”. Both the method (himself as data) and the result (govern the composition of the inputs) are unusual.
  3. The nuclear analogy used to show difference (2023-07-25). Against the Oppenheimer-moment rhetoric, he argues that the human power dynamics transfer and the containment does not. AI’s risks are dispersed, accessible and driven by “shadowy actors and naive developers”, and he keeps AI and nuclear weapons in proportion.
  4. Moral risk that runs toward the machine (2023-08-18, 2023-08-23). He puts AI welfare inside the risk landscape. He ties it to a long-standing concern about dehumanising whatever is different (FFTF p.139), and names “economic expediency” as the pressure to deny it.
  5. Equity as the lens on AI in education (2023-07-27, 2023-08-29). While institutions reached for bans and detectors, he asked who a ban protects. He read a product launch in terms of who can afford $20 a month.
  6. Play as a serious instrument. He reads his book aloud, keeps “playing with” a mind map, uses films as scenario engines, and sent Tesla a box of 27 books with a signed Banks novel hidden inside. The play carries the can-versus-should question, and it treats developers as potential allies rather than opponents.
  7. Long memory. He tests 2018 and 2021 work against 2023: “everything about artificial intelligence has changed over the past 12 months. In others, little has changed.” That lets him see continuity in a field that tends to treat each month as unprecedented.

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

  1. 2023-08-21 the-messiness-of-the-provenance-of-ideas. Serendipity, himself as the test case, discipline about metaphor, and scholarship and public work described as one mycelium.
  2. 2023-08-14 chatgpt-stimulates-creativity-critical-thinking. The classroom as evidence, flaws reframed as features, and enthusiasm checked against counter-evidence.
  3. 2023-08-28 dear-mr-musk-can-i-have-my-book-back. Play, students as agents, can versus should, and good-humoured engagement with industry.
  4. 2023-08-23 could-we-build-conscious-ais-in-the-future. A plausibility test, a visible change of mind, and harm and value extended to non-human minds.
  5. 2023-07-25 oppenheimer-and-ai. Film as a scenario engine, a structural analogy with its break points marked, a risk landscape full of people, and the rejection of bumper-sticker positions.
  6. 2023-08-02 fifteen-questions-about-generativeai. Refusing to answer, questions as the intervention, and responsibility spread across institutions.