B12 perspective notes: 2023-10-24 to 2023-12-01 (17 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. All seventeen posts were read in full. Quotes are exact, including his typos. There are no Modem Futura posts in this batch.

Evidence rules applied

2023-10-29 category-confusion-complicates-efforts and 2023-11-06 an-ipcc-for-ai-is-a-failure-mode are guest essays by his ASU colleague Brad Allenby. They are not evidence of Maynard’s thinking. What they do show is that he uses the Substack to host colleagues’ arguments. He links to Allenby’s piece once in his own words (2023-10-30 white-house-goes-all-in-on-responsible-ai), which counts as endorsing that one point.
2023-10-30 time-specific-actions-on-ai. The list was generated by Claude. Only his framing lines count, including the open disclosure of AI help and the instruction to cross-check.
2023-11-21 ai-and-risk-innovation. Most of the body is ChatGPT role-play output. Only his framing, design notes and reflections count. The Risk Innovation Planner itself, and its design brief, are his (and the Nexus team’s).
2023-11-11 a-fist-look-at-openais-gpts. The GPT conversations are not evidence. The instructions he wrote for the GPTs, and his commentary on building them, are.
2023-10-25 10-million-for-ai-safety-research and 2023-10-30 white-house-goes-all-in-on-responsible-ai each quote a line from his Nature Nanotechnology paper co-authored with Sean Dudley. These lines are treated as his published position because he quotes them as his own.
Event notices (2023-10-27 ai-and-consciousness, 2023-11-08 human-ai-symbiosis, 2023-12-01 dec-19-future-of-being-human-unplugged) are short. Only his framing lines are used. Speaker bios are ignored.

Context. These five weeks include: - the Biden AI Executive Order (30 Oct); - the Bletchley Park AI Safety Summit (1–2 Nov); - OpenAI DevDay and custom GPTs; - the Cruise robotaxi collapse; - Sam Altman’s firing and return (17–22 Nov); - ChatGPT’s first anniversary.

Maynard responds in several ways: - He builds things: five GPTs, and a role-play run of his own risk tool. - He checks things: he redoes the arithmetic on a Waymo safety study. - He brings back older work: the 2018 orphan-risks article. - He goes back to first principles on what “risk” means.

Across the batch, the answer to the news is rarely an opinion. It is usually an experiment, a reframing or a tool.

1. How he thinks here#

He opens by questioning his own ideas#

2023-10-24 flattening-the-learning-distribution-curve begins with doubt about a phrase he has been using on stage: “I sometimes come away from these events thinking to myself “is this really a thing?”” The post is him checking, in public, whether an intuition holds. He finds that it does: “It turns out that the concept does make sense”. Along the way he sketches a stylised diagram, and flags it as “admittedly, a somewhat stylized and over-simplified illustration.”

2023-11-26 everything-youve-heard-about-ai-risk-is-wrong runs the same move at a larger scale: - He leads with a provocation in the title. - He adapts a meme about overnight AI experts into one about overnight AI-risk experts, calling the original “snarkily accurate”. - He ends by walking the title back: “Maybe I was a little hubristic with my statement … there’s probably some stuff there that’s less wrong. Although the more I learn, the less sure I am.”

The provocation opens a door, and the self-correction shows the humility he is asking of others.

First principles, from a risk scientist’s foundations#

Faced with a year of AI-risk noise, his move in 2023-11-26 is to step back: “it’s worth going back to first principles and what we mean by risk in the first place.” He takes the textbook definition, “the probability of harm occurring from an action, process, or situation”, and unpacks it into five parts:

Cause and effect: “no cause, no risk”. Hazard is not risk: “bleach is hazardous, so is a piano”.
Magnitude: some sense of scale is needed, whether “quantitative, qualitative, or comparative”.
Harm: extended beyond health, wealth and the environment to “mental wellbeing, happiness, sense of worth, identity, beliefs, aspirations”.
Time: long timescales “mess up the lines between cause and effect”.
Perception: “an understanding of risk cannot be separated from an understanding of human behavior and societal dynamics.”

The foundations are conventional risk science. What he does with them is not. Each component, followed through for AI, shows how little we know. He uses first principles to open the question up, not to settle it.

The next day he adds an addendum because the classic formulation “risk = hazard x exposure” had been “bugging me”. He explains why he had left it out: - to build “a broader understanding of risk that extends beyond the hazard-exposure paradigm”; - because dose–response can be non-linear (“threshold responses, hormesis, and other low-dose responses”); - because “I also didn’t want to fall into the trap of implying that zero exposure — as in no AI — is a default risk management strategy.”

He still gives the old paradigm a role, as something that might “reveal new ways of thinking about and understanding risk”. This is quantitative risk science built on, not thrown away, and used as a mental model rather than a calculation.

Building and experimenting to find out#

2023-11-11 a-fist-look-at-openais-gpts. Within days of the launch he has built five GPTs: - Imagination Catalyst; - two personal learning guides, one on responsible innovation and one on advanced technology transitions; - two “book guides” for Films from the Future and Future Rising.

He records what he learned the way a lab notebook would: - tricks with conversation starters (an ellipsis at the end, and a “Tell me about yourself” starter); - a workaround for GPTs answering in character; - a playful probe test, “talk to me about peanuts”, chosen because “neither book explicitly mentions peanuts” (“idiosyncratic I know”); - the rules he wrote to keep a book GPT on topic.

His verdict is balanced. GPTs are “often not as impressive as they first seem”, and “good GPTs require a lot of tweaking and experimentation”. Yet the capability is “a game changer”.

2023-11-21 ai-and-risk-innovation. Three days after Altman’s firing he does not add another opinion. He runs a counterfactual game: “just to make things more interesting, I thought I’d experiment with using ChatGPT to explore how the planner might have been used.” - He casts ChatGPT as the board of “a hypothetical company” and walks it through his Risk Innovation Planner. - He publishes the opening prompt “typos and all”, along with the source documents and the completed planner. - His own reaction: “more illuminating than I expected”. What surprised him was “just how well ChatGPT did”. - He says what the exercise cannot show: whether it would have helped OpenAI “is, of course, impossible to tell.”

Play (role-play), experiment (a counterfactual), a tool (the Planner) and full transparency are all combined in one post.

Checking the evidence himself#

2023-11-09 waymo-safety-study-shows-benefits shows his quantitative instincts at work: - Plausibility test on the source. Swiss Re’s business depends on “cold, hard analysis of risk and liability”, so “I cannot imagine a scenario where Swiss Re sacrifices the very core of its business by putting out a report based on spin”. - Separating anecdote from data. His own rides were positive, but “anecdote is no substance for data.” - Checking the method. He spots a possible bias: the human baseline included interstates, where Waymo does not drive. He was “Not convinced” by the authors’ reassurance, so he ran “back-of-the-envelope calculations” on NHTSA and FHWA data. The check showed the baseline is probably conservative, so the benefit is likely understated. - Keeping the bigger question open. “there remains the serious question of how safe self-driving cars should be, and whether comparing them to humans is the right metric.”

Reading technology through stories#

2023-11-05 dont-panic-elon-musk-launches-grok reads a product through the fiction it borrows from: - He quotes the radio Hitch-Hiker’s Guide on the Guide being “occasionally accurate”, and dryly notes the fit with “a guide that leans more toward what you want to hear than what is, strictly speaking, true”. - He does not stop at the joke. Adams is “remarkably good at holding up a mirror to humanity’s absurdities, and in doing so helping us cut through the noise of entrenched ideas and narrow thinking.” - He traces “grok” back to Heinlein (“to understand something profoundly and intuitively”). - He ends open: “I guess we’re about to find out!”

The story here is a lens that can cut both ways, not just a weapon for mockery.

Framing through intersecting lenses and holding outcomes open#

2023-11-18 sam-altman-openai-impacts. He sets out “three frames, and the deep intersections between them”: the technology, its governance, and society. He sets aside the technical frame: “I’ll leave it to those who are more deeply involved in the technical foundations”. He keeps both outcomes open. The shake-up might be “a reset from the hype … a chance to take a breath (a pause even)”, or it “could put many plans into a tailspin”. He then updates twice as the story moves: “The story isn’t over yet!”
2023-11-29 the-year-that-generative-ai-changed-the-world. He moves the cause away from the technology: “it wasn’t so much the technology as the social and technological landscape surrounding it”. He gives a net judgement (“on balance, we’re collectively in a better place”) and a forecast (“I’d probably put my money on a cooling off of the hype”). Both are hedged by complex systems’ “nasty habit of throwing up the unexpected.”

Reframing a problem to find a better option#

In 2023-10-24 flattening-the-learning-distribution-curve he looks at three responses to students at the edges of the curve: - Narrow it, by selective admission. He rejects this because it works “by excluding students at the edges”. - Broaden it, through better teaching. This “only goes so far”. - Flatten it, by raising success at the edges.

AI is then placed as a possible way to flatten the curve, not as a threat to be contained. Even its flaws are reframed as useful for learning. In a Socratic-dialogue assignment, ChatGPT’s unreliable accuracy means “each student is challenged to question and test what they are exposed to as part of a dialogue that’s unique to them.”

Setting fields side by side and letting serendipity work#

The event notices show the same design habit: - AI with cooperation science and evolutionary biology (2023-11-08 human-ai-symbiosis): “What’s really interesting me though is how bringing together expertise on cooperation between biological organisms with deep understanding of evolutionary biology and behavior will reveal new insights”. - Conversations billed as “lively, serendipitous, unscripted”. - Events expected to “go places few other conversations around AI have gone” (2023-12-01 dec-19-future-of-being-human-unplugged).

2. What matters to him#

Inclusion and the people at the edges#

“This is not an approach that sits well with me as an educator” (2023-10-24), about selecting students so that success looks high. He anchors this in ASU’s charter, measuring success “by whom we include, not by whom we exclude”, and in “the ethos of learning and education at a public university”. The same value runs through his civic writing: - AI should serve “a better future for as many as possible” (2023-11-18). - Social-justice experts must be “partners in developing responsible and beneficial AI, and not just add-ons” (2023-10-30 white-house-goes-all-in-on-responsible-ai).

Technology that serves society, benefits included#

He counts the costs of the status quo. Waymo matters to him because he “worries about how we’ve normalized loss of life, limb, and property resulting from inattentive and poorly trained humans” behind the wheel (2023-11-09). His overall verdict on ChatGPT’s first year is that we are “in a better place” (2023-11-29). His governance commentary starts from “to realize the benefits of AI, we have to understand how to navigate the risks” (2023-10-30).

What makes a human life meaningful#

His idea of harm reaches into the inner life. It includes “our sense of identity, our beliefs, our worldview, relationships, and many other things that give our lives meaning”. Risk is also about the future: “not just about protecting what we have, but protecting pathways to futures we desire” (2023-11-26).

Opening capability to more people#

He is delighted that GPT building moves creation away from “something that computer scientists do to something that anyone with a good grasp of language and how to communicate can do” (2023-11-11). He is also delighted that a book can be explored “without reading it in a linear fashion”.

Imagination matched to the scale of the technology#

His hope for the Frontier Model Forum’s $10m fund is ideas “as radically transformative as the technologies they are hoping to steer”. His fear is “yet another attempt to be radically transformative in very conventional ways” (2023-10-25). The list of what he would like funded is itself a statement of values: - “novel conceptualizations of risk”; - “inspiration from the nexus between the arts, humanities and emerging capabilities”; - ways to “radically reimagine public participation and engagement in responsible innovation”.

What frustrates him#

Conventional thinking applied to unconventional problems. “unconventional challenges through rather conventional thinking”, and “outmoded models of risk management” (2023-10-25).
Overconfidence and ego. Commentary that fills the “understanding-vacuum” with “dogmatic overconfidence” (2023-11-26). “Half the world seem to have become AI experts overnight” (2023-11-29).
Big-tech capture of policy. “an over-emphasis on following the lead of large tech companies and their agendas”, and “vested interests wrapped in the clothing of social responsibility” (2023-10-30).
A reckless culture. “an ethos of go fast, break things, and don’t worry about the consequences” (2023-10-30).
Social naivety. Companies that are “technologically creative yet socially naive” (2023-11-15).
“Suck it and see” at scale. Naive experimentation in a bounded, non-linear world risks “pushing us past hard-to-spot tipping points” (2023-11-29).
Small irritations, admitted. The GPT’s “overly jolly tone … irritates me”. Waiting for access was “annoying!” (2023-11-11).

What delights him#

“Watching this, I still can’t get my head around just how easy this is!” and “wild!” on an image made from a conversation (2023-11-11).
“jaw-dropping moments (quite literally)” when showing ChatGPT to students (2023-11-29).
Upcoming conversations that “should be a doozy!” (2023-11-08) and are “going to be a lot of fun” (2023-12-01).
Robotaxis “driving themselves as if possessed by a hesitant 80 year old” (2023-11-09). The humour is affectionate, not dismissive.

3. Risk as a way of thinking#

This is the batch where risk innovation is first brought to bear directly on AI companies. The concepts work mainly as ways of seeing and of opening options, not as procedures.

Novel technology needs a new mindset#

“current developments in artificial intelligence are posing challenges that we simply don’t have the theories, models, mindsets, approaches, or policies, to navigate with clarity” (2023-11-26).
After thirty years in the field: “the less certain I am that we even know how to formulate the problems we face around AI, never mind manage the risks that may emerge from it.” The wording matters. Management assumes a problem has been formulated, and here it has not been.
He ends with a call for “broader transdisciplinary efforts to rethink and reformulate what we mean by risk in the face of a technology that defies conventional thinking.”
He also frames this as the foundation of risk innovation: “in order to thrive in a world driven by innovation and the changes it brings, we need parallel innovation in how we think about and act on risk” (2023-11-15). He calls it a “deceptively simple idea”.

Navigating, not managing#

In his own prose, “navigate” is the working verb throughout: - “Navigating AI risks” is a section heading (2023-11-26); - “navigate a risk landscape that might otherwise be hidden” (2023-11-21); - “learn together how to navigate the coming AI technology transition” (2023-11-26).

“Manage” turns up mostly where he is describing a limit or an older paradigm. The navigation options he gives are open-ended: “remove risks, help identify ways to circumnavigate them, or strategically absorb them” (2023-11-21). “Course-correct” and “agility” (2023-11-26, 2023-11-29) complete the picture of steering through changing terrain.

The risk landscape#

The landscape is the ground “that lies between new ideas and their successful implementation” (2023-11-15). He uses the image to explain failure: the Galactica model (which he calls Google’s; it was Meta’s) was “a ground-breaking idea with substantial potential that was scuppered by a lack of awareness of a deeply complex risk landscape”. AI’s landscape “is increasingly looking unlike anything we’ve had to navigate together before” (2023-11-26).

Risk as threat to value#

Admittedly subjective, and useful because it opens things up. “Approaching risk as a threat to value is, admittedly, a somewhat subjective way of thinking about risk. Yet it has the advantage of opening up conversations around factors that are often highly relevant” (2023-11-15).
An extension, not a replacement. It “extends conventional thinking rather than replacing it — health, wealth and the environment all fit comfortably into the “value” bucket.”
Reciprocal. It covers “the reciprocal dangers of threatening what is important to others through what they do.” The risk to a company is partly shaped by what it threatens in other people.
“Value” is not “values”. “values are not the same as value — a nuance that is often missed, and even caused ChatGPT some issues” (2023-11-21). The frame is “agnostic to particular worldviews, ideologies, or ethics. But it does open up thought processes”.
Future value. Risk includes threats to value “that we may aspire to” (2023-11-26). The frame covers creation as well as protection.

Orphan risks#

He reframes Rumsfeld’s taxonomy by adding a category of risks that are ““known knowns” if you’re looking in the right place, but aren’t taken as seriously as they should be” (2023-11-15). The subtitle states the purpose: these are risks “that have a habit of getting in the way of good ideas”. Risk here is not only harm to be prevented. It is what stops beneficial ideas from succeeding.

The ethical case sits alongside the strategic one: “Quite apart from the ethical inappropriateness of ignoring actions that could present deeply impactful social risks to others”. Even here he is funny about conventional risk mantras: “don’t break the law (or at the least, don’t get caught)”.

Tools as catalysts for a mindset#

The Risk Innovation Planner is a tool, but he is clear that its job is to shift thinking: - Design brief: “if we had 30 minutes to sit down with the founder of a startup to help them develop a risk innovation mindset that provided them with a competitive advantage, how could we do this?” - What it does not do: “What the Planner does not do is provide answers to problems.” - What it is: “a catalyst”, meant to “reveal pathways forward for users than they might not see”. - Letting users adapt it. When users reframe the orphan risks in their own terms: “This can be frustrating, but as the aim of the planner is to facilitate enterprise-relevant thinking, it’s also appropriate.” - Why it is needed for AI: it helps with challenges “that previous experience in tech innovation and formal business and risk management training simply do not prepare people for.”

Humility against false precision#

“a deeply complex risk landscape that only the foolish would claim to understand with certainty” (2023-11-26).
He applies the same humility to governance: “We don’t even know whether actions designed to manage potential risks — including regulations — could themselves constitute risks.”
He declines to take a side in the speed debate. The territory is uncharted “irrespective of whether it’s framed in terms of the risks of going too fast, the risks of not going fast enough, or the risks of simply assuming there are no risks.”
Quantitative skill is still used where data exist: the Waymo reanalysis (2023-11-09) is conventional risk comparison done carefully. He then asks whether the comparison metric is the right one, which is a social question.

Complex systems, tipping points, irreversibility#

In 2023-11-29 he names ChatGPT’s launch, and probably emergent LLM behaviours, as tipping points that “led to changes that cannot be undone, nor forgotten”. He does not condemn trial and error in general (“there’s nothing wrong with technology development that’s based on trial and error”). He worries about it in “an increasingly complex and bounded system”. The answer he gives is “agility” and “resilience in how we react to the unexpected”, not prediction.

4. Scholarship and public writing#

The Substack as a channel for his research#

The batch threads his scholarship into news commentary: - the Nature Nanotechnology paper with Sean Dudley on advanced technology transitions (2023-10-25, 2023-10-30); - the Risk Innovation Nexus, its Planner, templates and training materials (2023-11-15, 2023-11-21); - the 2018 orphan-risks article, reposted “pretty much as it was published back then”; - a National Academies-sponsored presentation, a JMIR paper on brain–machine interfaces, and the Nature Nanotechnology risk-innovation commentary (footnotes, 2023-11-15).

Old work is brought back when events make it newly relevant. The Galactica tweet triggers the orphan-risks repost, and the OpenAI crisis triggers the Planner post.

Tools given away#

He publishes the Planner template, the anonymised documents behind the role-play, and the completed planner as downloads (2023-11-21). He makes public GPTs and shares conversation links for readers without paid accounts (2023-11-11). The scholarship is meant to be used, not only cited.

Experimenting in public, method included#

He shows his workings in several ways: - the unedited GPT-building video, sped up only “to compensate for my slow typing (and stop you snoozing off)”; - prompts reproduced “typos and all”; - failed iterations: book GPTs “forgetting the book they were trained on”; - an open admission: “I’m still playing with this GPT and I’m not convinced yet” (2023-11-11).

Evidence, AI help and disclosure#

He discloses AI help up front: a list made “courtesy of some help from Anthropic’s Claude” that “should be cross-checked” (2023-10-30 time-specific-actions-on-ai).
He discloses a possible conflict with humour and precision: “Waymo once sent me a pair of socks … they are nice socks … Cruise have never sent me socks.” He adds that “I have never worked for or been paid by Waymo” (2023-11-09).
He adds dated updates rather than silent edits (2023-11-18; 2023-11-26 addendum).

Transdisciplinarity as a demand#

“our AI future cannot be left solely to AI experts” (2023-11-29). He goes beyond the usual call for social-science input and asks for expertise in “intelligence (irrespective of whether it’s biological or non-biological), consciousness, neuroscience, philosophy, the nature of personhood”.
Universities should explore “engaging in deeply novel ways across disciplines (including with the arts and humanities)” (2023-10-30).
The events practise this by putting AI next to cooperation science and evolutionary biology.
Hosting Brad Allenby’s essays is part of the same approach: a colleague trained in law and environmental science who now holds chairs in engineering and ethics, given the platform.

Accessibility#

A tl;dr for readers short of time (2023-11-11).
Everyday examples (“bleach is hazardous, so is a piano”).
A remixed meme, and simple diagrams (2023-10-24).
A “quick cheat sheet” on the Executive Order, posted the same day it came out.
A “sorry!” to readers without ChatGPT Plus.
Free, open events.

He works quickly: the Executive Order analysis goes up on the day it is released, and the Planner experiment three days after the OpenAI crisis. The aim is to be useful while it still matters.

5. His role as he sees it#

Authority used to show uncertainty#

He states his standing plainly: “I’ve been working in the field of risk for over thirty years and have been at the forefront of grappling with novel risks from emerging technologies for the past twenty of these” (2023-11-26). He uses that standing to say how uncertain the ground is, not to claim answers. With the OpenAI board he makes a practitioner’s claim: “this is the bread and butter of what I do” (2023-11-21).

Educator and experimenter with students#

He speaks as a teacher who designs assignments to “allow students to learn on their own terms and at their own pace” (2023-10-24). He also demonstrates new tools to “students and colleagues” (2023-11-29).

With industry: a constructive partner who sometimes pushes#

He says openly that he likes Waymo, with disclosure (2023-11-09).
He imagines “sitting down with folks from OpenAI” (2023-11-21).
He welcomes the Frontier Model Forum fund (“This is good news”) while warning against conventional thinking (2023-10-25).
His offer to companies is framed as helping good ideas succeed, not as policing them.

With policymakers and institutions: a supportive critic#

The Executive Order is “a start”. He also names its “naivety” (2023-10-30).
He is wry about universities positioning themselves as “intellectual and influential heavyweights” once funding appears.
He calls on them to “get their act together”.

With readers: host and co-explorer#

He invites more than he pronounces: - “Looking forward to seeing you there!” (2023-10-27); - “I really hope you’ll join us” (2023-12-01); - a recommendation “if you are even slightly curious”; - “I’ll be pushing at the boundaries of how we think about AI and the future as I lead the conversation.”

With people he disagrees with, or might mock#

He treats Musk’s Grok with irony (“the House of Elon”) but leaves room for it to live up to its name (2023-11-05).
Of AI-risk commentators he says they are wrong “not because these people don’t know their stuff” (2023-11-26).
His prescription sounds soft, and he says so: “This may sound like a bit of kumbaya thinking”. He then grounds it in “years of evidence around what hasn’t worked in the past.”

What he refuses to do#

Take sides in the speed binary (2023-11-26).
Treat “no AI” as the safe default (2023-11-26 addendum).
Claim certainty or read motives into a breaking story. “all of this is speculation at this point” (2023-11-18).
Preach or sermonise. Even his ethical point about orphan risks sits alongside the business case.
Exclude people. He rejects selective enrolment (2023-10-24).
Stretch beyond his expertise. He leaves the technical future of OpenAI to others (2023-11-18).

Changes of mind, shown in the text#

“is this really a thing?” becomes “It turns out that the concept does make sense” (2023-10-24).
The provocative title is walked back as “a little hubristic” (2023-11-26).
The addendum restores the hazard–exposure framing he had left out (2023-11-27).
“Not convinced” by the Waymo authors, he checks and finds the claim, if anything, conservative (2023-11-09).
ChatGPT’s performance with the Planner surprised him (2023-11-21).
Two updates follow the Altman story as it reverses (2023-11-18).

6. What is distinctive#

A quantitative risk scientist using first principles to widen the question. Most AI-risk voices either extrapolate to existential scenarios or list present harms. Maynard goes back to the definition of risk and shows two things. “No cause, no risk” limits speculation without a causal pathway. And “harm” and “perception” extend risk to identity, meaning and aspiration. He then refuses the “zero exposure” default (2023-11-26). The same fundamentals discipline the debate and open it up.
Risk that includes the futures people want. Treating risk as a threat to value, including value “that we may aspire to”, puts risks and benefits in one frame. It lets him ask what stops good ideas from succeeding (Galactica, OpenAI), not only what harms to prevent. The distinction between “value” and “values”, and the idea that threats are reciprocal, are uncommon in AI-governance discussion.
Answering a crisis with an experiment. While most commentators offered takes on the OpenAI board crisis, he ran a role-play counterfactual with a two-page planner and published every step (2023-11-21). The tool is presented as a catalyst for a mindset, not a compliance checklist.
Humility as a stance of expertise, not a lack of it. “the more I study … the less certain I am”. This comes from someone with thirty years in risk, and it sits against a landscape of “dogmatic overconfidence”. His humility carries over to governance: regulations “could themselves constitute risks”.
An equity lens on AI in education. In a year dominated by panic about cheating, he frames generative AI as a way to raise success “at the edges” of the learning curve. His measure is inclusion, and he treats AI’s unreliability as something that can help learning.
Fiction and play used as ways of knowing, not decoration. Douglas Adams becomes a mirror for Grok. A peanuts test probes a book GPT. A remixed meme becomes an argument about expertise. The socks disclosure turns an integrity practice into a light touch.
The widest call for expertise. He argues the ChatGPT year was driven by society as much as technology. So AI needs not only social scientists but experts in consciousness, personhood and biological intelligence, and he convenes cooperation scientists and evolutionary biologists to think about human–AI symbiosis.
Holding experimentation and caution together. He is a constant hands-on experimenter, yet he warns against society-scale “suck it and see” in “a bounded system” with tipping points. He separates learning by trying from pushing irreversible change on everyone. Few commentators hold both positions at once.

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

2023-11-26 everything-youve-heard-about-ai-risk-is-wrong. The clearest statement of his risk thinking: first principles from conventional risk science, risk as a social construct and a threat to value (including aspired futures), navigating rather than managing, and humility against false certainty. The addendum and the walked-back title show his self-correction.
2023-11-21 ai-and-risk-innovation. His method in action: responding to a crisis with a playful, transparent experiment. It also sets out the Planner’s design logic (“does not provide answers”) and the distinction between “value” and “values”.
2023-11-15 navigating-orphan-risks. The core risk-innovation concepts (risk landscape, threat to value, orphan risks, and innovation in how we think about risk) applied to AI. Risk is framed as what gets in the way of good ideas, and the new frame extends the old one rather than replacing it.
2023-11-09 waymo-safety-study-shows-benefits. His quantitative instincts: a plausibility test on the source, his own reanalysis, attention to benefits, and the open question of the right metric. It includes the socks disclosure.
2023-11-11 a-fist-look-at-openais-gpts. Building to find out: play, tinkering, delight and honest limits, including turning his own books into conversational guides.
2023-10-24 flattening-the-learning-distribution-curve. He questions his own idea in public, rejects exclusion, finds a third option by reframing, and experiments in the classroom.