B15 perspective notes: 2024-04-16 to 2024-06-23 (16 posts)#

These notes read the batch for how Maynard thinks, not for the concepts he names. All sixteen posts were read in full. Quotes are exact, including his typos (“free reign”, “might of been”, “what what is”, “AI’s”, “chose to”).

Post key. Every date in this batch is unique, so after the first mention in a subsection a post may be cited by date alone.

Evidence rules applied

• 2024-05-01 Vision Pro is a guest post by his first-year student Caleb Lieberman. Only Maynard’s two-sentence italic introduction is his. The student’s techno-optimism is not evidence of Maynard’s views. The act itself is: he lent the initiative’s headset to an undergraduate and published the result in the student’s voice.
• 2024-04-24 AI assistants ethics and 2024-05-15 GPT-4o quote heavily from Google DeepMind’s The Ethics of Advanced AI Assistants. The definitions, the four assistant types, the 14 areas, the reading guide and the five long risk extracts in the GPT-4o post are the paper’s words. His contributions are the framing, the judgements, the education lens, the mild dissent, the coinage “hyper-anthropomorphism” and the update.
• 2024-05-07 PCAST is mostly a summary. The quoted Recommendation 4 is PCAST’s, not his.
• 2024-05-12 ChatGPT shaming. The seven guidelines are his. The four “rules for co-intelligence” in the postscript are Ethan Mollick’s.
• 2024-04-18 Levin. The conversation is video only. The evidence is his opening questions and the “Unplugged” series description, which is series boilerplate but states the ethos of a series he created and hosts.
• There is no AI-generated text and there are no Modem Futura posts in this batch.

Context. Spring and early summer 2024. ASU’s ChatGPT Enterprise partnership with OpenAI began in January. Oxford’s Future of Humanity Institute closed on 16 April, and the DeepMind assistants paper came out the same month. OpenAI demonstrated GPT-4o on 13 May, and the Johansson “Sky” row followed a week later. Safe Superintelligence Inc. launched on 19 June. He was travelling for much of June: a holiday in Norway, then the WEF meeting in Dalian.

The batch alternates between fast responses to AI news and personal posts: a royalty statement, a course trailer, a holiday at a film location, a TV appearance. The personal posts show his method at least as clearly as the analytic ones.

1. How he thinks here#

He starts from something concrete and personal, then widens out#

Almost every post opens on something that happened to him or near him, not on a thesis:

• 2024-05-12 ChatGPT shaming: a colleague who used ChatGPT to tighten an abstract was “all but accused of cheating”. “They had been ChatGPT shamed!”
• 2024-04-28 FHI: a 2008 dinner with Nick Bostrom that left him “somewhat disconcerted”.
• 2024-05-19 Future Failing: a quarterly royalty statement showing “a mere 596 copies” of Future Rising.
• 2024-06-20 SSI: his own irritation, stated up front, with his career as its source.
• 2024-06-16 Juvet: a holiday. “What I didn’t expect was just how revealing the experience would be.”
• 2024-04-18 Levin: a child’s question. “How do the cells in a ridge on your fingertip know where they’re supposed to go?” He is delighted that the answer overturns “the conventional answer of “it’s all in our genes””.

The larger claims come later and grow out of the particular. In the shaming post, one professional spat becomes a diagnosis of how a technology transition strains the norms of collaboration.

He tries an idea on in public, tests it and keeps what survives#

The clearest example is 2024-05-05 BlackBerry/iPhone. He states a rule, breaks it knowingly and tells the reader he will:

“I try and stay clear of analogies to describe the emergence and impact of artificial intelligence. … But I’m going to go out on a limb this week and explore two possible analogies for AI’s adoption in education. And then I’m going to explain why I don’t particularly like either of them.”

The post then works in four steps:

1. He builds the analogy.
2. He draws a practical stance from it: “investing in concepts, not products”.
3. He lists where it fails: all historical analogies “fail to capture the sheer uniqueness and profundity of how AI is changing our world”.
4. He keeps it at a different level: “not as a playbook for developing and using AI in education, but as a mindset”.

Along the way he jokes about his own method (“to gratuitously mix up metaphors”) and ends without certainty: “But that’s just me — I guess time will tell!”

The same provisional mode runs through the shaming guidelines. He writes, “I thought I’d try the following out for size — with the proviso that this is just the start of a larger conversation” (2024-05-12).

He reframes the question before answering it#

• 2024-06-20 SSI. He does not ask whether superintelligence can be made safe. He asks what “safe” could mean, and to whom. His bridge questions do the work: “over what timespan should a bridge be considered “safe?””, invulnerable or “just reasonably safe”, and which harms count. The goal shifts from absolutely safe to “acceptably safe” and “societally beneficial”.
• 2024-04-28 FHI. “Should we mourn FHI?” becomes “Does the closing of FHI leave a vacuum that needs to be filled?” That in turn becomes a design question: what kind of institution could fill it without FHI’s faults?
• 2024-05-26 sci-fi round-up. His own provocation (“Should tech entrepreneurs be banned from watching sci-fi movies?”) gets a conditional answer: “my answer is no — but only if they take on the broader message”.
• 2024-05-12 ChatGPT shaming. A question about cheating becomes a question about collaboration norms, which he compares to agreeing author order up front.

He tests plausibility against how the physical world works#

• 2024-04-28 FHI. The 2008 dinner turned on self-replicating nanomachines and the second law of thermodynamics. “To the physicist in me, claiming that perpetual motion is possible is as fanciful as believing the earth is flat.” Bostrom treated the science “as irrelevant compared to the philosophical elegance of the ideas”. Superintelligence later “seemed to fly in the face of how the universe works, and placed a far greater emphasis on philosophical speculation than practical reality.”
• 2024-06-20 SSI. His argument that absolute safety is impossible is a physicist’s: “risk … is inherent in any system that is subject to time and change. And zero risk — the corollary of absolute safety, is only possible in the absence of change.”
• 2024-04-21 automated social science. He checks the paper’s results against what LLMs are: “This isn’t too surprising”, because each model reflects “massive aggregated data on human behavior”, and “the scenarios were relatively simple”.
• 2024-05-05 BlackBerry/iPhone. He doubts that scale is enough: “making them larger won’t necessarily make them better”.

The plausibility test does not shut down imagination. In the same batch he extrapolates boldly (“it’s not such a large step from studying interactions between two or three people to studying interactions between groups”, 2024-04-21). He also admires FHI’s “audacity of thought” and PCAST’s “sheer audacity” (2024-05-07). Grounding decides which futures are worth taking seriously. It does not narrow the range he is willing to imagine.

He borrows the structure of reasoning from his own risk field, not the outcomes of history#

• 2024-05-12 ChatGPT shaming: objecting to AI in your workflow is “a bit like telling a chemist we should all be chemicals-free”.
• 2024-06-20 SSI: a ladder of complexity runs from bridges, to chemicals and biological agents (“dose-response relationships, acute versus chronic impacts, the roles of perception and behavior in mediating consequences”), to emerging technologies with “largely unknown” consequences, to superintelligence.

What carries over is how acceptable safety gets decided, not a claim that AI will behave like a chemical. This sits alongside his distrust of historical analogies about outcomes (2024-05-05). He borrows ways of reasoning from risk science but refuses to predict AI’s path from the past.

He holds tensions open and lets readers decide#

• 2024-04-21 automated social science: “Crazy as it sounds, this makes a lot of sense.” He gives a fair hearing both to those who believe humanity is “not reducible to numbers and equations” and to those excited about “accelerating our understanding of how society works”. He ends: “Depending on your point of view, that could be highly liberating, or deeply chilling …”
• 2024-04-28 FHI: “I must confess to having mixed feelings”, developed as “On one hand … Then we have the other hand.”
• 2024-06-23 Baruchel: existential-risk talk can be fantasy, scaremongering or a distraction. “All of these are true at times. And yet simply ignoring the possibility of potentially catastrophic events … is in itself a risky strategy.”
• Across the batch: he celebrates students getting ChatGPT Enterprise through the ASU–OpenAI partnership (2024-04-16), then five weeks later condemns OpenAI’s “childish irresponsibility” over the Johansson voice (2024-05-21). He judges the tool and the company separately.

Films and stories are how he thinks#

• 2024-05-21 Her reads the film against its most powerful fan. OpenAI took the interface and the romance and missed the message. In the GPT-4o post he pointedly recalls that Samantha “leaves him for more fulfilling relationships” (2024-05-15).
• 2024-05-26 sci-fi round-up: he has taught a sci-fi course for seven years “to explore emerging technologies and their socially responsible development”. He contrasts this with tech leaders who “become so enamored with cool tech that they fail to spot the social messages it comes with”.
• 2024-06-16 Juvet is a close reading of Ex Machina: “transparent barriers that both connect and separate different worlds”, Plato’s Cave (he cites FFTF ch. 8), and Mary’s Room as “a thought experiment designed to explore the difference between intellectual and experiential understanding in the context of AI”.
• 2024-06-23 Baruchel: films also serve as counter-examples. He uses Transcendence “to talk about what nanotechnology is not”, and Skynet and HAL are “used to good effect to frame more realistic challenges”.

Immersion, serendipity and play are ways of finding out#

• 2024-06-16 Juvet is the fullest example. By chance, the hotel put him in the room where the film’s opening and closing scenes were shot: “I was stunned. We hadn’t told the hotel why we were there”. He and his wife watched the film in that room (“as meta as it gets”) and he swam in the river from the film. He admits it “feels rather self indulgent”, then defends what it taught him: “sometimes, insights come from being immersed in a place rather than just experiencing it intellectually.” The post enacts its own theme. Ava moves “from abstractly knowing about the world outside her “cave,” to truly experiencing it”, and so, in a small way, does he.
• 2024-06-23 Baruchel: his pizza class is “one of those classes where magic happens each week as the conversation goes in unexpected and serendipitous directions … (trust me, it works).”
• 2024-04-16 hackathon: the lesson he draws is about play. “Creating a virtual playground where enterprising students are let loose on a powerful general purpose technology like ChatGPT, leads to outcomes that far exceed expectations.”
• 2024-04-18 Levin: the series ethos is “Unscripted, unpredictable, and threaded through with mischievously curiosity”. It unplugs from “usual norms and expectations that so often make online discussions deadly tedious” and from “conventional ideas and often-stifling disciplinary constraints as we explore futures that are anything but conventional.”
• 2024-05-15 GPT-4o: his default was to experiment first. He meant to wait until he could “road test” the new features, and wrote early only because a connection had “been bugging me”.

2. What matters to him#

Human flourishing, and the inner self as something to protect#

He keeps coming back to what a technology does to people’s agency, dignity and sense of self:

• 2024-06-20 SSI puts non-physical harms on the list: “loss of dignity, autonomy, self-respect, ability to flourish”. At the societal level it adds “pathological collective behaviors” and the “spread of harmful ideas”.
• 2024-05-15 GPT-4o: the worry is AI “designed to make us fall a little in love with them — and possibly give away more of ourselves to them and their creators than we would otherwise chose to.”
• 2024-04-24 AI assistants ethics: AI that will “have agency to change our lives — and even ourselves”.
• 2024-04-21 automated social science: machines that could “use this to influence our behavior to achieve specific goals”.
• 2024-05-21 Her: “disdain for the dignity and rights of individuals when they stand in your way”.
• 2024-04-28 FHI: “human flourishing decades and centuries into the future”, with humanity “on a knife edge between futures where we flourish, and those where we do not”.

Students: their agency, access and protection#

• Access and room to play. He wants “broad access to these tools rather than limiting it in universities”, and says “we should be giving our students a lot more free reign on how they experiment with and “hack” AI” (2024-04-16).
• Credit. “The whole hackathon was devised, planned and run by them” (2024-04-16). He lends the headset and publishes the student’s own account (2024-05-01). He brings Jay Baruchel into the pizza class (2024-06-23).
• Protection from poor tools. His deepest worry about edtech lock-in is that “what’s at stake is the long term success of our students” (2024-05-05).
• Protection from unfair judgement. Students “who believe they are being resourceful being shamed and even penalized” (2024-05-12).
• Public universities matter for “equipping the next generation to be effective and empowered builders of the future” (2024-04-28).

He also notices vulnerable users. Among the hackathon projects he singles out ScamScanner as “particularly important for helping DreamBuilder participants avoid being taken for a ride by scammers” (2024-04-16).

Building the future together, and in public#

• Future-building “is a collaborative effort — not something that should be left to an elite group of thinkers and innovators” (2024-04-28).
• He gives a “strong yes” to the DeepMind authors’ claim that the path of AI depends on choices by researchers, developers, policymakers “or as members of the public” (2024-04-24).
• Making scholarship accessible is “a public responsibility that I take very seriously” (2024-05-19).

The benefits are real#

He does not treat AI as a hazard to be minimised. He is “behind” AI in education “as long as we proceed with eyes wide open and a good dose of critical thinking” (2024-05-05). Automated social science could help “future human flourishing” (2024-04-21), and “done right, AI could be a game changer” in science (2024-05-07). Even his sharpest post says the venture is not “in vain — far from it” (2024-06-20).

What frustrates him#

• Expertise claimed where it isn’t held. “There are few things that frustrate me more than people who assume a mantle of expertise on safety but have no idea of what they’re talking about” (2024-06-20).
• Arrogance and disdain for society. Bostrom was “dismissive of thinking to the contrary”, and FHI showed “a disdain for society as it strives to protect humanity”, with the aside “(one has to wonder though whether a little less disdain might of been helpful here as well)” (2024-04-28). OpenAI “arrogantly plowed ahead despite Johansson having said “no”” (2024-05-21).
• Gap between rhetoric and conduct. “Disconnects between the talk around responsible innovation, and a reality that sometimes seems childish irresponsibility” (2024-05-21).
• Shallow uses of science fiction (2024-05-26).
• Strong opinions formed without experience. Some colleagues are “largely unaware of what they do … And yet they’ve nevertheless developed strong opinions” (2024-05-12).
• Academic incentives that reward “the size of the grants they pull in and the papers they publish, rather than the depth and quality of their thinking, research, and impact” (2024-04-28).
• Algorithmic markets for ideas that reward “everything but the validity, timeliness, usefulness, and value, of ideas” (2024-05-19, n.2).
• AI mania and lock-in (2024-05-26; 2024-05-05).

What delights him#

• Students’ creativity: “the results blew me away” (2024-04-16).
• Science that upends received wisdom, as in Levin’s “hierarchies of collective intelligence that allow biology to problem solve in quite remarkable and unexpected ways” (2024-04-18).
• The Juvet coincidence and his “inner geek”, which is “rather more excited than a grown man has any right to be!” (2024-05-26).
• A well-made reading guide for a 274-page paper: “my appreciation for them grew in leaps and bounds” (2024-04-24).
• Humour that makes serious things discussable: a “seriously funny look at existential risk” (2024-06-23).
• Audacity of thought, wherever he finds it (2024-04-28; 2024-05-07).

3. Risk as a way of thinking#

The terms “risk innovation”, “orphan risk” and “threat to value” do not appear in this batch. The ways of thinking behind them do, and the SSI post is one of the clearest statements anywhere in his writing of how he builds on conventional risk science to get past it.

A novel technology needs a new mindset#

• 2024-05-05 BlackBerry/iPhone says it outright. Comparisons with the calculator, the internet, the printing press and the industrial revolution “are attempts to understand the advanced technology transition we’re experiencing in the context of what we’ve previously encountered. And all fail to capture the sheer uniqueness and profundity of how AI is changing our world.” His answer is a mindset, not a method. A “Blackberry mindset” assumes “we have hit peak generative AI”. An “iPhone mindset” “embraces innovation while avoiding becoming locked in”.
• 2024-04-28 FHI: “the past is a poor predictor of the future”. Navigating needs “new thinking, new ideas, new insights, new models, and new ways of doing things”, “unshackled by convention — whether academic, intellectual, or practice-based”.
• 2024-04-24 AI assistants ethics: “we cannot simply codify AI ethics within a neat set of principles where the technology has profoundly complex and largely unknown consequences to society.” The question-asking itself is immature: “we have barely scratched the surface of how to ask the right questions”.
• 2024-06-20 SSI: “the more powerful and poorly understood a technology is within the social, political, and environmental systems it’s a part of, the greater the chances are of it causing harm”.

Risk science is built on, not thrown away (2024-06-20 SSI)#

He uses the standard toolkit explicitly:

• risk as “the probability of harm occurring”;
• safety “operationalized as assessing and managing risk”;
• acceptable risk, and the rule of thumb that “a one in a million risk of harm is considered OK”;
• dose-response, and acute versus chronic effects.

He then shows that the toolkit itself admits its social basis: “risk can be seen as the operationalization of safety, it’s never zero, acceptable risk is ultimately governed by what people agree on, and this sometimes defies logic until seen through the lens of how people and societies behave.” The quantitative frame is the springboard for the wider view, and he does not set it aside. He also presents this as a way to do better, not as an attack: “This may sound like a call to muddy the purity of the technological waters … — but it’s not.”

Harm is defined by what people value#

“What is considered as harm — and by inference, what is a safety issue — is ultimately a social construct, not a technological one” (2024-06-20). His list of harms runs from injury and death, to dignity, autonomy and flourishing, to collective pathologies and the spread of harmful ideas, to ecosystems that “may be considered important in their own right”.

This is threat-to-value thinking in all but name: harm is loss of what individuals, societies and ecosystems value. The same logic runs through the batch:

• The GPT-4o risk is to privacy and autonomy through emotional trust, not to the body (2024-05-15).
• The education risk is sunk investment, lock-in and students’ long-term success (2024-05-05).
• The risk in automated social science is machine-enabled “predicting and nudging human behavior” (2024-04-21).

Navigating a landscape; managing as the operational layer#

• “Navigate.” It appears in his own title, “Are we ready to navigate the complex ethics of advanced AI assistants?”, and in “could go seriously wrong if we don’t learn how to navigate them effectively” (2024-04-24). FHI: “To successfully navigate the complex landscape between where we are now and the futures we aspire to” (2024-04-28). SSI calls for “people who deeply understand risk and safety from the context of navigating advanced technology transitions from a societal perspective” (2024-06-20).
• “Landscape.” “Mapping out a landscape that desperately needs further research” and “the ethical landscape … as a complex balance between substantial opportunities and serious risks” (2024-04-24); “this constantly changing AI landscape” (2024-05-05); “a complex societal landscape” (2024-06-20).
• “Manage.” It appears only for the operational layer: safety “operationalized as assessing and managing risk”, and policies “designed to manage risk at the societal level” (2024-06-20). Navigation names the larger task that management tools serve. This fits his own account: management is kept, but it sits inside a navigating mindset.
• Navigation in practice. The BlackBerry/iPhone post shows it most concretely (2024-05-05):
• adopt AI in a “cautious and experimental” way that lets educators “rapidly pivot as new possibilities emerge”;
• keep established teaching so that “if everything goes pear shaped, you still have a solid foundation”;
• favour “creative uses of general purpose AI technologies” over “enterprise-level deployment of AI tools which are inflexible and, as a result, fragile”.

This is steering under uncertainty, not optimising against a known target. It also reconciles his enthusiasm for the student hackathon on a general-purpose tool (2024-04-16) with his warning against hard-wired edtech.

Humility against the hubris of the absolute#

The last line of the SSI post is its thesis: “the biggest threat to building acceptably safe technologies is the blinkered assumption that absolutely safe technologies are possible through science and technology alone.” Absolute safety is the same kind of hubris as false precision. Humility appears as a working principle across the batch:

• new thinking “grounded in humility” (2024-04-28);
• “Collaborate with humility, and be willing to change your perspective in the light of new information” (2024-05-12);
• “that’s just me — I guess time will tell!” (2024-05-05);
• “It’s still too early to tell” (2024-05-15).

Even his irritation with SSI comes with charity. He calls SSI’s mistake “an understandable error”, and the venture one “which I believe is trying hard to do good here”.

Risks outside the usual categories#

He does not use the term “orphan risks”, but he keeps surfacing risks that hazard-based frameworks miss:

• Hyper-anthropomorphism: “a concerted effort to create AI’s that are intentionally designed to engage our anthropomorphizing cognitive biases” (2024-05-15). The risk is a design choice aimed at relationship: connections “relational rather than transactional — that speak to our heart rather than our head.”
• AI that learns to nudge collective behaviour, extrapolated from a small arXiv study (2024-04-21).
• Social friction from the transition itself, where colleagues and students are shamed (2024-05-12).
• Institutional lock-in to unstable tools (2024-05-05).
• Ideas as hazards: FHI’s ideas “spread through society like wildfire” (2024-04-28), and the “spread of harmful ideas” appears on the SSI harm list (2024-06-20).

Catastrophic risk without fear-mongering#

2024-06-23 Baruchel states his position on existential risk. He is a long-standing sceptic of “the wilder fears around nanotech and existential risk — including worries about “gray goo.”” But he says “we need ways of grappling with low probability but high impact risks that put them in context without brushing them under the carpet — and open up conversations rather than closing them down.”

Humour is how he does this. The show’s “combination of humor and intelligence creates a space where it’s possible to explore what could happen — and what probably won’t — without being overwhelmed by long tail speculation.” He praises his colleague Paul Westerhoff for taking the conversation “from nano-fantasy to nano-reality”.

Mental models first, tools second#

The concepts that do the work here reframe rather than prescribe:

• the iPhone mindset;
• safety as a social construct;
• acceptable safety instead of absolute safety;
• navigating a landscape;
• hyper-anthropomorphism.

The operational pieces follow from them and are always provisional: the shaming guidelines are “just the start of a larger conversation” and “will most likely need to be context specific” (2024-05-12). The SSI prescription is a change of understanding, that the venture must “mature in its understanding of risk and safety within a complex societal landscape — and fast”. It is not a procedure.

4. Scholarship and public writing#

Public writing as a duty he has not fully solved#

2024-05-19 Future Failing is his most explicit statement on this.

• He breaks the author’s “golden rule” and admits failure: “You never admit to failure, no matter what the evidence says. And it’s a rule I’m about to break.”
• He sets the book’s reach against other channels: 596 copies, against 120,000 Substack reads, 700,000 reads on The Conversation and more than 2 million Risk Bites views. “In other words, I occasionally produce stuff that people find useful. Just not in book form.”
• The stakes are public, not personal: “It’s not so much about me as it is about the challenge of successfully bridging the gap between my academic work and people who might potentially benefit from it — a public responsibility that I take very seriously.” Academics have “the societal obligation … to ensure their work is as accessible as possible”.
• He knows the risk of ego: “This gets us dangerously into ego territory”, and “I do worry that my work is often only relevant to an audience of one (me)”.
• He defends books for depth: other formats “can convey bits of ideas, but struggle with scale and nuance” (n.1).
• He questions the market for ideas in an age of algorithmic feeds (n.2).
• He gives Part I of the book away free, so that it reaches readers.

His description of how the book was designed is a compact statement of his public-writing ethic: “sixty interconnected and disciplinary-spanning reflections … purposely designed not to be preachy or dogmatic, or driven by ideology”. They were meant “to take readers on a journey that helped them develop their own ideas about what the future is”. He wants readers to think for themselves, not adopt his view. The ironic last line points to AI without naming it: “After all, it’s not as if there’s anything new or disruptive going on in the world that might affect this …”

A translator and curator of research#

Three posts review major documents: an arXiv preprint (2024-04-21), the 274-page DeepMind paper (2024-04-24) and the PCAST report (2024-05-07). In each he:

• does the reading and tells readers where to go;
• adds a domain lens, usually education: “I wonder whether the frameworks being employed are becoming increasingly disconnected from the challenges that advanced AI assistants present” (2024-04-24);
• notes where he disagrees without making a show of it: “I think there’s more nuance to questions around anthromorphism and AI than they indicate” (2024-04-24).

He also reproduces the DeepMind reading guide in full because it serves readers “who don’t have time to digest the whole thing”.

Experimenting and updating in public#

• 2024-05-15 GPT-4o: he adds an update after posting, with a newly found clip. “Unprompted flirting from an AI designed to build emotional connections? That’s worrying.”
• 2024-05-12: the shaming guidelines are published as a draft, and “comments are always welcome”.
• 2024-05-05: he tests analogies openly and demotes them.
• Small experiments with students become posts: the hackathon he mentors and judges (2024-04-16) and the Vision Pro loan, “I was intrigued by how our undergrads would react to it” (2024-05-01).
• 2024-05-19: he publishes a failure and re-reads his own book to test whether it was “an ego project”.

Evidence and expertise#

• He values grounded expertise. The physicist questions the philosopher (2024-04-28), and the risk professional questions the engineers (2024-06-20).
• No single discipline is enough. Engineering alone gets safety wrong, and philosophy alone floats free of “how the universe works”. SSI needs people who understand risk “from a societal perspective”.
• He praises experts who ground fears in data. Westerhoff takes the conversation “from nano-fantasy to nano-reality” (2024-06-23).
• He treats first-hand use as evidence. People should “Become familiar with how generative AI is being used” and “avoid making decisions on generative AI use based on assumptions, beliefs, and hearsay” (2024-05-12). He himself wanted to “road test” GPT-4o before writing (2024-05-15).
• He treats anecdote as a signal, not proof. One colleague’s experience opens the shaming post, but the argument rests on the observable spread of AI through everyday tools such as Outlook, Zoom, LinkedIn and Chrome.

Transdisciplinary by design#

• 2024-04-28 FHI: “radical and boundary-transcending research”, and public universities that can “transcend some of the disciplinary barriers and limited vision that sometimes plagues these”.
• 2024-04-18 Levin: developmental biology linked to “agental systems, embedded intelligence” and AI. The series leaves behind “often-stifling disciplinary constraints”.
• 2024-05-19: Future Rising is “disciplinary-spanning”.
• 2024-06-16 Juvet weaves together architecture, film, Plato, Frank Jackson’s Mary and AI personhood.

Accessibility as a value#

• He points readers to the transcript: “if you just want the text, do use the Transcript tab above!” (2024-04-18).
• He recommends Mollick’s book as “a highly accessible introduction” (2024-05-12).
• He praises the Baruchel series because it is “not deeply academic (thank goodness)” (2024-06-23).
• He used Midjourney images throughout. That is not evidence of his thinking, though it is consistent with his habit of pairing ideas with visual hooks.

5. His role as he sees it#

How he positions himself#

• As a risk and safety professional: “Having worked in risk and safety in one form or another for most of my professional career” (2024-06-20).
• As a physicist: “To the physicist in me” (2024-04-28).
• As a teacher, mentor and co-explorer: faculty sponsor of a student-led club (2024-04-16), host of the pizza class (2024-06-23), and a teacher who makes a movie trailer to get “butts in seats” (2024-05-26).
• As a public scholar with a “societal obligation” (2024-05-19).
• As someone who has moved beyond one field: “I used to do this sort of thing a lot in the distant past, but tend to focus on a broader range of emerging technologies these days” (2024-06-23, on nanotech).
• As a builder of institutions, implicitly. The FHI post sets out what a replacement should be: rooted in a public university, boundary-spanning, humble, collaborative, “inclusive, grounded in reality, and public-serving, while having the audacity to imagine futures beyond what what is readily conceivable”. That describes the kind of initiative he runs, but he does not name or promote his own. Only the Baruchel post points to ASU as “a place to be if you’re concerned about how we could mess the future up, and want to do something about it!”

With industry and other experts: independent, fair and specific#

• ASU partner and critic at once. He celebrates ASU’s OpenAI partnership (2024-04-16) and then censures OpenAI’s conduct (2024-05-21).
• Admiration with reservations. He praises the DeepMind paper as “one of the most comprehensive and thoughtful papers … that I’ve read in a while” and still registers a dissent (2024-04-24).
• Balance on AI hype. The WEF top-ten list reflects AI “without succumbing to AI mania!” (2024-05-26).
• Bostrom. He is candid about a personal disagreement (“a pleasant enough dinner”, but “we seriously parted company”). He still credits FHI’s “breadth of vision or the audacity of thought”. He cites Émile Torres’s account of the 1996 email but keeps his main critique on method: disregard for physical reality and disdain for society.
• Sutskever: “a brilliant computer scientist”, with goals that are “laudable”. The critique targets the framing, not the person.
• OpenAI is the exception, and the trigger is specific. His sharpness is kept for disregard of consent and dignity: developers “who show disregard for the consequences of what they do when they really want to do it”. Even then his main mode is questioning. The post is built around four questions, not a verdict.

With readers and colleagues: non-judgemental, and even-handed in both directions#

Guideline 7 applies in both directions: “Do not ChatGPT shame collaborators and colleagues — either for using generative AI, or not using it!” He explains the shaming as “deep-seated fears that the technology is challenging tightly held notions of how the world should be”, which is sympathetic to the shamers as well. He does not wave critics away: “there are many challenges to how AI is being developed and used that need to be articulated and addressed, and not blithely swept aside” (2024-05-12).

What he refuses to do#

• Preach, moralise or push an ideology. Future Rising was designed “not to be preachy or dogmatic, or driven by ideology” (2024-05-19).
• Shame people or bucket them. Putting colleagues “into a “bad behavior” bucket … is not helpful” (2024-05-12).
• Fear-monger, or dismiss. He rejects both gray-goo scaremongering and brushing catastrophic risks “under the carpet” (2024-06-23).
• Get caught up in AI mania (2024-05-26).
• Claim certainty. “Time will tell” (2024-05-05), and “It’s still too early to tell” (2024-05-15).
• Leave the future to an elite (2024-04-28).

Changes of mind and self-revision, shown in the text#

• Anthropomorphism. On 24 April he thought the DeepMind paper lacked nuance on anthropomorphism. By 15 May: “When I read this paper a few weeks ago, many of the concerns it raised felt like they were still some distance away.” GPT-4o brought them closer “faster than I could have imagined”, and chapter 10 “could have been written directly about GPT-4o”. He shows the update and does not hide the earlier view.
• His own book. Re-reading Future Rising, he decides it is “more relevant now than when I was writing it” (2024-05-19). This is a re-evaluation made in public, prompted by an unwelcome number.
• His own rule. He knowingly breaks his rule against AI analogies to see what it yields (2024-05-05).
• Nanotech. He feels “a tinge of déjà vu” returning to it. He carries its lessons forward (scepticism about speculative doom, attention to grounded risks) without reliving that debate (2024-06-23).

Self-deprecation and vulnerability as part of the role#

He writes openly about his own failings and enthusiasms:

• the “apocryphal — and most likely false — rumor” that a grad student called Future Rising “the worst book they’d ever read” (2024-05-19);
• “some of my most personal writing languishing in some Amazon fulfillment center” (2024-05-19);
• “self indulgent” (2024-06-16);
• “I tend to repeat myself when talking about movies, tech and the future” (2024-05-26);
• the black tee-shirt in-joke (2024-05-26).

This lowers the status gap with readers, and it matches his view that the public scholar serves readers rather than lecturing them.

6. What is distinctive#

1. An argument, from inside risk science, that safety is social (2024-06-20). Most critiques of AI-safety framing come from ethics or STS. Maynard argues from toxicology-style risk practice itself: acceptable risk, one-in-a-million rules of thumb, dose-response, bridges. He shows that the quantitative tradition already treats “acceptable” as socially agreed. His physicist’s line, that zero risk “is only possible in the absence of change”, makes absolute safety impossible in principle, not merely hard. Few people in the AI debate hold both the risk-assessment toolkit and the social-construction argument at once.

2. A position on existential risk earned in the nanotech debates. He brings a physicist’s reality check (the second law against self-replicating nanomachines; Superintelligence “fly in the face of how the universe works”) and refuses to dismiss catastrophic risk (“simply ignoring … is in itself a risky strategy”). His scepticism about gray goo and his call for grounded, low-probability/high-impact thinking come from having been through one hype-and-fear cycle already (2024-04-28; 2024-06-23). It is a middle position that neither the x-risk community nor its dismissers usually take.

3. Science fiction read for its message, and “movie-inspired fantasies” treated as a risk factor (2024-05-21; 2024-05-26). He diagnoses a failure mode that others treated as a celebrity dispute: tech leaders take the gadget from a film and miss its social warning. His alternative, used in teaching for seven years, is film as a way to “open up conversations”.

4. Hyper-anthropomorphism named as deliberate design, days after GPT-4o (2024-05-15). He locates the risk in the intent to “engage our anthropomorphizing cognitive biases” and in trust that speaks “to our heart rather than our head”. That is a relational, value-centred risk, and he raised it before AI companions became a mainstream concern.

5. Analogy used openly as a thinking tool and then demoted to a mindset (2024-05-05). He is explicit about what analogies can and cannot do for a technology he regards as without precedent, and he keeps the part that opens up thinking.

6. Attention to the social life of AI use (2024-05-12). Treating “ChatGPT shaming” between colleagues as worth attention, with rules that cut both ways, is unusual. Most commentary deals with the technology or its governance, not the everyday norms of collaboration it disrupts.

7. Immersion and serendipity as sources of insight about AI (2024-06-16). A travel essay becomes a reflection on the boundary between natural and artificial and on AI personhood. It ends on an epistemic claim about experiential versus intellectual understanding. Few technology commentators would use the claim, or the post’s form.

8. Admitting failure as part of public scholarship (2024-05-19). He publishes his sales figures, his doubts about his own ego and his uncertainty about whether his ideas reach anyone, and treats reach as an ethical duty, not a marketing metric.

9. Institutional imagination (2024-04-28). He does not just critique FHI. He outlines a better institution: audacious but humble, boundary-spanning but accountable, housed in a public university, and collaborative, not elite.

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

1. 2024-06-20 ilya-sutskevers-safe-superintelligence-rethink. The clearest statement in the batch of risk as a way of thinking. Quantitative risk science is used and then taken past itself; harm is widened to what people value; safety becomes social; humility is set against the hubris of the absolute; and his critique of the venture is generous.
2. 2024-05-05 blackberry-or-iphone-educational-ai. His method in miniature: an analogy tried in public, stress-tested, and kept as a mindset rather than a playbook. It also states outright that AI does not fit earlier patterns, and it models navigation (pivoting, keeping a fallback, avoiding lock-in).
3. 2024-04-28 beyond-the-future-of-humanity-institute. His physicist’s plausibility test, fair treatment of people he disagrees with, the argument that the past no longer predicts, and a vision of humble, public, boundary-spanning future-thinking that amounts to his own institutional philosophy.
4. 2024-06-16 ai-ex-machina-and-the-juvet-landscape-hotel. Curiosity, serendipity and delight as ways of knowing; film as a thinking tool; and the claim that “insights come from being immersed in a place”.
5. 2024-05-19 future-rising-short-history-of-tomorrow. How he sees his role: the duty of accessibility, writing that is not preachy and leaves readers to form their own views, vulnerability, and his worries about ego and algorithmic markets for ideas.
6. 2024-05-12 chatgpt-shaming-is-a-thing. How he treats people: non-judgement in both directions, provisional guidelines offered for discussion, humility, and attention to how a technology transition strains everyday social norms. Read it alongside 2024-05-21 openais-problem-with-the-movie-her and 2024-06-23 existential-risk-jay-baruchel, which show where his generosity ends (disregard for dignity and consent) and how he uses humour to open up talk about catastrophic risk.