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

Reading notes on Andrew Maynard’s Substack posts in batch B12. Only his own prose counts as evidence. Quotes are exact, including his typos and curly punctuation.

Batch context. These five weeks run from just before the Biden AI Executive Order (30 Oct 2023) and the UK Bletchley Park AI Safety Summit (1–2 Nov), through OpenAI DevDay and custom GPTs (6 Nov), the Cruise robotaxi suspension, and the firing and return of Sam Altman at OpenAI (17–22 Nov). The batch ends at ChatGPT’s first anniversary. Maynard responds in three ways: - He reframes AI risk from first principles. - He brings his ASU Risk Innovation work (risk as a threat to value, orphan risks, the Risk Innovation Planner) to bear on AI companies for the first time. - He keeps arguing that the AI transition is socio-technical and cannot be left to AI experts or large companies.

Two posts are guest essays by his ASU colleague Brad Allenby. Several are event notices for the ASU Future of Being Human … Unplugged livestreams. There are no Modem Futura podcast posts in this batch; the user asked for those to be skipped, so nothing had to be left out.

Relevance summary:

HIGH#

2023-11-26 — everything-youve-heard-about-ai-risk-is-wrong — “Why everything you’ve ever heard about AI risk is wrong”#

Provenance. His own prose throughout, including an addendum dated 27 November. The only other material is a meme image he adapted.

Argument. The title is deliberately provocative, and he later softens it himself. The claim is that almost all of the past year’s commentary on AI risk is wrong. This is not because commentators lack expertise. It is because the challenges are “so novel, so complex, and so cross-cutting” that no single person or discipline can handle them. He writes with the authority of more than 30 years in risk and 20 on the novel risks of emerging technologies, but the more he studies AI, “the less certain I am that we even know how to formulate the problems we face around AI”. He describes the result as an “understanding-vacuum that is proving exceptionally easy for people to fill with their own ideas, ideology, or speculations”, which social media amplifies. Individual contributions can be useful, but “few if any are trustworthy on their own”, and at worst they hide complexity behind “dogmatic overconfidence”.

Risk from first principles. This is the most systematic statement of his risk concept in the batch. He starts from the standard definition, “the probability of harm occurring from an action, process, or situation”, and unpacks it into five elements:

1. Cause and effect. “no cause, no risk.” He separates hazard, the potential to cause harm, from risk (“bleach is hazardous, so is a piano”). He treats speculative scenarios such as “AGI going rogue” as hazard or speculation: without a causal pathway to outcomes, there is no risk. This is an implicit critique of existential-risk framings that skip the causal pathway.
2. Magnitude. Risk only makes sense with some sense of how large the harm is, whether quantitative, qualitative or comparative. Otherwise resources flow to negligible risks while substantial ones are ignored. Because “negligible” and “substantial” must be agreed, “risk is ultimately a social construct as it reflects broad societal values and norms.”
3. Harm. Harm is defined by social values and norms more than by science. If harm means diminishing something of value, the concept stretches to “mental wellbeing, happiness, sense of worth, identity, beliefs, aspirations”. It can also cover harm to future generations, as effective altruism frames it. The further it stretches, the harder the causal links are to define. This connects directly to his idea of risk as a threat to value.
4. Time. He draws on toxicology: long-term effects of acute exposure, and chronic exposure, blur cause and effect even in conventional risk science. It is much harder when neither cause nor effect is well defined, “and this is where we are with most emerging technologies, AI included.”
5. Perception. Responses to risk are rooted in evolution, cognitive biases and heuristics, so “an understanding of risk cannot be separated from an understanding of human behavior and societal dynamics”. This matters most when the harm touches “what makes us human”: identity, beliefs, worldview, relationships. Risk also covers threats to value we aspire to. It is “not just about protecting what we have, but protecting pathways to futures we desire.”

Applied to AI. “If risk in general is messy, AI takes this messiness to a whole new level.” Relevant effects are poorly understood, causes are uncertain, and the causal threads between them may be “unknown (and possibly unknowable)” across timescales. Beyond the minimal premise that a powerful technology will cause some harms, we don’t know “pretty much everything else”. His list includes harms from not developing capabilities, and whether regulation itself could be a risk: “We don’t even know whether actions designed to manage potential risks — including regulations — could themselves constitute risks.” He frames the uncertainty symmetrically: “the risks of going too fast, the risks of not going fast enough, or the risks of simply assuming there are no risks”.

Prescription. He calls for humility, collaboration and agility. We can navigate this “only if we drop our egos, recognize the limits of our own understanding, and have the humility to listen to and learn from others”. He anticipates the “kumbaya” objection and says the call rests on “years of evidence around what hasn’t worked in the past”. He asks for collaboration across disciplines “that far exceed what we currently achieve”, ways of understanding “that transcend science and technology alone”, and agility to course-correct. He applies the humility to himself: “Maybe I was a little hubristic … there’s probably some stuff there that’s less wrong. Although the more I learn, the less sure I am.” He is firm on the diagnosis (deep uncertainty, overconfident commentary) and on the process (humility, transdisciplinarity), and tentative about substance.

Addendum: risk = Fn(hazard, exposure). Here he explicitly uses the chemical risk-assessment paradigm, structurally and conceptually rather than literally. He explains why he left it out at first: - He wanted a broader risk concept than “the hazard-exposure paradigm that dominates approaches to chemical risk assessment”. - The paradigm “can get gnarly” with non-linear dose-response: thresholds, hormesis, low-dose effects. - “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.” This is an explicit rejection of abstention as the precautionary default.

He then tries the paradigm on AI. Exposure is necessary to turn hazard into risk, but the transforming function might be linear, thresholded, inverse or time-dependent. For AI: - Hazard could be “as obvious as disrupting financial services, or as subtle as influencing human behavior”. - Exposure could be an AI with “access to and the agency to manipulate critical systems”, or “as intangible as hints of ideas encountered over hours of social media use”.

This is an early statement that exposure to AI can be cognitive and cumulative, in line with his later concerns about language and formation. He concludes that there is not “even the beginnings of a framework” for AI hazard, exposure and transfer function. The paradigm may still have “mileage”, but only within broader transdisciplinary efforts to reformulate risk “in the face of a technology that defies conventional thinking”.

Views. - On AI: a powerful, novel technology whose risk landscape is “increasingly looking unlike anything we’ve had to navigate together before”. - On existential risk: agnostic. He lists “whether AI holds the potential to crash human civilization” alongside mundane risks and the chance that we need AI and AGI to thrive, and does not take a side. - On experts: he criticises the overnight AI-risk experts, and also established AI experts who are “in danger falling into a similar trap”.

Quotes. - “no cause, no risk.” - “risk is ultimately a social construct as it reflects broad societal values and norms” - “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.” - “We don’t even know whether actions designed to manage potential risks — including regulations — could themselves constitute risks.”

2023-10-30 — white-house-goes-all-in-on-responsible-ai — “White House goes all in on responsible innovation and artificial intelligence”#

Provenance. His own prose, written from the EO fact sheet on the day of release. It quotes the fact sheet, and also quotes his co-authored 2023 Nature Nanotechnology commentary with Sean Dudley on advanced technology transitions, which is co-authored prose.

Argument. The Executive Order is “not your run-of-the-mill angst over generative AI” or classroom cheating. It is a forward-looking attempt to navigate “some of the most transformative advances in advanced tech in decades — possibly centuries”, and should be taken seriously. He welcomes its message that benefits require navigating “deeply novel and highly challenging” risks. He puts this in historical context through the co-authored line that “transformative technologies always come with unintended and often hard-to-anticipate uses and consequences, and we often ignore or trivialize these at our peril”. He adds that AI’s transformation is “in many ways, substantively different from those associated with previous technologies.” So he claims both continuity (every transformative technology brings unintended consequences) and novelty.

He names three pitfalls in the new safety awareness: - regulatory capture; - “narrow thinking around governance and regulation”; - a lack of awareness of the “complex dynamics surrounding integrated sociotechnical systems”. He links this last point to Allenby’s guest post, which suggests he endorses it.

AI is “a foundational set of technologies”, tied to climate, infrastructure, cyber, equity, health, sustainability and discovery.

He highlights five themes: 1. Responsible innovation. The US has been hesitant compared with Europe, “perhaps because there’s still very much an ethos of go fast, break things, and don’t worry about the consequences here”. He hopes experts will step up and AI leaders will “recognize and respect this expertise”. 2. Public Interest Technology, implicit and explicit in the EO. 3. Critical infrastructure, including chemical, biological, radiological, nuclear and cyber risks. 4. Equity, justice and civil rights. He notes the EO’s phrase “irresponsible uses” and argues that social-justice experts must be “partners in developing responsible and beneficial AI, and not just add-ons.” 5. Transdisciplinary research on advanced technology transitions. This is his main takeaway.

Critique. The EO reflects “a naivety that is, in part, a result of a gaping lack of understanding around how we collectively navigate such transformative technologies”. He itemises that naivety: - “an over-emphasis on following the lead of large tech companies and their agendas”; - too little diversity of expertise; - “a lack of evidence for learning from past technology transitions”; - little recognition of innovative governance; - weak support for new thinking on AI’s transformative societal impact.

He sees this as an opportunity for universities, with arts and humanities included. He is also wry that institutions will scramble to become “intellectual and influential heavyweights” when funding arrives. He closes with a warning against “the trap of conventional thinking and vested interests wrapped in the clothing of social responsibility”.

Views. - On governance: broadly supportive of federal action, calling it “a start”; unsure whether it is “strong enough or informed enough”. - On companies: government leans too heavily on big tech. - On past technologies: he says the EO fails to learn from past transitions, but gives no specific cases here.

Quotes. - “an ethos of go fast, break things, and don’t worry about the consequences here” - “an over-emphasis on following the lead of large tech companies and their agendas” - “a lack of evidence for learning from past technology transitions” - “the trap of conventional thinking and vested interests wrapped in the clothing of social responsibility”

2023-11-29 — the-year-that-generative-ai-changed-the-world — “The Year that Generative AI Changed the World”#

Provenance. His own prose.

Argument. On ChatGPT’s first anniversary, he argues that the world changed not mainly because of the technology but because of its social setting: “it wasn’t so much the technology as the social and technological landscape surrounding it”. GPTs existed before as a niche technology. ChatGPT made them accessible and showed, “with meme-like virality”, how human-like natural-language interfaces had become. His own demonstrations produced “jaw-dropping moments”. He lists twenty changes. Among them: - “A whole new way of interacting with tech … that more closely mimics human to human interactions”; - education “shaken to the core”; - people who “think for a living” fearing replacement; - AI experts as “overnight celebrities”; - “a dramatic shift from talking about AI ethics to addressing AI risks”; - worries about democracy; - “exponential risks” going mainstream; - “Responsible AI has become a thing”; - OpenAI politics “more compelling than reality TV”.

Main lessons. 1. AI is part of “a highly complex ecosystem of people, communities, organizations, and governments”, so “our AI future cannot be left solely to AI experts”. He goes further: broader scientific expertise is needed, including on intelligence (biological or not), consciousness, neuroscience, philosophy, “the nature of personhood”, responsible innovation and ethics. This is because we are building machines that may have “their own version of intelligence — and possibly consciousness and self-awareness along with it.” 2. Against “suck it and see.” Trial and error is how innovation has worked “for most of human history”. But in “an increasingly complex and bounded system” with non-linear cause and effect, there is “an increasing likelihood of naive experimentation pushing us past hard-to-spot tipping points”. He names ChatGPT’s launch as one tipping point and emergent LLM behaviours as probably another: “Both led to changes that cannot be undone, nor forgotten.” This is irreversibility language, and it echoes his earlier tipping-point and broken-symmetry framing. 3. Net judgement: mixed, but “on balance, we’re collectively in a better place than we were a year ago”, though that could not have been predicted. 4. Forecast: “a cooling off of the hype and a consolidation of current capabilities … with no AGI emerging in the near future”. He hedges that complex systems throw up the unexpected, so we need agility and resilience at every scale.

Views. - On AI: profoundly transformative, with emergent properties not to be underestimated. Possible machine consciousness is treated as a serious possibility, not dismissed. - On governance: needs creativity and expertise “that goes far beyond the science and technology of AI”. - Framing: an “advanced technology transition”.

Quotes. - “our AI future cannot be left solely to AI experts” - “naive experimentation pushing us past hard-to-spot tipping points” - “Both led to changes that cannot be undone, nor forgotten.” - “There’s been a dramatic shift from talking about AI ethics to addressing AI risks.”

2023-11-15 — navigating-orphan-risks — “Navigating “orphan risks” has never been more important for tech companies — especially around AI”#

Provenance. His own prose. A short 2023 introduction is followed by a repost of his 2018 article on orphan risks and tech startups, “pretty much as it was published back then, with a couple of editorial tweaks”. It includes the ASU Risk Innovation Nexus graphic of eighteen orphan risks. The 2023 footnotes point to riskinnovation.org, his earlier Substack posts, and a Nature Nanotechnology commentary on risk innovation. Treat the body as 2018 thinking that he explicitly re-endorsed and redirected at AI in 2023.

Argument (2023 frame). In 2018 the focus was not explicitly AI. Over the past year, though, AI researchers, developers and policymakers have needed to learn the concept. His trigger is a tweet by Ross Taylor, first author of the Galactica paper, on the model’s failed launch. Maynard calls it “a classic case of a ground-breaking idea with substantial potential that was scuppered by a lack of awareness of a deeply complex risk landscape”. He calls Galactica a “Google” model; it was in fact Meta’s. He frames orphan risks as risks “often ignored because they’re hard to quantify, and yet have a habit of stymying tech companied that are technologically creative yet socially naive.”

Argument (2018 body). - Where orphan risks sit. Rumsfeld’s known knowns and unknowns miss a category: risks that are ““known knowns” if you’re looking in the right place, but aren’t taken as seriously as they should be”. - Definition. Orphan risks are “risks that are perceived as being too ill-defined, too complex, or too irrelevant to be worth paying attention to, yet have the power to derail entire enterprises”. Examples include ethically ambivalent behaviour, threats to deeply held values, and technologies perceived as carrying unusual and uncertain dangers. They are mostly social risks that resist quantification. Uber and Facebook are his cases, and he quotes Kara Swisher on Silicon Valley needing “adult supervision”. - Old tools, new technologies. AI, gene editing and advanced manufacturing, combined with global interconnection, are transforming the risk landscape beyond what conventional tools handle. - Risk innovation. Defined as the understanding that “we need parallel innovation in how we think about and act on risk”. - Risk as a threat to value. Admittedly subjective, but it lets people discuss hard-to-quantify threats “to autonomy, dignity, trustworthiness, self-esteem, way of life, and even deeply held beliefs and convictions”. It “extends conventional thinking rather than replacing it”: health, wealth and environment still fit in the value bucket. It also brings in “the reciprocal dangers of threatening what is important to others through what they do”. - The case for attending to orphan risks. Ethics and commercial self-interest point the same way. Orphan risks include discrimination, threats to equity and livelihoods, and loss of public trust. “Few of these orphan risks are overtly governed by laws and regulations”, yet ignoring them raises the chance of failure.

Audience and framing. The frame is enterprise-facing: startups, investors, tech companies. Attending to social risk is presented as competitive advantage and survival, not just compliance or ethics. Governance here is soft and voluntary: tools and a mindset, not regulation. He cites Mitchell Baker (Mozilla) on technologists’ blind spots that will “come back to bite us.”

Quotes. - “risks that are perceived as being too ill-defined, too complex, or too irrelevant to be worth paying attention to, yet have the power to derail entire enterprises” - “we need parallel innovation in how we think about and act on risk” - “This way of thinking about risk extends conventional thinking rather than replacing it” - “Few of these orphan risks are overtly governed by laws and regulations.”

2023-11-21 — ai-and-risk-innovation — “Could OpenAI have benefitted from this tool for navigating complex risks?”#

Provenance. Mixed. - His prose: the introduction, the explanation of the Risk Innovation Planner and the value/values distinction, his definition of “community”, his brief evaluative comments between steps, and the closing “So what can be learned from this?” section. - Not his: everything presented as ChatGPT output, which is most of the body. That covers the value lists for enterprise, investors, customers and community; the seven orphan risks; the ~50-word threat summary; the three next steps; the imagined three-month reflection; and the “board’s” assessment of the exercise. ChatGPT role-played the board of a hypothetical company closely modelled on OpenAI, primed with anonymised OpenAI documents. None of this output is evidence of his views. Only his judgements on it count, for example that the value areas are “not bad” and the risk list “quite insightful”.

Argument (his parts). OpenAI’s board crisis prompted him to ask whether the Nexus’s two-page Risk Innovation Planner might have helped. - The tool. It grew out of risk innovation. The key move is “reframing risk as threat to value”, covering both future value an organisation wants to create and existing value it wants to protect. - Value versus values. This distinction is newly explicit here. Value means worth (“products, ideas, and processes, that have worth”). Values are what is considered good or bad. Values often underpin value, “But values are not the same as value — a nuance that is often missed”. ChatGPT also stumbled on it. - Neutrality. The framing is “agnostic to particular worldviews, ideologies, or ethics”, yet “it forces a company to think about how its actions may threaten what is of value to others”. Others’ sense of threat shapes the risks the enterprise faces, which it can then remove, circumnavigate or “strategically absorb”. - Stakeholders. Four groups: enterprise, investors, customers and community. His definition of community is broad: users, people disadvantaged by not using the product, those who gain or lose from wider societal effects, the marginalised, and people affected by bias. - Design brief. What could be done in 30 minutes with a startup founder to build a risk innovation mindset “that provided them with a competitive advantage”? “What the Planner does not do is provide answers to problems.” It prompts reflection on goals, hidden barriers and simple actions: three actions for the next quarter, then a reflection after three months. People reframe the eighteen orphan risks in their own terms, which he calls “appropriate”.

His conclusions. The results “weren’t unexpected to me — this is the bread and butter of what I do”. What surprised him was how well ChatGPT drew out insights. The Planner is “a catalyst for helping people and organizations who are highly time constrained and lack breadth of context or expertise”. Whether it would have helped OpenAI is “impossible to tell”, though it “might have taken the edge off some of the messiness”. His “bigger takeaway”: it is a tool for AI-development challenges “that previous experience in tech innovation and formal business and risk management training simply do not prepare people for.”

Notes. - AI companies are treated as enterprises whose governance failures are orphan-risk failures: mission drift, employee sentiment, novel governance structures. - The post is also an experiment in using an LLM as a simulated organisation or thinking partner. He finds this “more illuminating than I expected”, which fits the ai-in-scholarship-writing thread. - The tone toward OpenAI is constructive and consultative, not adversarial.

Quotes. - “But values are not the same as value — a nuance that is often missed” - “It’s agnostic to particular worldviews, ideologies, or ethics.” - “it forces a company to think about how its actions may threaten what is of value to others”

MEDIUM#

2023-10-25 — 10-million-for-ai-safety-research — “$10 million for AI safety research”#

Provenance. His own prose. It includes one quote from his co-authored Nature Nanotechnology commentary with Sean Dudley (co-authored prose). He also mentions the managing-ai-risks.com statement by Bengio, Hinton, Russell, Harari, Kahneman and others, calling its authors “luminaries” without comment.

Argument. He welcomes the Frontier Model Forum’s $10m AI Safety Fund; the Forum was founded by OpenAI, Google, Microsoft and Anthropic. He explains the terms “foundation models” and “frontier models”, the latter meaning models with potentially dangerous capabilities. He quotes the co-authored line that AI conversations are “still being driven by technological experts with little understanding of the complexity of the social, economic and political landscape”. He hopes the fund narrows that gap, but warns of “a very real danger of the Frontier Model Forum attempting to address unconventional challenges through rather conventional thinking”. The danger lies in projects that “prioritize technical expertise, build on outmoded models of risk management”, or follow disciplinary norms.

Bold technologies demand equally bold safety thinking. He wants work on advanced technology transitions that goes beyond linear progress and disciplinary silos, and names what he would like funded: - projects that “explore novel conceptualizations of risk”; - work at the nexus of arts, humanities and emerging capabilities; - radically reimagined public participation in responsible innovation; - blended models of thought leadership and knowledge mobilisation.

He closes: “Here’s hoping … this opportunity won’t be yet another attempt to be radically transformative in very conventional ways.”

Views. He is cautiously positive about an industry-led safety fund and sceptical of technocratic, discipline-bound “safety”. “Safety” in his sense includes risk conceptualisation, public engagement and the humanities.

Quotes. - “attempting to address unconventional challenges through rather conventional thinking” - “build on outmoded models of risk management” - “yet another attempt to be radically transformative in very conventional ways”

2023-11-18 — sam-altman-openai-impacts — “What could Sam Altman’s departure from OpenAI mean for societally beneficial AI?”#

Provenance. His own prose, with updates on 20 and 22 November. It quotes the OpenAI board, Jeremy Howard and Ilya Sutskever (not his words).

Argument. He looks at the ouster through three intersecting frames: 1. Technology. Little effect, because Claude, Llama, open models and university work mean OpenAI “isn’t the only game in town”. The change may show up instead in the “ethos” of development and the speed of commercialisation. 2. Governance. Despite pushes for open source, the field has consolidated around closed, closely regulated frontier models. “leading companies — OpenAI amongst them — have had an outsized influence in guiding the framing of regulations that seemingly favor commercial leaders in the field”. Whether policymakers listened to the company or the man is unclear: “My suspicion is that Altman was the “shiny person” they were interested in.” The shake-up may open “a sliver of a window of opportunity”. 3. Society. This is where he expects the biggest effects. He uses STS language: AI is “part of a complex socio-technical system”, and “AI isn’t just about the technology, but about the society in which it’s developed”. OpenAI’s social impact “isn’t commensurate with technological breakthroughs alone”. Educators restructured teaching. Universities scrambled, “driven by a hype and promise filled” revolution. Competitors raced “often far faster than a measured and responsible approach would suggest is wise”. ChatGPT’s launch “was a social and commercial as much as a technological step”.

A reset might be good: “a chance to take a breath (a pause even)”. Or it might send plans into a tailspin. The updates stress long-term ripple effects on responsible AI.

Views on companies and leaders. He is critical of industry influence on regulation and of the competitive race. He is neutral to mildly sceptical about Altman personally. He hopes OpenAI’s charter commitments (broad benefit, long-term safety, cooperation) will matter.

Quotes. - “leading companies — OpenAI amongst them — have had an outsized influence in guiding the framing of regulations that seemingly favor commercial leaders in the field” - “My suspicion is that Altman was the “shiny person” they were interested in.” - “often far faster than a measured and responsible approach would suggest is wise”

2023-10-24 — flattening-the-learning-distribution-curve — “Flattening the learning distribution curve using ChatGPT”#

Provenance. His own prose. The diagrams are his; the lead image is from Midjourney.

Argument. In one-to-many teaching, especially large undergraduate classes, educators teach “to the mean”, and students at both edges of the learning distribution are under-served. One response is to exclude the edges through selective enrolment: “This is not an approach that sits well with me as an educator”. It also conflicts with ASU’s charter of measuring success by “whom we include”. Broadening the curve through differentiated teaching only goes so far. Generative AI could instead flatten it by raising success at the edges.

His own classes are the evidence. In his prompt-engineering course, students hold individual conversations with ChatGPT on bias and failure modes, and Socratic dialogues with it. He turns inaccuracy into a feature: “Of course the absolute accuracy of ChatGPT’s responses cannot be guaranteed. But this adds to the personalized learning environment”, because students must question and test what they hear. Students also use it as an in-class “ideas-translator” and to decode poorly written assignments. The limits are “the imagination of educators” and institutions’ willingness to experiment. Caveats are “bias, persuasive incorrectness, and a whole host of intellectual property issues.” He is enthusiastic (“game changer”) while admitting early doubt (“is this really a thing?”).

Notes. This is a justice and inclusion frame for AI in education, and a positive view of AI’s cognitive role: a personalised interlocutor that builds critical questioning. There is no concern here about cognitive offloading or formation. “Persuasive incorrectness” is an early, compact name for a risk in epistemic AI.

Quotes. - “This is not an approach that sits well with me as an educator.” - “bias, persuasive incorrectness, and a whole host of intellectual property issues”

2023-11-09 — waymo-safety-study-shows-benefits — “Waymo safety study shows not all self-driving cars are created equal”#

Provenance. His own prose, including his own back-of-envelope analysis of NHTSA and FHWA 2021 data, and a disclosure footnote.

Argument. Against the Cruise crisis (a pedestrian dragged, California’s licence suspension, the fleet recall, Gary Marcus’s “Theranos of AI” question), he reads the Waymo–Swiss Re claims study: - zero bodily-injury claims over 3.87 million driverless miles, against 1.11 per million for humans; - 0.78 property-damage claims per million miles, against 3.26 for humans.

His method: - He trusts the data because the reinsurer’s business depends on accurate risk analysis, an argument from institutional incentives. - He insists “anecdote is no substance for data” (sic), despite his own positive rides. - He tests the study’s weak point himself: the human baseline includes interstates, where Waymo does not drive. His calculation shows non-interstate crash rates about three times higher per mile, so the reported benefit is probably understated.

He then leaves the normative question open: “how safe self-driving cars should be, and whether comparing them to humans is the right metric.” He also names a normalised risk: “how we’ve normalized loss of life, limb, and property resulting from inattentive and poorly trained humans being allowed behind the wheel”. He praises Waymo’s “very cautious, slow-and-steady, and responsible approach”, and concludes that “not all technologies — or companies — are created equal.” The footnote discloses Waymo gifts (“a pair of socks”).

Notes. This is the risk scientist at work: comparative, evidence-based risk assessment, checking baselines, trusting data over anecdote. The post also shows his view that corporate conduct, meaning cautious and responsible deployment, is what separates outcomes within one technology. Here autonomous AI is framed as a potential risk reducer relative to the human status quo.

Quotes. - “anecdote is no substance for data” - “how safe self-driving cars should be, and whether comparing them to humans is the right metric” - “not all technologies — or companies — are created equal”

2023-11-11 — a-fist-look-at-openais-gpts — “A first look at OpenAI’s new customizable versions of ChatGPT”#

Provenance. His own prose. It includes screenshots, links to shared ChatGPT conversations (not reproduced in the text), and quoted instructions he wrote for his GPTs.

Argument. This is a hands-on review of custom GPTs. They are very easy to build and impressive at first, but “they are often not as impressive as they first seem”; good ones need a lot of tweaking. His builds: - “Imagination Catalyst”; - tutors for “Learn about Responsible Innovation” and “Learn about Advanced Technology Transitions”. He doubts the latter works well, because little source material exists and the model must “infer and interpolate”; - book GPTs for Films from the Future and Future Rising, which kept drifting off-topic. His “talk to me about peanuts” test caught this.

His conclusion is about language as the new programming medium: coding is becoming “something that anyone with a good grasp of language and how to communicate can do”, and “that is a game change.”

Notes. This fits his view of LLMs as language-mediated tools that change who can build technology. The post is enthusiastic, practical and education-oriented, with no risk analysis.

Quote. - “something that anyone with a good grasp of language and how to communicate can do”

2023-11-05 — dont-panic-elon-musk-launches-grok — “Don’t Panic: Elon Musk launches a new AI based on the Hitch-Hikers Guide to the Galaxy”#

Provenance. His own prose. It quotes xAI’s description and passages from Douglas Adams’s radio series (not his).

Argument. Days after talking AI risk with Rishi Sunak, Musk launches Grok, modelled on the Hitch-Hiker’s Guide. Using Adams’s own descriptions of the Guide as “apocryphal” and “wildly inaccurate”, Maynard notes wryly that “a guide that leans more toward what you want to hear than what is, strictly speaking, true” may be “par for the course for AI coming out of the House of Elon”. He sees “a palpable tension” between the rhetoric of responsible foundation-model development and a product inspired by something “anything but responsible”. He is even-handed: Adams holds up a mirror to human absurdity, and Grok may live up to Heinlein’s “grok” of deep intuitive understanding, “for all the hype and hubris”.

Notes. He is critical of Musk’s mix of safety rhetoric and flippant product design, and there is an early hint of concern about AI that tells people what they want to hear. He uses science fiction as a cultural lens.

Quotes. - “a guide that leans more toward what you want to hear than what is, strictly speaking, true” - “for all the hype and hubris”

LOW / NONE#

• 2023-10-27 — ai-and-consciousness — “Exploring the cutting edge of AI and consciousness” (low): A short recap and link for an ASU Future of Being Human … Unplugged livestream on AI and consciousness with Mark Daley and Blake Richards, plus a notice of the next event on “Living with your AI Symbiont”. His prose, but no substantive argument.
• 2023-10-29 — category-confusion-complicates-efforts — “Category Confusion Complicates Efforts to Regulate AI” (low; guest post by Brad Allenby, not evidence of Maynard’s thinking). Allenby argues that AI works at three levels: enabling infrastructure, infrastructure in itself, and a subsystem of an emerging “cognitive ecosystem”. Regulation that conflates these levels risks dysfunction, and “Regulation by hypothesis is a dangerous game”. Only the subtitle is plausibly Maynard’s. Note that he chose to publish it and links it approvingly in his 2023-10-30 White House post as evidence of poor awareness of “integrated sociotechnical systems”.
• 2023-10-30 — time-specific-actions-on-ai — “Time-specific actions in the White House AI executive order” (low): A short intro by him, then a timeline of EO deadlines produced by Anthropic’s Claude. He flags that “any information below should be cross-checked”. The list is AI-generated and not evidence; the post shows his practice of using LLMs openly for policy summarisation.
• 2023-11-06 — an-ipcc-for-ai-is-a-failure-mode — “An IPCC For AI Is A Failure Mode” (low; guest post by Brad Allenby). Allenby argues that an IPCC-style body for AI would fail. AI changes faster than institutional cycles, is unknowable in a way climate is not, is geopolitically and commercially entangled, and is “post-Westphalian”. He proposes informal, networked, real-time working groups focused on “pure perception” of what AI in the wild is doing. None of this is Maynard’s own prose; he published it and presumably wrote the subtitle.
• 2023-11-08 — human-ai-symbiosis — “Is deep and meaningful human-AI cooperation possible?” (low): An event notice and recording link for an Unplugged conversation with Athena Aktipis and Paul Rainey on human-AI symbiosis. He frames it as asking whether symbiosis “should excite us or have us deeply worried”, with bios.
• 2023-12-01 — dec-19-future-of-being-human-unplugged — “What happens when AI and cooperation science collide?” (low): An event registration notice for an Unplugged conversation with Mark Daley and Athena Aktipis on AI and cooperation science, with bios. No argument.