B31 digest: 2026-07-16 to 2026-09-20 (9 posts)#

What this batch is#

Summer 2026, during his sabbatical. Four threads: his risk innovation thinking applied in scholarly form to frontier AI companies; a self-questioning experiment in AI-written scholarship; a turn toward play and “learning by not trying to learn”; and two public pieces, on universities in the “AI transition” and on whether AI will “kill us all”. He also announces his move to ASU’s Thunderbird School of Global Management. None of the posts is a Modem Futura podcast post.

Provenance matters more than usual. He rewrote the 2026-07-16 orphan-risks paper himself, but says three days later that the “ideas, analysis and insights that Fable generated remain intact”; its AI use statement claims the “argument architecture” as his. Fable designed the Hyperbubble game, wrote most of mull.chat’s documents, and is sole author of the constitutional AI paper announced on 2026-09-04. Only his framing and judgements count as evidence.

Main ideas#

1. How AI companies choose their risks (2026-07-16). Developers keep several accounts of AI risk: narrow, capability-based safety frameworks; broader compliance frameworks that law requires; broader still, securities filings. The case study is persuasion, which OpenAI dropped in 2025 and which came back as “harmful manipulation” only when law required it.

He explains the pattern with four filters: “Can we measure it? Is it big enough? Can we evidence it? And can we afford to keep it?” He traces the first three to a definition of risk as “the probability of a specified, severe harm event”. The frameworks are “working as designed. It’s just that the design itself may be flawed.”

His remedy joins his long-standing reframing of risk as a threat to value to new accountability tools: the “safety differential” (the gap between a firm’s safety and compliance frameworks), an “orphan-risk register”, an “aperture log”, and regulators requiring disclosure of how risks are selected rather than coverage of every risk.

He explicitly refuses a cynical reading: “sincerity almost always operates inside an incentive field”. He is “not optimistic” that regulation alone will close the gap. He also adds a justice caveat: the channels that turn stakeholder harm into cost for the firm “are not equally open to everyone”.

2. AI risk: plural, insidious and long known (2026-09-15). In response to a September 2026 wave of extinction alarm, he goes back to his 2018 Risk Bites list of ten AI risks. The list includes dependency (“harder to think for ourselves”), heuristic manipulation, value misalignment, re-writable goals and superintelligence. He says it still holds.

Risks that have risen since 2018: cybersecurity, water and energy impacts, privacy, deep fakes, systemic disruption of education and political infrastructure, frontier AI governance, children’s development, and “psychological/cognitive disruption”.

On existential risk: low probability, not to be dismissed, and to be handled “without running around like headless chickens”. He criticises developers for “acting as if they’re the first people to notice” these risks, and for saying they should go slower while not doing so. Talking about risk is not fear-mongering, and ignoring “people and institutions who know a thing or two about risk” creates risk of its own.

3. AI and writing: machines resetting human standards (2026-07-19, 2026-09-04). His view has changed: a year earlier he was “blown away” by Claude 4.5’s eloquence. Now he finds frontier-model prose “superficially profound yet substantively hollow”. His explanation: AI “cannot understand what it feels like to read as a human”. The worry that follows is a formation risk: LLM feedback loops pushing human writing toward “an LLM-view of what good writing is”. He sums it up this way: “the AIs we have trained to “think” like us are now beginning to train us to think like them.”

He still respects the substance: Fable 5.1’s paper is an “original knowledge contribution”, though “incremental and combinatorial”. He names Fable as sole author with a CRediT-AI statement, and treats any AI-assisted paper made in under 10–20 hours of human labour as “highly suspect”.

4. Play, joy and formation (2026-08-02, 2026-09-20, 2026-08-23). His sabbatical thinking: “play without purpose”, joy and serendipity are “critical skills for thriving in an age of AI”. Professional culture treats them as unserious, and learning may best happen with “no learning expectations”.

The Hyperbubble game, which he endorses as capturing how he sees the world, turns his risk worldview into mechanics: orphan risks to adopt, moral panics that grow if engaged, a “hype line” where the bubble swells until it bursts, and “flourishing” as the goal. Both fast and cautious strategies work, given “risk navigation skills”.

mull.chat parodies LLM “reasoning” streams as performance.

5. Universities and the AI transition (2026-08-30). Bill Gates’s essay aligns with his own call for agile, “boundary-transcending” approaches to the AI transition. But Gates treats universities as users, not leaders, and Maynard agrees that is what they have been: “followers and users of the technology”, “guardians of the past more than leaders toward the future”. He holds “fiercely” that universities have a responsibility to use their freedoms for public good. He is uncertain, and somewhat disillusioned, about whether they can: “Sadly, this has been my experience so far.”

Concepts appearing#

• Long-standing, restated: orphan risks; risk as threat to value; risk innovation; the Risk Innovation Planner; responsible innovation (the RRI dimensions and the Garbee “entrepreneurial culture” lesson); the cognitive Trojan horse; playground not playpen; the 2018 ten AI risks; advanced technology transitions; flourishing.
• New in this batch: the four filters; the safety differential; severity floor as a “boundary of accountability”; risk aperture; the orphan-risk register and aperture log; “your risk is my risk” and conversion channels; safety frameworks as “the de facto governance layer”; values drift (from the co-written book); cargo-cult paper; LLM-view of good writing; learning by not trying to learn; joy as a metric; CRediT-AI.
• Adopted from others: Kasirzadeh (decisive versus accumulative harms), Power (audit society), Vaughan (normalisation of deviance), Kasperson (social amplification), “cognitive surrender”.

New or changed#

• Risk innovation, fully applied. The first full scholarly application of his 2013–2020 risk innovation work to frontier AI governance, with testable predictions and a justice caveat. The 2023 versions were short posts about OpenAI.
• A change in his relationship with AI writing. Explicit disenchantment with AI prose, a new worry that AI is setting human standards, and an openly voiced (though not believed) doubt about his own writing.
• AI named as sole author. A new practice for him.
• A heavier emphasis on play and joy as method. Tied to his sabbatical and his move to a management school (leaders’ “mindsets and skills”).
• Continuity on AI risk. Anti-alarmist but serious; newly on his radar are environmental impacts, children’s development and cognitive disruption.
• Writing for AI readers. He increasingly tells readers to have an AI summarise his work (beinghuman.fyi, the text mirror).

Most important posts in this batch#

1. 2026-07-16 orphan-risks-frontier-ai-maynard. The core statement of his risk thinking applied to AI companies and governance. Read it with the provenance caveat.
2. 2026-09-15 will-ai-really-kill-us-all. His current map of the AI risk landscape and his stance on existential risk.
3. 2026-07-19 publish-or-perish-ai-vs-human-vs-human. Cognition and formation; his change of view on AI writing.
4. 2026-08-30 do-universities-have-a-place-in-bill. Universities, leadership and the AI transition.
5. 2026-08-02 what-we-can-learn-with-ai-by-not-trying-to-learn. Play and formation; his risk worldview as a game.