B22 digest: 2025-03-02 to 2025-03-27#

What this batch is#

Four weeks in March 2025. Deep Research has just arrived, Manus has opened early access, and DOGE is at work in Washington. Of the ten posts, four are Modem Futura episode notes and are set aside here, following the user’s instruction. The other six fall into three groups: - a re-reading of his 2018 critique of permissionless innovation; - a new responsible-innovation idea, the “hard” concept of care; - four posts on AI and higher education, two of them hands-on tests of the agentic AI Manus.

Two posts contain large amounts of AI text: a ChatGPT-written first-person op-ed (2025-03-16), and a Manus-written report on synthetic survey data (2025-03-22). Only his framing, footnotes and prompts count as evidence.

Main ideas#

1. Permissionless innovation, re-endorsed and sharpened (2025-03-02). Prompted by Reid Hoffman, he republishes his 2018 Films from the Future argument and says it is “more relevant now than it was then”. His objection is to fixing harms “after the fact” rather than “anticipating them and navigating around them”. The new 2025 footnotes add three things: - A test for when permission matters. Experiments in low-risk, reversible, linear systems are fine. Breaking “people, governance, society, and the planet”, which are complex systems with unpredictable and potentially catastrophic outcomes, is not. - Precaution. He places Thierer’s idea as a reaction against the precautionary principle, and calls the transatlantic history of precaution “fraught with misunderstanding, misinterpretation” and trade-barrier accusations. - Musk. He retracts his 2018 view of Musk as “naive” and names DOGE as permissionless innovation applied to government.

The 2018 core stays in place: the single innovator trapped in Plato’s Cave, technologies of hubris, irreversibility rising from the Industrial Revolution to the “nuclear and digital age”, and “checks and balances around who gets to do what”. So does its candour about the benefits of speed and his own lab-bench rule-breaking.

2. The “hard” concept of care (2025-03-09). This is a new node in his responsible-innovation thinking. It draws on Alison Gopnik and on his ASU colleague Emma Frow’s work on care in synthetic biology. He coins “hard” care in two senses: - care as “something of substance” that can ground “policy, governance, and decision-making”, as opposed to soft, touchy-feely care; - “hard” as in difficult.

In his own commentary, care stands against control, which reduces agility and responsiveness. It stands against performative box-ticking. It attends to the vulnerable rather than to aggregate benefit, and he criticises people with “little time for the individuals who make up that aggregate”. He endorses Tronto’s “privileged irresponsibility”. Care also links back to hubris: we overlook what sustains well-being because we are “too arrogant”. His conclusion is that we cannot think about human–technology relationships “without placing the hard concept of care at the very heart”. The detailed definition he ends on is Deep Research’s, not his.

3. AI is categorically different because it simulates what defines us (2025-03-15). In education he states his strongest nature-of-AI claim in this batch. Models increasingly simulate “the ability to think, to reason, and to solve problems with agency”, so they “stand apart from pretty much any previous technology”. Treating them as a learning aid is “a categorical error”. AI is changing “who we are”. This grounds his playpen-to-playground argument: - give students access and “permission to play” under norms that avoid “unacceptable harm”; - educators become “fellow-travelers”; - the “greater danger” is holding students back.

The tension with his critique of permissionless innovation is not addressed. His own reversibility test resolves it: a curated, recoverable playground is not the same as breaking complex systems.

4. Agentic AI: “the keys to the digital kingdom” (2025-03-22, 2025-03-27). Testing Manus leads to his first sustained reflections on general AI agents: - hierarchical agents on their own virtual machine are “game-changing”; - they infer intent and change goals “without asking for permission first”; - “I’m not sure how ready most people are for machines that decide for themselves what we actually want”

He expects the move “from clunky gimmick to deeply disruptive technology in a matter of months”. He foresees agents completing whole courses for students, making ChatGPT-cheating worries “quaint anachronisms”. He welcomes cheap, agent-built courses as a possible “revolution in how knowledge and training flow through society”, bad news only for those treating knowledge “as a commodity”. He is candid about unreliability and failed replications. Missing: any governance response, and any concern about privacy or data. He treats Manus’s ethical refusal lightly and works around it. He warns only implicitly that fabricated research is persuasive.

5. AI in his own scholarship and writing (2025-03-09, 2025-03-16). Deep Research is “a game changer” as “a catalyst to human-initiated thinking and research, rather than as a substitute”. It still “lacks the professional intuition of a good polymathic scholar”. Letting ChatGPT write as him makes him uneasy, because “so much of my professional and academic identity is embedded in my writing”. His footnotes note that the AI strips out his caveats and turns his support for ASU into PR.

Concepts in this batch#

Permissionless innovation (critiqued). The lure of permissionless innovation. Technologies of hubris. The single-innovator problem and Plato’s Cave. The paradox of innovation (speed against caution). Rising irreversibility across eras. Reversibility and system complexity as the test for permission. Precaution as a politically distorted debate. The “hard” concept of care. Care against control. Privileged irresponsibility. Aggregate against individual benefit. Technologies as “monsters” needing stewardship (via Frankenstein). AI as catalyst, not substitute. Playpens against playgrounds. AI as categorically different, simulating defining human capacities. The “categorical error” of AI as a learning aid. General AI agents and “keys to the digital kingdom”. Machines deciding what we want. Human-less human-subjects research. Navigating advanced technology transitions. Knowledge as commodity against democratised access. Authorial identity and AI ghostwriting.

What is new or changed#

• Explicit revision of Musk. He moves from humble AI-risk voice (2018) to naive permissionless innovator (DOGE). The 2018 claim that AI was only narrowly smart is implicitly withdrawn, while the governance argument is kept and strengthened.
• A reversibility and complexity criterion is stated more crisply than before, as the line between healthy experimentation and dangerous permissionlessness.
• “Hard” care is new. It extends responsible innovation towards an operational, justice-oriented ethic, and he developed it using Deep Research as a catalyst.
• Agentic AI becomes a hands-on focus. His time horizon for disruption shortens to “months”.
• Education framing escalates from ChatGPT and cheating to agents doing whole courses, and from AI as a tool to AI as a challenge to human identity.
• He publishes AI text written in his own voice, with open ambivalence.

Most important posts#

1. 2025-03-02 the-lure-of-permissionless-innovation: his core governance stance on after-the-fact fixes, irreversibility and hubris, plus the Musk revision and the precaution footnote.
2. 2025-03-09 the-hard-concept-of-care-in-technology-innovation: a new responsible-innovation concept with governance ambitions.
3. 2025-03-15 ai-playgrounds-in-higher-education: his clearest statement in this batch on AI’s nature and on formation.
4. 2025-03-22 when-agentic-ai-takes-charge-manus: his first sustained take on agentic AI and the handing-over of intent to machines.
5. 2025-03-27 ai-agent-creates-online-course-in-minutes: agentic AI, access to knowledge, and the teaching of technology transitions.