Late Lessons, Jensen Huang and AI
Markdown version

B25 digest: 2025-04-13 to 2025-05-25#

What this batch is#

Spring 2025: Colossal’s “dire wolves”, US and Chinese plans for AI in schools, agentic AI (Manus), and the end of the ASU teaching year. Three of the ten posts are Modem Futura episode posts, set aside per the user’s instruction. The Substack anniversary post is low relevance. The other six are his own prose, with caveats: the de-extinction post is mostly a republished 2018 Films from the Future chapter; the US–China post’s Deep Research report is not used; the parasocial essay quotes ChatGPT o3 as “reviewer 2”.

The batch is less about AI risk as such than about the tools and dispositions he thinks responsible navigation needs: frameworks for characterising agents, heuristics for anticipating consequences, a culture of care, and relational communication between experts and publics.

Main ideas#

1. Responsible AI as a mindset of care, grounded in risk as a threat to value. The Responsible AI Trajectories Tool (05-18) is the batch’s most developed original contribution: six cause–effect models (linear, S-curve, exponential, hysteresis, jagged, chaotic) for anticipating AI’s consequences. The aim is formation of judgement, “fostering a responsible use mindset rather than simply focusing on facts and figures”. The non-linear models carry the moral weight: - hysteresis: sticky effects that persist when the cause is removed (lost research depth; workplace paranoia); - jagged: uneven adoption and unevenly distributed risk (an AI sentencing tool); - chaotic: tipping points and irreversible harm, led by emotional dependency on an AI companion bot and, at scale, threats to grids, finance and “societal cohesion”.

He redefines “effect” as “what we consider to be of value — or what we care for” and footnotes Risk Innovation. This carries his long-standing “risk as a threat to value” into AI literacy teaching, joined to a vocabulary of “care”. Responsibility belongs to everyone affected. Responsible AI is a “wicked problem” in the strict sense. He admits the frame can seem “too cut and dried”.

2. Agentic AI is new in kind, and governance must first learn to describe it. In 05-04 he says we have “never had the ability to create machines that can decide on their own how to solve problems” and act unsupervised. The first obstacle is conceptual: “we’re not even sure yet how to formulate the problem”. His own definition of an agent includes acting on “behavioral” and “social” environments. He backs Kasirzadeh and Gabriel’s four dimensions (autonomy, efficacy, goal complexity, generality) as a way to fit oversight to the technology rather than regulate it one-size-fits-all. He adds two things: - the framework has no place for AI’s “direct causal effects on the beliefs, understanding, and behaviors of individuals and groups”; - risk probably scales “closer to exponential than linear”.

3. Hubris, power and responsibility in commercial deep tech. Colossal’s hubris is “palpable” (04-13), yet it “may be onto something — at least technologically” as gene editing merges with AI. AI companies are his reference point for being “full of themselves”. The republished 2018 chapter restates commitments he still holds: - could vs should; - responsible science as humility and listening to other expertise; - normal accidents and bounded chaos (the Arkema chemical-plant cascade); - power used responsibly, not abdicated; - past technologies (Industrial Revolution, the bomb) as “just a rehearsal”.

4. Governance under uncertainty: resilience over speed. On US vs China K-12 AI policy (04-24), he is even-handed. China’s centralised model is fast but may be “fragile”. The US distributed model risks leaving students “behind” but offers “far greater resilience under uncertainty”. He faults the US order for leaving ethics implicit, sees “win-win” collaboration, and doubts near-term AGI.

5. Expertise and publics. The parasocial essay (05-25) is a significant statement of his theory of engagement: - The deficit model persists in how experts are trained, though it has “been repeatedly shown not to be effective”. - Relationships, including parasocial ones, shape decisions (Trump’s 2024 podcasts). - The same mechanism can serve “widespread social manipulation and control” or empowerment. - He sets out four purposes of expert communication (instruction, ego, impact, empowerment) and argues that parasocial communication can join scale to impact in “complex technology transitions”.

He does not link this to AI companions here, though it sits naturally beside (1) and his 2024 work on AI social influence.

6. Futures and being human. The course trailer (05-11) restates his framing questions (“What are the technologies we want in our lives?”) and warns “we risk losing everything” through technologies “we do not understand, and cannot handle”.

7. AI in his own practice. He uses AI throughout: apps built with AI, “vibe coding”, reader buttons, a “developed by AI, checked by humans” workflow, and ChatGPT o3 as a sparring reviewer that he corrects.

New or changed#

Concepts in this batch#

Responsible AI Trajectories Tool; six cause–effect models (linear, S-curve, exponential, hysteresis, jagged, chaotic); tipping points; runaway impacts; sticky or irreversible effects; effect as threat to (or enhancement of) value; culture and mindset of care; responsible-use mindset; wicked problem (strict sense); responsibility shared by all stakeholders; AI agent (his definition); autonomy, efficacy, goal complexity and generality; the “human dimension”; formulating the governance problem; exponential risk scaling; commercial evolution (Colossal’s term, examined); biology by design; could vs should; humility and listening; normal accidents; bounded chaos; responsible use of power; resilience vs fragility in governance; public value beyond economic growth; deficit model; transactional, relational and relationship-based communication; parasocial communication; the four purposes (instruction, ego, impact, empowerment); empowerment “on their own terms”; “developed by AI, checked by humans”.

Most important posts#

  1. 2025-05-18 exploring-ai-through-cause-and-effect: his fullest statement of responsible AI in this period. It covers non-linear and irreversible effects, care, and risk as a threat to value applied to AI.
  2. 2025-05-04 an-important-new-model-for-guiding-agentic-ai-oversight: what is new about agentic AI, his definition of an agent, framework-first governance, and the gap around epistemic influence.
  3. 2025-05-25 why-parasocial-communication-is-important: the deficit-model critique and relational and parasocial communication. This is his theory of expertise and publics in technology transitions.
  4. 2025-04-13 de-extinction-conservation-futures: hubris, power and responsibility, AI–biotech convergence, and AI companies as a reference point, plus his reaffirmed 2018 framework.
  5. 2025-04-24 us-and-china-vie-for-ai-k-12-leadership: governance as resilience under uncertainty, and near-term AGI scepticism.