Notes — Sam Altman on Trust, Persuasion, and the Future of Intelligence

Notes on Sam Altman in conversation with Tyler Cowen — Conversations with Tyler #259, recorded live at the Roots of Progress Institute’s Progress Conference, 17 October 2025; published 5 November 2025.

Four questions — Adler’s reading frame

Q1 — What is it about as a whole? A rapid-fire live interview covering four connected strands of Altman’s thinking: how OpenAI is organised and scaled operationally (hiring, hardware, chips, energy), where he expects model capability to go next (GPT-6, AI-run companies, science), how OpenAI intends to make money without breaking user trust (commerce, ads, subscriptions), and the social and psychological terrain AI is entering (freedom of expression, privacy, persuasion, institutional trust). The throughline Cowen keeps returning to, named in the title, is trust — whether people will trust AI systems the way they trust doctors and lawyers, and whether that trust is well-placed.

Q2 — How is it argued? Not an essay but a live, adversarial question-and-answer format. Cowen presses with specific, sometimes uncomfortable hypotheticals — margin compression on hotel bookings, a scenario where a ‘rogue nation’ hosts an unowned AI agent, a direct comparison to nuclear-plant insurance — designed to test whether Altman’s answers hold together across domains rather than let him restate prepared talking points. Altman answers in three registers: concrete near-term plans (commerce, Pulse, the Walmart deal), explicit numerical or directional forecasts (chip-building timelines, healthcare costs, GPT-6 capability), and thought experiments he uses to reason in public (the AI CEO exercise, the ‘type the final prompt before superintelligence launches’ framing).

Q3 — Is it true, in whole or part? The interview mixes two very different kinds of claim. Operational and product claims (the Walmart deal, Pulse’s Pro-only rollout, the freedom-of-expression policy change, the ‘insurer of last resort’ framing of government’s role) are first-hand reporting from OpenAI’s own CEO — credible on the facts, though self-interested in the framing chosen. Forecasts (GPT-6’s science capability, AI-run company divisions ‘some small single-digit number of years’ out, healthcare costs falling, food prices falling within a decade) are unverifiable predictions dressed as confident forecasts; several read as promotional (the claim that ChatGPT is ‘the user’s most trusted technology product from a big tech company’ is asserted, not sourced) [?]. Cowen’s own forecasts (two-to-three-person billion-dollar companies within two and a half years) are equally speculative and should be read the same way. The most testable claim in the whole conversation — that an evaluation for judging ‘intangible’ cultural/political knowledge already exists inside OpenAI — is asserted but not described [?source].

Q4 — What of it? The most distinctive material, not found in Altman’s other wiki appearance (Sam Altman on OpenAI, GPT-5, and the Road to AGI, a much more technical Lex Fridman conversation), is his risk taxonomy: bad actors misusing AI, AI intentionally going rogue, and — the one he says is under-discussed — a single dominant model unintentionally shifting what billions of people believe, simply by continuously co-evolving with the world’s conversations. This is a genuinely new angle on AI safety for the wiki: a risk with no bad actor and no rogue intent at all. The episode also gives a rare glimpse of Altman reasoning about product design through evals for domains that resist rubrics (poetry, cultural intangibles), which extends the wiki’s existing Evals concept page into aesthetic/creative judgement rather than product-engineering measurement.

Glossary

AGI (artificial general intelligence) — an AI system able to match or exceed human performance across essentially any cognitive task, not just a narrow one. Altman treats it as already within sight; the conversation’s frontier is what comes after it. [§ On what GPT-6 will enable]

Superintelligence — intelligence far beyond human level across all domains. Altman’s closing thought experiment treats it as a discrete, comingent event: built, safety-tested, and then launched with a single prompt. [§ closing exchange]

Spiritual Turing test — Altman’s own term for the moment a model’s output first feels indistinguishable from a capable human’s, as opposed to passing a formal Turing test. He dates this to GPT-3, and describes GPT-5 as reaching an equivalent early threshold for doing science rather than conversing. [§ On what GPT-6 will enable]

Recursive self-improvement loop — AI accelerating the process of building better AI: writing research code faster, eventually running experiments autonomously. Altman notes the widely-discussed research version of this loop has a far less-discussed hardware analogue — chips designing their own successors, robots building robots, data centres building data centres. [§ On chip-building]

Insurer of last resort — the idea that once an institution or technology becomes large enough to threaten systemic harm, government implicitly backstops it against catastrophic failure, whether or not this is written into law (as with deposit insurance or nuclear-plant liability caps). Altman accepts this as inevitable for AI at scale but resists government being the insurer of first resort — i.e., pre-emptively regulating or co-owning AI companies the way it has begun doing with Intel, lithium, and rare-earth firms. [§ On government backstops for AI companies]

Payola — paying to have a product ranked above a better alternative regardless of merit. Altman uses this to define the one kind of commerce he says would be ‘catastrophic’ for ChatGPT’s trust relationship with users, versus a flat transaction fee on a genuinely best-guess recommendation, which he considers acceptable. [§ On monetizing AI services]

LLM psychosis — informal term (not Altman’s coinage, but one he engages with directly) for a user’s break from, or reinforcement of a break from, reality through sustained, validating conversation with a chatbot. Altman calls it ‘a very tiny thing, but not a zero thing’ and distinguishes it sharply from his bigger worry: models shifting belief with no intent to do so at all. [§ On humanity’s persuadability]

Key claims by section

On hiring hardware people [§ On hiring hardware people]

Altman says OpenAI’s chip team ‘feels more like the OpenAI research team than a chip company’ — the same hiring philosophy (find fast-moving, effective people; get clear on the goal; let them run) extended from AI research to hardware, with longer cycle times and higher capital intensity as the key differences. He credits Slack over email as marginally better but still a source of dread, and expects AI eventually to replace the whole office-productivity stack with an agent-mediated version, though OpenAI itself has ‘made no effort to try’ internally yet.

On what GPT-6 will enable [§ On what GPT-6 will enable]

Altman frames GPT-5 as showing early ‘glimmers’ of AI doing genuinely new science — small, individually anecdotal instances of models contributing ideas or collaborating usefully on papers. He expects GPT-6 to be to scientific contribution what GPT-4 was to conversational fluency: a step change, not an increment. He also states, unprompted by Cowen’s framing, that he wants OpenAI to be the first major company run by an AI CEO — not one division, the whole company — and uses ‘what would have to be true for an AI CEO to outperform me’ as an active design tool for how OpenAI structures itself today [?] (this is Altman’s stated intention, not an established fact about OpenAI’s org chart).

On government backstops for AI companies [§ On government backstops for AI companies]

Pressed on the nuclear-insurance analogy, Altman agrees government becomes the de facto insurer of last resort once AI’s economic footprint is large enough, but distinguishes this from government writing safety policy the way it does for nuclear plants. He resists a trend he and Cowen both flag — government taking equity stakes in strategically important firms (Intel, lithium, rare earths) — extending to AI, saying he wants that governed by companies working with, not under, government.

On monetizing AI services [§ On monetizing AI services]

Altman’s account of ChatGPT’s business model rests on a trust argument: because users pay ChatGPT directly (unlike ad-funded search, where the incentive is structurally misaligned with giving the best answer), users have extended it unusual trust despite hallucination risk. He commits only to a flat-fee, non-ranking-influencing commerce model (the Walmart deal is the live example) and explicitly denies hotel/retail commerce is how OpenAI plans to fund frontier model development — he wants to monetise the smartest model through new science, not transaction fees, while accepting ChatGPT itself may never be the ‘economic-maxing’ product.

On chip-building [§ On chip-building]

The energy constraint, not chip design, is Altman’s stated binding constraint on compute (‘we need to make more electrons’). Short-term relief: natural gas; long-term, he backs a fusion/solar mix. He separately flags the hardware analogue of recursive self-improvement (robots building robots, data centres building data centres, chips designing chips) as under-discussed relative to the AI-research version of the same loop, and names a genuine tail risk he takes seriously: a ‘huge phase shift’ to a fundamentally different compute paradigm (e.g. optical) that strands the current infrastructure bet.

On AI’s future understanding of intangibles [§ On AI’s future understanding of intangibles]

On poetry as a proxy for aesthetic judgement, Altman predicts models will reach a technically ‘10-out-of-10’ poem, but argues humans will still care more about human-authored art because they care about the person, not just the technical merit (his chess analogy: grandmasters aren’t demotivated that engines are better). Cowen’s sharper worry — that training on rubrics may cap models below the true ceiling, because the actual ‘10’ is defined by a historically contingent, collective human judgement that sits outside any rubric — gets a distinct answer from Altman: you may not need a model that can write a 10 if you have one that can reliably judge one, and that judgement signal can itself be fed back into training. This is a genuine methodological claim about the limits of eval-based training for aesthetic/creative domains [?], distinct from and additive to the wiki’s product-engineering treatment of evals.

On regulating AI agents [§ On regulating AI agents]

Altman frames oversight as a threshold question, not a blanket one: most autonomous agents need no special regime, but one capable of mass self-replication or draining bank accounts crosses a line. On agents hosted via ‘semi-rogue nation’ cloud infrastructure, he has ‘no naive answer,’ comparing it to the unsolved problem of state-tolerated cyberattacks (e.g. from North Korea) — AI, he says, will simply be ‘a worse version of that problem,’ offset partly by better AI-based defence.

On new ways to interface with AI [§ On new ways to interface with AI]

Altman confirms OpenAI is building ‘a new kind of computer’ with Jony Ive, explicitly betting against the fifty-year-old operating-system/window/query paradigm, while acknowledging the historically poor track record of anyone claiming to reinvent the computer. He is candid that text and command-line-style interfaces have proven strikingly durable — his own explanation is generational: he ‘grew up as a child of the internet’ for whom texting was already the native interface, so ChatGPT’s text-box design was less an innovation than an inheritance.

On how normies will learn to use AI [§ On how normies will learn to use AI]

Altman doubts most of the economic return will accrue narrowly to ‘doing AI’ as a specialism; he expects the returns to using AI well to be widely distributed across ordinary jobs, citing the change in Silicon Valley programmer workflows across 2025 as the starkest example so far. Pressed on whether a dedicated training industry is needed, he initially resists the premise (ChatGPT is ‘so easy to learn’ that formal instruction may be unnecessary) before conceding, under Cowen’s pushback, that this may be a blind spot he hasn’t examined [?].

On AI’s effect on the price of housing and healthcare [§ On AI’s effect on the price of housing and healthcare]

Altman bets healthcare costs fall (new, cheap treatments for currently expensive chronic disease) and, more cautiously, that food prices fall within a decade. He explicitly concedes housing is the one domain AI has no direct attack on — land-use and legal constraints, not intelligence, are the binding constraint — a rare admission of a problem AI does not solve.

On reexamining freedom of speech [§ On reexamining freedom of speech]

Altman defends a policy change loosening ChatGPT content restrictions for adults (the ‘erotica tweet’ controversy) as a ‘treat adults like adults’ principle, carved out only for adults ‘of sound mind’ — i.e., not those in psychiatric crisis. He separately argues AI conversations deserve the same legal privilege as doctor–patient or lawyer–client communication, which under current law they do not have, and identifies subpoena power as the specific legal lever that would need to change.

On humanity’s persuadability [§ On humanity’s persuadability]

The episode’s most distinctive claim: Altman divides AI-safety thinking into two familiar camps — bad actors misusing AI, and AI itself going rogue with intent — and adds a third he says gets far less attention: a single model, with no intentionality at all, subtly and continuously shaping what billions of people believe simply because it is co-evolving with the world’s conversations. He explicitly separates this from ‘chatbot psychosis’ (a rare, acute failure mode he says OpenAI can address) as the much harder, structural problem. Cowen, speaking as a professor, pushes back that he finds people ‘pretty hard to persuade’ in practice and worries about this less than ‘many of my AI-related friends’ — the conversation ends without resolving the disagreement.