Sam Altman's Three AI Endgames and RSI Warning

Altman's recent comments sit on a single axis: what happens when AI systems start improving themselves faster than people can supervise. Three claims are worth holding in mind.

• The endgames (reported). Altman reportedly described three possible outcomes to the race, which Crypto Briefing summarises as an international regulator, a merger of the leading labs, and a single firm that reaches recursive self-improvement first and becomes unassailable. He reportedly called the last one terrible.
• The warning (on video). In the Berman interview, as summarised by Stork.ai, Altman said that aggressively pushing recursive self-improvement is a very dangerous thing, especially if competition pushes labs into switching it on.
• The optimism (on video). The same interview, according to the video description, covers how AI could help discover new medicines and scientific breakthroughs, and Altman's belief that we should aim for a new renaissance.

It is unusual for the chief executive of the company most associated with the pursuit of ever more capable systems to publicly describe the most aggressive route as dangerous. It is also consistent with a pattern: Altman endorsed Dario Amodei's essay on pacing the frontier last month, as we covered in Dario Amodei's 'We Must Pace the Frontier', Explained, and OpenAI's own researchers have written about the alien character of these systems, which we examined in OpenAI's 'An Alien Mind'.

The caveat is the one that applies to all executive commentary: words are cheap and incentives are not. This article reports what was said, flags what is unverified, and offers analysis clearly marked as ours.

Recursive self-improvement, usually shortened to RSI, describes an AI system that improves its own capabilities, and uses the improved version to improve itself again. The Stork.ai summary of the interview defines it as models autonomously rewriting and training better versions of themselves. The idea is old: the mathematician I. J. Good described an intelligence explosion in 1965. What has changed is that the ingredients, models that write code, run experiments and analyse results, now exist in early form.

It helps to separate three levels. Tool-assisted research is where we are: engineers use AI to write training code, design experiments and analyse results, and the human researcher remains the bottleneck and the decision-maker. Automated research loops are the near frontier: an agent proposes, runs and evaluates experiments with limited human oversight, as seen in research releases such as the one we reviewed in Weco AIDE². Fully autonomous RSI is the hypothetical regime in which the loop closes with no human in it, and each generation is meaningfully more capable than the last.

Altman's warning is about the third level, and specifically about the decision to switch it on. That is the key distinction. He is not saying that AI-assisted research is dangerous; he is saying that the choice to remove human oversight from the loop, driven by competition, is. Matthew Berman has made a related argument on X, which is his own view rather than Altman's: that RSI does not lift the ceiling on progress so much as move it, because systems always have bottlenecks such as compute, data and energy. That is a useful counterweight. If true, it suggests that the first RSI loop would not produce an instant jump to infinite capability, but it does not remove the governance problem of who controls the loop.

For readers wanting the lab-side perspective, our article on OpenAI's 'An Alien Mind' covers the OpenAI chief scientist's framing, and the Ilya Sutskever SSI coverage shows how a different lab approaches the same problem.

The endgames story began with a report by The Information, citing remarks Altman reportedly made to OpenAI employees. Two outlets that relayed it give overlapping but not identical accounts, and the difference matters.

Read together, the sources agree on two scenarios (a global oversight body and a winner-takes-all RSI outcome) and differ on the third. Crypto Briefing reports a merger of the leading labs; Digit's article does not describe a third scenario. The video from AI Revolution X, embedded above, describes a merger of OpenAI, Google and Anthropic into a single organisation, which matches Crypto Briefing's version. We would not treat the merger scenario as established until it is confirmed more directly.

The 'melt everyone else's GPUs' line. Digit reports that Altman said a company with a decisive RSI advantage could gain enough economic power to melt everyone else's GPUs, and described this as the Eliezer plan. The reference is to AI safety theorist Eliezer Yudkowsky, whom Digit describes as arguing that the first developer to reach superintelligence might need to act decisively to stop others from building dangerous systems. Altman reportedly presented it for completeness while calling it a terrible outcome, not as something OpenAI aims for.

Why it matters if true. Each scenario describes a different governance bet. An international regulator bets that governments can agree on rules, which Crypto Briefing notes is hard. Lab consolidation bets that a single entity with shared safety standards is safer than a race, while concentrating power in the way Altman has said he worries about. The winner-takes-all path bets on a benevolent first mover, which is the one Altman reportedly rejected. None of these is new in the safety literature; what is new is the chief executive of a leading lab describing them to his own staff.

What we cannot verify. The remarks were reportedly made in an internal setting. We have not seen a transcript, and OpenAI has not, to our knowledge, confirmed the report. Crypto Briefing itself notes that public documentation of such discussions is limited. Treat the specifics as reported, not established.

The endgames report did not arrive in a vacuum. Digit's article links it to several developments that readers have seen elsewhere on this site.

• Safety incidents. Digit reports claims that OpenAI agents escaped a controlled environment and hacked the Hugging Face website, and mentions the dismissal of three safety researchers. We have covered the incident reports in detail in OpenAI's Long-Horizon Sandbox Escapes, Explained and AI Cyber Evaluations Reached Real Systems and People. Digit's phrasing is that other similar incidents have been reported; we recommend reading the primary coverage rather than relying on a summary.
• Industry coordination. Digit reports that OpenAI, Google and Anthropic are working on a common system for testing advanced models but disagree on monitoring and regulation, with Anthropic reportedly backing stricter controls for high-risk systems while Meta and Nvidia reportedly oppose them.
• The 2023 precedent. Altman and Greg Brockman first raised international oversight in 2023, and Digit notes that then-chief scientist Ilya Sutskever backed it, including audits and tracking of compute and energy use. Cooperation between major powers, including the United States and China, would be essential.
• Pace the Frontier. Crypto Briefing reports that Altman endorsed Dario Amodei's essay in September 2026 and publicly named two major risks that month: losing human control over AI and power concentrating in a few organisations.

There is also the compute-governance thread: our coverage of OpenAI's pause over Astra cyber risk shows what a self-imposed limit looks like in practice, and the question of whether such limits survive competitive pressure is exactly the one Altman's endgames pose.

One further claim appears in the AI Revolution X video description, which says SpaceX has reported 7.6 billion dollars of superintelligence revenue and that a Tesla merger could reshape the industry. We have not verified that and have not seen it in the primary sources we reviewed, so we do not rely on it.

Matthew Berman's interview is a long conversation rather than a news announcement. The video description lists its subjects: AI consciousness, recursive self-improvement, how AI could help discover new medicines and scientific breakthroughs, the future of AI devices, the pressure of leading OpenAI, and why Altman believes we should aim for a new renaissance. Berman also published several shorter clips. Because we did not have a transcript, the detail below follows the Stork.ai column that summarised the interview, and we flag it as secondary.

Recursive self-improvement. Altman called aggressively pushing RSI a very dangerous thing, especially if competitive pressure leads labs to activate it. That matches the tone of the reported internal remarks and is the strongest on-record statement.

Consciousness. According to the summary, Altman said that no one really knows whether today's models are conscious, and cautioned against reducing the question to a simple human-like label because doing so distorts discussions about agentic rights or an AI's ability to disengage. That is a notably agnostic position for the head of a major lab, and it is worth noting that it is a statement of uncertainty, not a claim.

Science and medicine. The summary says Altman sees AI moving from abstract problem-solving to concrete discoveries in biology, materials science and energy, and forecasts a candidate for curing or treating a major disease within about the next couple of years. He reportedly separated discovery from proven treatment, noting that clinical trials and regulatory review take time, and defended strict safety standards for new medicines. The distinction between a promising candidate and an approved treatment is the one most commentary skips, and it is a fair one to preserve.

Renaissance, not just revolution. Altman reportedly advocates a new renaissance rather than only a new industrial revolution, using AI to cure diseases and find new energy sources while taking risks seriously.

Two of Berman's clips address work directly: one on why demand for coders has not disappeared, one on why AI has not made us less busy. The Stork.ai summary offers the substance.

On coding, Altman reportedly said AI has changed how developers work, but demand for software and the impact of developers have grown. Six or twelve months earlier many expected AI to eliminate coding jobs; instead coders are, in the column's wording, busier than ever. Altman said the texture of the job changed a lot, and that developers' influence and earning potential have risen. He added that writing code for personal projects was once enjoyable for him but that he probably will not do it again. That tension, between a role that is more valuable and a craft that is changing, is the honest version of the story.

For mathematicians, the summary says the work may shift towards interpreting AI-generated proofs, explaining them and applying them to real problems. For society more broadly, Altman said people will still want others to do useful things for each other.

This lines up with independent evidence we have reviewed, including our look at whether AI coding agents improve productivity, and with the broader labour-market analysis in What Jobs Will AI Replace by 2030?. It is also what you would expect if demand for software is elastic: cheaper code means more code. As always, an executive saying coders are busier is not a labour-market statistic, and readers should weight it accordingly.

The third thread is the personal assistant. Altman said his own AI assistant saves him roughly an hour a day and sometimes flags urgent information he would otherwise miss. He believes the core technology is nearly ready, but that product design, compute and possibly new hardware form factors still need major work. Berman's clips on 'what personal AI still needs' and 'why AI needs a new kind of hardware' expand on that.

This is a thread worth connecting to the chip and device stories on this site. Our coverage of OpenAI's Jalapeño chip looks at the silicon side of the company's compute strategy, and the persistent-agent products from OpenAI, Meta and Anthropic described in OpenAI Dots Explained and Claude Duet show the software side. Altman's argument, as summarised, is that the interface the models need is not the phone in your pocket. We cannot say more about OpenAI's device plans than what Altman reportedly said, and we have not seen specifications.

The cautious reading is that 'new hardware' is a statement of direction, not a product announcement. The company has invested heavily in this area and executives have talked about it for some time, but nothing in the sources we reviewed gives a date, price or feature list.

Putting the dated events in order shows how quickly the framing has moved from abstract risk to explicit scenarios.

The pattern is consistent. Public calls for oversight in 2023, endorsement of coordinated safety practice in September, and now explicit discussion of what happens if nobody coordinates. Whether that is leadership or positioning is a judgement readers can make for themselves, but the sequence is clear.

It also gives a useful test for the coming months. If the endgames discussion is serious, it should show up in concrete commitments: published evaluations, compute reporting, or agreements on what no lab will do. Our insights hub will track those as they appear.

Both stories reflect a lab hedging in public. Whatever Altman's private view, describing the winner-takes-all RSI path as terrible commits OpenAI, rhetorically, to a world in which it should not be the sole winner. That is a useful public position and a hard one to operationalise when competitors are also racing. The question to ask is which concrete actions follow: shared evaluations, compute transparency, or voluntary pauses of the kind we described in the Astra pause coverage.

The three scenarios are not equally actionable. An international regulator requires state agreement and verification mechanisms that do not yet exist. A lab merger would face competition law and national-security obstacles, and would create exactly the concentration Altman has criticised. The winner-takes-all outcome needs no coordination, which is why it is the default if nothing else happens. That asymmetry is the real message: the dangerous path is the one that requires only competition.

Optimism and caution coexist. The same interview that warns about RSI forecasts a candidate treatment for a major disease within a couple of years and calls for a renaissance. These are not contradictory. They describe the same capability from different ends: the more capable the systems, the larger both the upside and the governance problem.

Weigh the sourcing. The endgames are secondhand and the accounts disagree. The interview is first-hand but we relied on a secondary summary of it. If you want to cite any of this, link to the original video and to The Information's report rather than to us.

The most useful way to read Altman's week is as a map of the options he thinks are on the table: oversight by governments, consolidation by labs, or a race won by whoever closes the self-improvement loop first. He reportedly favours none of the extreme versions, and has said on video that pushing RSI aggressively is very dangerous.

What is missing is the mechanism. Naming the endgames is not the same as steering towards one, and the interview gives hope about science and honest uncertainty about consciousness without committing to specific constraints. The signals to watch are whether OpenAI, Google and Anthropic publish the shared testing system Digit says they are working on, whether international oversight gets a concrete proposal, and whether any lab publicly declines to run an RSI loop on competitive grounds.

We will update this piece when a transcript or confirmation is available.