A charitable reading of Sam Altman

Sam Altman is often read as either a hypocrite, who warned about AI and then raced to build it, or a salesman, whose warnings are marketing. This entry tries the most charitable reading available: that his choices follow from one strategic bet, made early and stated in public, that the safest way through is for AI to arrive soon and gradually, in plain view.

This is an exercise, not Eddie’s view. Eddie finds Altman untrustworthy and reads much of his career as a pursuit of power, a judgement many people who have worked closely with him share; OpenAI’s own board removed him in 2023, briefly, saying he had not been “consistently candid in his communications.” But a position is worth understanding at its strongest before rejecting it, and the pieces of this one, scattered across a decade of Altman’s writing, fit together better than either caricature suggests. The case comes first; where it is weakest comes after.

He took the danger seriously before he had a stake in it

In February 2015, ten months before OpenAI was founded, Altman wrote that superhuman machine intelligence “is probably the greatest threat to the continued existence of humanity” (Machine intelligence, part 1). The argument was the standard one: such a system need not be evil to kill us, only indifferent while pursuing some other goal, and “programs often achieve their fitness function in unpredicted ways.” He called the belief that the danger is real but far off “sloppy, dangerous thinking.”

The next day he asked for regulation (part 2): a new agency to monitor anyone training such systems, safeguards for development past a certain point, and funding for safety research. That is the position of someone persuaded by the case for AI risk, not of someone looking for a product.

The premise: it will be built anyway

The step that turns a worried observer into a builder is a judgement that AI cannot be stopped. Altman made it early. In a May 2015 email to Elon Musk, he wrote that he doubted humanity could be stopped from developing AI, and that if it was going to happen, it would be good for someone other than Google to do it first. A month later he proposed the lab’s mission: build the first general AI and use it for individual empowerment, “the distributed version of the future that seems the safest”, with safety as “a first-class requirement.”

Everything that follows depends on this premise. If AI cannot be stopped, the question is not whether it is built but how, when, and by whom. See Why people build AGI despite the danger.

The two-by-two

Given that AGI is coming, Altman frames the choice as a two-by-two: AGI could arrive soon or late, and the takeoff from the first AGI to far more powerful systems could be slow or fast. In Planning for AGI and beyond (February 2023) he wrote:

Many of us think the safest quadrant in this two-by-two matrix is short timelines and slow takeoff speeds; shorter timelines seem more amenable to coordination and more likely to lead to a slower takeoff due to less of a compute overhang, and a slower takeoff gives us more time to figure out empirically how to solve the safety problem and how to adapt.

There is a lot packed in here.

• Slow takeoff is the prize. If capability rises gradually, each generation of systems is only somewhat stronger than the last. Problems show up in systems weak enough to fail visibly and survivably, and alignment can be worked out by experiment rather than gotten right on the first try. This is the hope behind Can we learn from AI failures before they are fatal.
• Waiting can make takeoff faster. If the ideas for AGI arrive late, after decades of chip production, the first working method meets an enormous stockpile of compute and can scale almost at once. Building early, while compute is scarce, means each step is limited by how fast new chips and data centers come online.
• Fewer players are easier to coordinate. Soon, only a handful of well-funded labs can build frontier systems, and their training runs are large enough for governments to see. Later, the field could be crowded and the cost of entry low.
• The “now” half is a means, not a goal. On this reading Altman does not prefer short timelines for their own sake. He accepts them as the price of the slow takeoff he wants.

Shipping as a warning system

The bet also explains OpenAI’s habit of releasing products. “A gradual transition gives people, policymakers, and institutions time to understand what’s happening,” the same post argues, so that society can adapt “while the stakes are relatively low.”

The alternative was not a world without advanced AI. It was a world where it was developed inside a few labs, mostly out of public view, until it was very strong. Google had capable chatbots internally before ChatGPT and did not release them. ChatGPT, whatever else it did, gave hundreds of millions of people direct experience of what the technology could do, and moved AI risk from a niche worry to a subject for heads of state. Under the bet, that public awareness is part of the safety plan: governments cannot regulate, and voters cannot demand regulation of, something they have never seen.

Now he is warning, as the plan said he would

The 2023 post also said that “coordination among AGI efforts to slow down at critical junctures will likely be important.” In 2026 Altman’s public statements read like that stage arriving.

• September 2026, Fortune. He called it “unacceptable to be taking a 10% chance of killing everybody by the end of the decade,” and said OpenAI had been “pausing training runs until we can make a safety case that we’re more comfortable with” (interview).
• September 2026, UN Security Council. “We could lose control of the future to AI,” he told the Council, alongside Dario Amodei, as both asked governments to set safeguards (AP).
• November 2026, on X. “People should always be in control of the future… People should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress” (post).

Taken together, the position is consistent from 2015 to 2026: superintelligence could end humanity; it cannot be stopped; the safest path is soon, gradual, and public; so build it in the open and, as the stakes rise, slow down and ask governments to set limits.

Where the bet is weakest

• The quadrant argument conveniently endorses what OpenAI does anyway. “Short timelines are safer” is a very comfortable conclusion for the company shortening them. The compute-overhang reasoning also cuts against OpenAI’s own push for ever-larger compute buildouts: if scarcity of compute is what keeps takeoff slow, racing to remove that scarcity removes the safeguard.
• Takeoff speed may not be anyone’s choice. Building early does not guarantee a slow takeoff. If AI systems come to drive AI research, the speed is set by that loop, not by a lab’s intentions. See A feedback loop is not a speed estimate and The step from AGI to superintelligence could be short.
• Awareness came with a race. ChatGPT woke up the public, and it also pushed every large technology company to compete harder. A strategy meant to buy time spent some of it.
• The safeguards that made the bet credible have weakened. The 2023 post relied on a nonprofit that “can override any for-profit interests” and a cap on investor returns; the post now carries a note that its description of OpenAI’s structure is outdated. Researchers who led safety work left in 2024 saying safety had taken a back seat to products.
• “Ultimately” does a lot of work. The November post says human control “should ultimately constrain AI progress,” not that it does now. Like Musk, Altman has not made a conditional offer, such as stopping frontier training if other leading labs and governments agree to stop, that would make his warnings cost him something.
• It asks us to trust him. Every step of this reading takes Altman’s stated reasons at face value. The same record fits a less flattering story, in which the risk talk wins talent, money, and a seat at the table where the rules get written, and each safeguard lasts only until it gets in the way. Nothing in his public statements distinguishes the two stories; only costly actions would.
• The premise is self-fulfilling. “It will be built anyway” is true partly because people like Altman build it. That weakness is common to every lab leader’s case, and is discussed in Should safety-focused labs keep building.

What would make this reading credible

Evidence that the slowing-down half of the plan is real, not rhetorical. Published safety cases for OpenAI’s largest training runs, with outside audits. Public, verifiable criteria for when a run is paused, and examples of runs that were. Support for binding international limits, with OpenAI agreeing in advance to stop if others do. If those arrive, the bet looks like a strategy carried out in good faith; if the warnings keep arriving without them, the less charitable reading, Eddie’s own, stands.