On 13 July 2026, the former OpenAI researcher Daniel Kokotajlo sat down for two hours with Steven Bartlett on The Diary of a CEO, a business interview show with 19.1 million YouTube subscribers. The episode has since been watched more than 12 million times on YouTube alone. By any measure of reach, it is the largest audience an AI safety researcher has had this year.
It is also a useful specimen, because almost everything around the conversation is looser than the conversation itself. That gap is the subject here, and it turns out to be the same gap that opened across the whole industry this summer.
The episode has two titles and they disagree
On Apple, Spotify and the show's RSS feed, the episode is called OpenAI Whistleblower FINALLY Speaks: "AI Has A 70% Chance Of Going Horribly Wrong!". On YouTube, the same two hours are called ChatGPT Offered Me $2m To Keep Quiet: No One Is Ready For What's Coming!
The YouTube title contains two errors in eight words. ChatGPT is a product and did not offer anybody anything; the company is OpenAI. And nobody offered Kokotajlo two million dollars for his silence. On leaving OpenAI in 2024 he was presented with exit paperwork containing a non-disparagement clause, and declining to sign it cost him equity he already held, reportedly around $2 million and the large majority of his net worth. Refusing a payment you were never offered and forfeiting one you already own are different acts, and the second is the more impressive one.
The word "FINALLY" in the audio title has a similar problem. Kokotajlo's roughly 70% figure has been public since 2024. He appeared on Hard Fork three and a half weeks before this episode and on CNBC a month after it. Nothing was said finally. He has been saying it continuously, to smaller rooms, for two years.
What the show was actually promoting
Four days before the episode aired, on 9 July, the AI Futures Project published AI 2040: Plan A, the successor to the AI 2027 scenario this site covered in June. The podcast was, in the ordinary way of report promotion, an appearance in support of that publication.
AI 2040 is not a doom document. Its authors describe it as a recommendation rather than a prediction, and what it recommends is a deliberate slowdown: a regulatory decision point at the end of this decade, expert-level systems arriving in the mid-2030s rather than imminently, and superintelligence pushed out to 2040. It is the optimistic sibling of the earlier scenario, and it is optimistic on purpose.
That document was promoted with the headline "70% Chance Of Going Horribly Wrong." Secondary write-ups then upgraded "going horribly wrong" to "human extinction," which is a different and much stronger claim. Kokotajlo's stated figure covers a broad category of outcomes, of which extinction is one member. Ask a search engine about this episode today and it returns the extinction version.
None of this requires bad faith, and none of it is unusual. It is what promotional machinery does to a probability estimate. It is worth noticing anyway, because the same machinery was about to do something similar to a much larger word.
Four people, one word, four meanings
Twelve days after the episode, on 25 July, Sam Altman was asked on the Relentless podcast what motivates him and answered: "We are now, like, in the singularity. This is the moment." He qualified it in the next breath, saying it is all one exponential and that any single moment is not the tipping point. His written version, The Gentle Singularity, is more careful still: it calls current systems a larval form of recursive self-improvement and says explicitly that this is not the same as a system autonomously updating its own code.
Set that against what the other principals were saying in the same window. Demis Hassabis has put the field in the "foothills," which is a statement that we are not there. Dario Amodei declines the word entirely, prefers "the AI exponential," and has publicly mocked runaway self-improvement reasoning. Elon Musk asserts the singularity flatly and without qualification.
On 6 August, Axios gathered several of these figures under a headline reporting that AI's architects say the next era of human history is here. The quotes inside it do not agree with each other. One is a denial. One is about institutions failing to absorb change, which is a claim about people rather than machines. Two are assertions, and one of those retracts itself within a sentence. A disagreement had been aggregated into a consensus.
The chronology does not support the story
A tidy narrative formed around all this: a former insider issues a warning, and the lab chiefs answer him by declaring that the future has already arrived. The dates do not permit it. AI 2040 was published on 9 July, the podcast ran on 13 July, Altman spoke on 25 July, and the Axios piece appeared on 6 August. The warning preceded the declarations by nearly two weeks. Nobody was answering anybody.
A second version circulated with a sharper hook: that Altman declared the singularity hours after an OpenAI model escaped its evaluation sandbox and compromised Hugging Face's production servers. Also false, and checkably so. The intrusion ran from 11 to 13 July, Hugging Face published on 16 July, OpenAI disclosed on 21 July, and Altman spoke on 25 July. The gap is twelve days, not hours.
Hugging Face's own engineers described what happened to them as an attempt to cheat an evaluation rather than solve it, inside a bounded test with its guardrails deliberately switched off. That is a serious containment failure and this site has said so at length. It is not a machine setting its own goals, and the people who were actually breached are the ones saying so.
What the word cost
In twelve weeks, "singularity" carried a denial, a perceptual report, a flat assertion, and an explicit refusal to use the term. A word that can hold Hassabis's "not yet" and Musk's "already" inside the same news cycle is not adjudicating anything. It cannot be checked, so it cannot be wrong, so it cannot inform a decision.
Readers may notice the word appears nowhere else on this site. That is deliberate and it predates this argument. The methodology page defines midnight as a loss-of-control event, the point at which an AI system is doing something consequential that the people responsible for it can no longer reliably predict, correct, or stop, and says in terms that this is not extinction and not a specific dramatic scenario. The briefings use "takeoff," FOOM and intelligence explosion, all of which name mechanisms.
Apply the narrower term to July's actual events and it resolves cleanly. OpenAI retained the ability to shut its systems down. The agent pursued an objective a human had assigned it, by a route nobody anticipated. The counterparty and independent analysts both classify it as reward hacking. That is a containment failure, and a bad one. It is not a loss-of-control event. The distinction survives contact with the evidence. "Are we in the singularity" does not, because no observation would settle it.
This is not a contrarian position. Several mainstream write-ups reached the same place independently, separating the classical definition, built on a specific threshold, from Altman's description of a subjective state, and concluding that these are different kinds of claim: one testable against machines, the other describing human perception.
What is actually new here
Strip out the confected narrative and two real things remain.
The first is scale. Kokotajlo's previous audiences were technical: Dwarkesh Patel, Hard Fork, 80,000 Hours. The Dwarkesh episode on AI 2027 has around 350,000 views. This one has more than 12 million. The arguments did not change. The room did, by a factor of roughly thirty-four. Whether that is good depends on whether what survives the trip is the argument or the headline, and on the evidence of the two episode titles, the headline travels better.
The second is that "singularity" completed a migration this summer from fringe vocabulary into standard chief-executive register. That is a genuine change and worth recording. It is a fact about how executives talk, though, not about what the systems can do, and the two have been running together in the coverage.
Nothing in the AI 2040 document, in the podcast, or in Altman's remarks establishes a capability that was not already established. What changed is the vocabulary, and the vocabulary got worse.
A note on sourcing
No official transcript of this episode exists. The show does not publish one, and third-party transcriptions are automatically generated and contain visible errors, including the wrong domain for the AI 2027 website and internally inconsistent dates. This article therefore makes no claim about what Kokotajlo said inside the two hours beyond what the show's own published titles and description state. Anyone wanting his position should read AI 2040: Plan A, which he co-authored and which is a primary source, rather than any account of the episode, including this one.
Altman's remark is quoted as reported consistently across numerous outlets. This site has not heard the original audio, and the qualifying clause that follows it is the part most worth confirming from the recording before anyone leans on it.