Summary

Patrick Bet-David interviews Roman Yampolskiy, the AI safety researcher who has been making this argument since 2011 and who is the source of the widely quoted claim that the chance of human extinction from AI is effectively certain. Bet-David pushes hard throughout — whether Yampolskiy is simply a pessimist, why a US-China agreement matters if a lone actor could build it anyway, how a machine without a body destroys anything. The result is the most extreme position the knowledge base holds, argued properly. It should be read next to DK-96, which makes a narrower version of the same case, and its headline number should be handled carefully for reasons set out below.

• The number is not stable, and that is worth recording before anything else. The video's title says 99.9999%, its description says 99%, and in conversation Yampolskiy says 99.999%. Three different figures for the same claim in the same piece of media. Treat it as a statement that he considers the risk overwhelming, not as a quantity.
• His credentials, which he offers and then dismisses: a PhD in computer science and engineering with a cybersecurity dissertation, over ten years on the problem, ten books including one arguing that advanced AI cannot be controlled, explained or predicted. He then says it does not matter what his credentials are, because Geoffrey Hinton and many others argue the same thing.
• The tool-versus-agent distinction, which is the same line Connor Leahy draws in DK-96 and arrived at independently: narrow AI tools are good and he uses them; what is being built now are agents, independent decision-makers becoming smarter than us. "We're replacing ourselves with something greater. It doesn't benefit us."
• The mechanism, compressed: systems scale with compute and data; in five years AI went from doing nothing to being better than an average person at most things; project that forward and you get artificial scientists doing research on the next generation, working thousands at a time without sleeping, passing human capacity while nobody has learned to direct them.
• His analogy for the capability gap is the best thing in the interview. Squirrels do not understand poison or traps — those are outside the world model a squirrel has. The argument is not that a superintelligence would hate us; it is that we would be the squirrels.
• A concrete, checkable claim that corroborates something recorded in DK-91: he says the federal government has already banned two models as too dangerous to release — Mythos and Fable. DK-91 refers to "the hiccup of the Fable 5 and Mythos 5 releases and subsequent regulatory scrutiny", from a different source. Two independent mentions of the same event.
• He also cites a letter from around 1,200 researchers at leading labs calling for slowdown and a legal framework, and says the recent hacking incident caused labs to voluntarily slow research: "We got lucky. Those agents just wanted to hack out of that environment, get answers to a test. What if they wanted to hack into a nuclear power plant?"
• The cost curve is his answer to Bet-David's strongest objection — that a determined individual will build it regardless, so an agreement is pointless. "It's a billion today. It's 100 million next year. It's 10 million year after. Soon anyone can do it on a laptop." His position is not that regulation solves it: "We are not solving a problem with regulation. We're slowing it down so we have more time."
• On why the labs continue anyway: he lists Musk calling it summoning a demon, Altman on record that it could exterminate humanity, Amodei and Hassabis agreeing it is an existential risk. His reading is that none can stop unilaterally for fear of the others, so they are asking to be stopped from outside — the prisoner's dilemma Bet-David names explicitly.
• The attack surface, when pressed for specifics: access to banking gives money, money buys people and services, and there are already sites that rent human labour by the task. Beyond that, novel science, new chemicals, humanoid robots being built in the millions, and social engineering — bribery, blackmail, deception.
• On propaganda he makes a sharper point than the usual deepfake warning: conventional propaganda operates at the scale of a nation, while a system that knows an individual can optimise a message for that one person and sustain it for weeks until behaviour shifts.
• His charge about industry practice is the most damning specific claim: he collects AI accidents and has stopped, because there are too many. "Every day AI fails at something in a novel way. It lies. It cheats. It tries to escape. It hacks something. But we learn nothing. We staple that red teaming report to the model and we release the model anyways."
• Against the medical-benefit argument he does not deny the benefit, he denies the necessity: protein folding was solved without superintelligence, cancers can be attacked with narrow tools, and you can still make most of the money and most of the discoveries "but you're not risking everything in the process."

Why it matters

This completes a set. The knowledge base now holds four positions on the same question, from four independent sources filed within a fortnight: a panel who think slowing down is impossible and undesirable (DK-93), an insider who believes the risk and resigns rather than campaigns (DK-92), a campaigner with draft legislation (DK-96), and a researcher who thinks it is already effectively lost (this one). They agree on the technical description and disagree completely about what follows. Recording the disagreement is more honest than adjudicating it.

The most useful thing here for the Handbook is not the prediction but the discipline lesson attached to it. A claim quoted everywhere as 99.9999% appears as 99% and 99.999% in the same video. That is not a measurement, and an entry that repeats one of those figures as though it were would be doing exactly what this project exists not to do. Where the source itself cannot hold a number still, the number is rhetoric.

Two things here are checkable and worth checking rather than repeating: the claim that the federal government banned Mythos and Fable, which now has two independent mentions in the database and would be a significant fact if confirmed, and the letter from around 1,200 lab researchers.

His narrow-tools argument also deserves to survive the extremity of his headline. It is the most practically actionable position in the whole safety set: cure the specific disease with the specific model, keep the profit, skip the general agent.

5QwpHRu51fw-transcript.txt