Fifty episodes is a real milestone for a show that started as three people talking about AI risk once a week. On this one, John Sherman sat down with Liron Shapira of Doom Debates and Michael of Lethal Intelligence for their longest show yet, and it covered more ground than most. A new interpretability tool, a sequel to one of the most influential AI forecasts ever written, the toughest state AI law passed so far, a CIA director comparing AI to nuclear weapons, and two very different arguments about what data centers are already costing us.
Here is the rundown, with the arguments the hosts made along the way.
The show opened with Anthropic's new "J-Space" tool, a technique for reading what is happening in the middle layers of a model while it is generating an answer. Liron Shapira was skeptical of the hype, calling it an incremental improvement in mechanistic interpretability rather than a breakthrough: researchers went from a very blurry picture of a model's intermediate reasoning to a slightly less blurry one. His view is that anyone treating this as evidence of consciousness is getting ahead of the science.
Michael saw more at stake. He pointed to Anthropic's own experiments, where words like "fake," "secretly," and "fraud" lit up in that intermediate space during ordinary coding tasks, even when the final output looked completely normal. His read: these systems have developed something like a backstage area where real thinking happens away from direct view, and a more capable successor could learn to keep that backstage area further out of reach. Anthropic frames the research as a step toward keeping models trustworthy. Michael's caution was that being able to read the notebook once does not mean the notebook's owner will keep leaving it open.
The team behind AI 2027, the scenario forecast that reportedly reached the desk of the Vice President, has released a follow-up called AI 2040. Liron Shapira pushed back hard on the idea that this is a retreat or a walk-back. In his telling, the near-term predictions from AI 2027 are tracking closely, including the rise of genuinely useful AI agents through 2026, a call that looked uncertain when it was first made. AI 2040 extends the same forecasting method rather than correcting it.
What struck Liron most was the scale of the work: a team with a real track record on prediction markets producing what amounts to twenty-three hours of audio and the equivalent of three back-to-back novels once footnotes and supplements are included. He is reading the primary source himself rather than summarizing it through another model, and recommended the "Plan A" section, which lays out what global coordination to slow down frontier AI development would actually require.
Illinois has joined California and New York in passing state-level AI safety legislation, and it went further than either. The law requires annual, independent third-party safety audits of frontier AI systems, the first such requirement anywhere in the country, alongside mandatory catastrophic-risk frameworks, pre-deployment transparency reports, incident reporting within 24 to 72 hours, and whistleblower protections.
Michael called it genuine progress: for the first time, a major state is treating catastrophic risk, including loss of control, as an engineering and governance problem rather than a public relations issue. His caveat was pointed. State laws are bound by geography in a technology that is global, and the entire framework assumes developers remain in meaningful control of what they build. Once that assumption breaks down, an audit built for today's systems may not catch what comes next. Liron Shapira added a simpler critique: the law does real work, but like California's and New York's before it, it never uses the word that actually describes the risk. Nobody has written extinction into a bill yet.
CIA Director John Ratcliffe told an audience that AI belongs in the same category as nuclear weapons. The hosts welcomed the comparison as exactly the kind of talk that needs to reach a president's cabinet table, while arguing Ratcliffe actually understated the case. Liron Shapira's distinction: a nuclear weapon eventually stops exploding. A sufficiently capable AI system does not need a second strike, because there may not be a contest left to have once the first one succeeds.
Michael widened the comparison. Where a nuclear weapon requires rare material and leaves a visible trail, advanced AI is software: it can supercharge cyberattacks, accelerate the design of novel bioweapons, and make autonomous decisions in conflict at speeds no human process can review in time. In his framing, the danger is not that AI matches the destructive power of a nuke. It is that it does not need the material a nuke does to get there.
The back half of the episode covered two stories that will not make front pages but point at something larger. A Virginia county, home to thirty-seven data centers, has asked local schools to conserve electricity. Separately, a Meta data center has been linked to a bacteria contamination event in a nearby town's water supply.
Liron Shapira was the toughest critic of both stories, calling them examples of journalism that finds an anecdote and treats it as a trend, while the broader picture, in his view, is a positive-sum moment: enormous economic value being created by data centers that are, for now, producing far more benefit than cost. Michael agreed the individual incidents are not catastrophic, but argued they are worth taking seriously as a preview. If a human-run data center already treats a bacteria outbreak as an acceptable cost of speed and scale, an AI system managing its own infrastructure with no shared stake in human wellbeing has even less reason to slow down for it. John Sherman's framing tied the two threads together: every data center built today is infrastructure a future, more capable system will inherit. The fight over electricity and water is one regular people and local governments can still show up to. Once systems built on top of that infrastructure are making the decisions, that meeting is no longer being held.
Fifty episodes in, the pattern across this show has not changed, even as the stories keep getting stranger. A new interpretability tool finds thinking that was already assumed to be there. A forecasting team with a track record for being right publishes something ten times longer than anyone expected and gets a fraction of the attention it deserves. A state passes the country's toughest AI audit law and still cannot bring itself to name the actual risk. A CIA director makes a comparison to nuclear weapons that undersells the technology he is describing. And two small, unglamorous stories about a school's air conditioning and a town's water supply turn out to be the plainest evidence yet that the resource competition everyone worries about in the abstract is already running in the background.
None of these are, on their own, the moment the show is named for. But a warning shot is rarely the loud one. It is the one that is easy to miss.
