Warning Shots #44 captures a moment when the financial story and the safety story became the same story. A near trillion-dollar valuation, a wage debate no one could win, an explicit corporate goal of building AI that improves itself, and a fleet of new cameras pointed at private life. As John Sherman put it, the AI race is missing the one thing every other high-speed race takes for granted: a way to slow everyone down when the track is no longer safe.
The biggest argument in artificial intelligence right now is not about how smart the models are getting. It is about money, and where it goes. On Warning Shots #44, John Sherman is joined by Michael (Lethal Intelligence) and Liron Shapira (Doom Debates) to work through a week in which the economics of AI moved to center stage: corporate AI bills came due, Anthropic reached a near trillion-dollar valuation, and the conversation kept returning to one unanswered question about AI job loss and where future wages come from.
It is a useful episode precisely because the three hosts disagree. What follows are the key discussion points, with each claim attributed to the person who made it.
The hosts opened with a run of reports suggesting that AI spending is outpacing AI returns. As discussed on the show, Microsoft reportedly canceled certain Claude Code licenses over cost, Uber is said to have spent its 2026 AI budget in four months, and one large company, rumored to be Amazon, reportedly ran up roughly half a billion dollars in a single month because no usage limit had been set. A Pizza Hut franchisee is reportedly suing over AI that mishandled a wave of orders.
John Sherman reads these stories as evidence that deploying AI is harder than the plug-and-play promise suggested. Liron Shapira pushes back. He argues that this is an early optimization gap rather than a true AI bubble. In his view, the technology is new and barely tuned, and the cost per unit of useful work keeps falling. He points to Anthropic cutting the price of its fast-mode coding tool by roughly a third as a sign that the same budget will buy far more output within a year.
Michael offers the caution that frames the whole show. The dot-com comparison only goes so far, he argues. If that bubble burst, the world lost some search engines and online stores. If the current AI build-out overshoots toward systems that can plan and improve themselves, the failure mode is far harder to reverse.
The strongest section of the episode is an argument the hosts never fully settle. Liron Shapira makes the optimistic case on AI job loss: the economic pie grows. For two centuries, he notes, more productive workers have earned more, and he expects that pattern to continue as AI makes individuals dramatically more capable.
John Sherman presses on the mechanism. He offers a blunt example: take one hundred doctors, automate most of the work, and keep five. The ninety-five who are let go do not disappear. They compete for the remaining roles, and they compete by accepting lower pay. That dynamic, he argues, pushes wages down, not up.
Michael sharpens the concern. If a person's labor is no longer needed, he asks, what actually carries the growing pie back to them? Without a redistribution mechanism such as universal basic income or taxes on the companies capturing the gains, the result could look like a demand collapse, because people who are not earning are not spending.
To his credit, Liron concedes the endgame. He describes a scenario he calls gradual disempowerment, in which robots and models can do nearly everything and most people have little left to offer the market. In his words, that is a scary situation, and he hopes a basic income is waiting on the other side of it. The disagreement, then, is less about the destination than about how soon we arrive and whether anyone is building the bridge.
The conversation then turned to Anthropic raising roughly sixty-five billion dollars at a post-money valuation approaching one trillion. Liron Shapira argues that, if you believe AI will capture a large share of what human labor currently earns, even that Anthropic valuation could be considered low.
Michael names the cost of pricing the company that highly. A valuation of that size, he argues, manufactures pressure to ship faster, deploy more widely, and treat safety as something to address after the next milestone. Because every major lab faces the same incentive, no single one feels able to slow down.
Liron then highlights what he sees as the real prize underneath the race between Anthropic, OpenAI, and Google. The labs, he says, are explicitly aiming at recursive self-improvement: building an AI capable enough to improve the AI itself, running it, and returning to a system that has advanced the equivalent of years because each improved version improves the next. He describes this as the move the labs are intentionally building toward.
John Sherman responds with an analogy from car racing. When a hazard appears on the track, a caution flag drops and every driver slows from full speed to a crawl until it is safe. The AI race, he argues, has no caution flag, no agreed trigger, and no one whose job it is to throw it.
The episode closed on a trend toward putting cameras into everyday life. As discussed on the show, Apple is reportedly adding cameras to its AirPods, and OpenAI has reportedly run a program placing cameras inside homes in New York to record ordinary life for training data.
Michael offers the most vivid framing of the show, describing a system with billions of eyes assembling footage from millions of homes, learning how people behave when they believe no one is watching. Liron, often the more alarmed voice, pushes back here. He considers privacy concerns secondary to the central question of whether humanity retains the ability to pause and reach a stop button at all.
The hosts also noted Pope Leo's recent encyclical on AI, which described investment in AI-powered weapons as feeding a spiral of annihilation. The AI Risk Network shares that alarm. Wiring autonomous systems into weapons is among the most dangerous paths available, and it deserves far more public scrutiny than it currently receives. Our commitment remains to peaceful, lawful, democratic advocacy.
Warning Shots #44 captures a moment when the financial story and the safety story became the same story. A near trillion-dollar valuation, a wage debate no one could win, an explicit corporate goal of building AI that improves itself, and a fleet of new cameras pointed at private life. As John Sherman put it, the AI race is missing the one thing every other high-speed race takes for granted: a way to slow everyone down when the track is no longer safe.
You can help raise the salience of AI risk and push for sensible guardrails. Take action at https://safe.ai/act.
Watch Warning Shots #44 in full on The AI Risk Network: https://www.youtube.com/@theairisknetwork
