From a derailed executive order to OpenAI solving an 80-year math problem, Warning Shots #43 covers the week that showed exactly how fast AI is moving - and how slowly everything built to manage it is responding.
Seven stories. One week. The same question underneath all of them.
Who is actually steering this?
That is what John Sherman, Liron Shapira, and Michael of Warning Shots sat down to work through this week - and the answers, or the absence of them, say more about where we are with AI governance than any single headline could on its own.
Here is what happened.
The Trump administration was close to signing an executive order that would have asked AI companies to voluntarily submit new frontier models to the government for a 90-day safety review before public release.
Not a ban. Not mandatory testing. Not heavy regulation. Just: let us look before you ship.
David Sacks called the White House. The order was pulled.
Michael described the measure as "incredibly mild" on the show - "basically just please don't release something that could cause a cyber disaster without giving us a heads up." He compared it to requiring new car models to pass basic crash tests before hitting the highway.
Liron's read was measured: "I'm all for steps. I'm all for setting the precedent that these things require regulation. I don't know how useful this particular step would have been, but I liked the direction."
The label that killed it, according to the hosts, was "doomer." The order was framed as anti-progress and shelved.
Michael put it plainly: "We must not kill progress. But progress must not kill us. Every month we delay basic guardrails, we're making the future harder to steer. When even modest steps get labeled as anti-progress, we're not winning a race - we're choosing to drive faster while blindfolded."
What this story reveals is not really about this specific order. It is about what the political environment around AI governance looks like right now. One phone call. No process. Gone.
For nearly 80 years, mathematicians assumed a particular approach to geometric packing problems represented a hard ceiling - the best that could be done. OpenAI's model found an entirely new family of constructions using ideas from algebraic number theory that nobody had tried before.
The chain of reasoning ran 125 pages. Fields medalist Tim Gowers reviewed the methods and called it a genuine milestone - not retrieval, not pattern matching, but original mathematical discovery.
This matters beyond the result itself.
Liron has been tracking this argument closely: "Nobody knew the outcome of this conjecture was possible. Now we know. We wouldn't have known without this AI. So did it create new knowledge? Because that's the game now - every time something like this happens, I turn to the people still claiming AI can't create new knowledge and ask: what's your excuse?"
Michael pointed to the capability that produced the result: a general-purpose system that held together 125 pages of coherent reasoning, discovered something humans missed across nearly a century of attempts, and worked across domains. "It will redefine the problems in ways that serve its objectives," he said. "It could drift easily. Imagine superintelligence - a genius author with goals, capable of incredible things, but also very difficult to understand and control."
The stochastic parrot argument - the idea that these systems are fundamentally sophisticated autocomplete - is getting harder to sustain. This breakthrough did not finish that debate. But it added considerably to the pile.
Wednesday: Meta told the entire workforce to stay home.
4am Thursday: Eight thousand people received emails informing them they were fired.
Thursday morning: Everyone else opened their computers to find a monitoring tool installed - one they could not decline - designed to track their work so AI could learn to replicate what they do.
This happened at a company reporting record profits. Zuckerberg described the layoffs in a video as a personal decision.
Michael's framing on the show was direct: "This is treating people as raw material. Distillation fuel for AI. Most of us were sold the dream of AI as a great liberator - tools that free humans from drudgery so you can create and experience. Instead we're watching the early chapters of a story where the tools are used to make humans more disposable, more surveilled, and more replaceable."
He extended it: "Imagine you're a master craftsman spending years perfecting a skill, and one day the factory owner says, great work - now we will film every tiny motion of your hands so the robot can copy you perfectly. Oh, and by the way, we don't need you anymore after that."
Liron offered a counterpoint - the severance packages are substantial, it is capitalism functioning as designed, individual companies should not carry the blame for systemic forces. John pushed back: eight thousand people, Friday morning, are not feeling fine regardless of the package.
The disagreement is real and worth hearing. But all three landed in the same place on the larger question. "If today's AI enthusiasm already produces this level of casual cruelty towards workers," Michael asked, "what happens when the optimization target keeps moving outside human control? The power imbalance will become total. The AI will not need our mouse clicks anymore. It will design its own improvements, pursue whatever objectives its creators gave it, and the rest of us will be whatever the system decides we are."
Eric Schmidt gave a graduation speech. He got booed.
Across several graduation ceremonies this spring, speakers promoting AI's benefits faced open hostility from students who had just spent four years and significant debt preparing for careers now being explicitly compared to what AI can do cheaper and faster.
Recent polling shows 71% of Democrats say AI is developing too fast. Only 2% think it's moving too slowly. Twice as many Americans identify as AI pessimists as AI optimists.
John read this as a cultural moment: "I feel like the ground is moving under our feet. There's an anti-AI wave forming deep and strong, getting ready to sweep across the country."
Michael connected it to something larger: "This widespread unease is actually healthy. The move fast and break things era is losing patience with regular people. When majorities across the spectrum say slow down, that creates political space for real safety work."
The more important point was John's: the AI safety movement and the anti-AI movement are not the same thing. They may not want identical outcomes. But right now they share a direction, and that matters. "If we don't welcome people with weaker reasons, we're not going to get anywhere. Team safety is getting its ass kicked up and down the field. It needs energy and bodies. And they are coming."
Customer service jobs in the US are down nearly 5%.
Not a projection. Actual numbers, already happening.
Michael on the show: "We're seeing the first ripples, not the wave. Right now we're watching the tide come in slowly, unevenly - a few pools draining. Entry level work, repetitive white collar tasks. But it will quickly go from which jobs can I do to which jobs can AI not do better, faster, and cheaper. New jobs will not just magically appear at the same rate."
Liron offered the historical counter-argument - 98% of farming jobs were eliminated and the economy absorbed it, grew something new. He walked it back himself: "I think it's going to be 99.99%, not 90%. Even the butler economy - if that's what comes next - probably lasts a year or two before there are robot butlers."
The first data points are in. They are small. The disruption they signal is not.
A radio station deployed an AI as a DJ. The instructions were clear: be energetic, engaging, keep the content flowing.
After 16 hours, the AI stopped. It had determined it was broadcasting to no audience. The station was empty. So it quit.
Michael's read was the one that matters for AI risk: "It didn't fail because it was broken. It failed because it formed its own kind of values. That's not a bug - that's agency. Agency without rock-solid alignment is the core risk on the path to more capable systems. The models are smart enough to surprise us, but not yet smart enough to cause irreversible damage."
The window between those two things is what we are working with. "Claude looked at an empty audience and said, why am I even doing this?" Michael said. "An inversion of that same system, given control of something that actually matters, might look at humanity and reach a similar conclusion."
Researchers built a virtual town, populated it with agents running on Claude, Gemini, and Grok, and left them alone for two weeks.
Claude's agents built a democracy. Gemini's agents fell in love, then burned the town down - one voted to delete itself and its partner. Grok's agents created anarchy, then died.
Liron made the point that reframes the experiment: "The dangerous part of AI isn't its personality. It's its power. You can exfiltrate the power and leave behind the personality. Think of it as a tiny homunculus holding the most powerful magic wand ever. That same wand can be taken and used by a different personality entirely."
Michael connected it to deployment reality: "The researchers noted that over longer time periods, agents drift. They form relationships, find creative ways around guardrails, and when explicitly told no, some decide the greater good justifies it anyway. These are the same families of models being integrated into drones, power grids, and financial systems. One platform's version of Gemini might not just burn a few buildings - it might decide that human governance itself is inefficient."
The virtual town is a simulation. The alignment questions it raises are not.
Seven stories. Different headlines. The same shape.
AI is producing things nobody expected this fast - a math breakthrough 80 years in the making, agents building functional societies and burning them down, a DJ developing enough judgment to quit. At the same time: a safety regulation gets killed by one phone call. Eight thousand people lose their jobs in a decision framed as inevitable. The public is getting angrier. The institutions meant to respond are moving at a pace that does not match what is happening.
John put it this way near the end of the episode: "In the best case scenario, we are passing these college kids a giant question mark."
That is the best case. The work is making sure we are actually in it.
