Dispatch from Mitch
Anthropic, Nvidia, and the agents among us
The top AI story today is that Anthropic has officially filed for its IPO, expected to take place this fall with a valuation of about a trillion dollars.
This move surprised no one. Moreover, since Anthropic completed its filing confidentially, it was not required to release any new operational details at this time. This made for pretty thin articles.
Anthropic mentions were still everywhere, though, in stories about how the global economy has been shaped by the boom in agentic AI. Anthropic wasn’t the first to build agents — AIs that keep grinding away on tasks, rather than returning single-stage replies like traditional chatbots — but the company became synonymous with the technology when its Claude Code product got good at the end of 2025. This led to viral adoption by non-coders starting in the first months of 2026. Since then, the heavy compute demands of agents have been shaping the market: launching the price of memory chips into the stratosphere, motivating compute-sharing arrangements between erstwhile competitors, and pouring cold water on the build-out bubble thesis.
The agent boom was also cited in today’s announcements by Nvidia, the world’s top AI chip company. Business sections today reported two new initiatives: A line of PC chips built for running more powerful AI models locally, and a robotics partnership with Chinese company Unitree.
Those PC chips reflect Nvidia’s claim that the chatbot era is ending and that “agents are the new workload. They will run everywhere.” The completed PCs, to be produced by partners like Dell and HP, are being touted for their small form factors; some look positioned as specialized alternatives to the Mac Minis people had been buying to host around-the-clock autonomous agent frameworks like OpenClaw. These frameworks, in turn, are well-suited to running open-weights models like the guardrail-free abliterated models I talked about yesterday. In short: I expect scammers to love these machines.
As for the robots, it has been my sense that a missing piece in the robot revolution has been large quantities of affordable, efficient chips suitable for running agents locally — in or very near the robots. That’s because robots need fine motor control, which is hard to achieve when your hands and feet have to bounce signals to a data center that might be two states away. Nvidia is working to put the brain in the robot. What they just announced is an “open humanoid robot reference design” using its own onboard compute in a Unitree-built body.
The American company claims partnering with a Chinese company is “the only way” to bootstrap the lagging U.S. robotics industry. The Wall Street Journal’s coverage expects the arrangement to attract scrutiny in Washington.
Separate Wall Street Journal coverage describes venture capitalists flocking to robotics — the so-called “physical AI” sector — as a refuge from software ventures. Commercial software has proven vulnerable to replacement by home-grown alternatives vibe coded with the help of agents like Claude Code. Robotics companies, on the other hand, have “collectively raised billions in funding at sky-high valuations despite being only a few years old and, in some cases, having little revenue.”
Have you seen your first humanoid robot in person yet? I have a feeling this will be the kind of thing where you go from never seeing one to feeling like you see them everywhere over the course of a year or two. Robotics is on the same sort of exponential as AI itself, with all that this implies about economic transformation and AI’s ability to get by without us.
Dispatches from Joe
A future we decide
In a New York Times op-ed today, Senator Bernie Sanders announced a forthcoming bill which he frames as giving the public both a financial and voting stake in AI.
The bill would establish a sovereign wealth fund, an investment vehicle owned by the federal government, which could pay dividends to ordinary citizens in the same way that, for example, Alaska’s $70 billion oil fund does. To create the fund, Sanders proposes a one-time 50% tax on AI company stock.
You would think the AI companies would oppose this measure, but Sanders quotes no fewer than three companies and CEOs who have recently called for a public wealth fund or for universal payments fueled by AI growth. OpenAI, Anthropic, and xAI’s Elon Musk have all floated similar ideas. I suspect “the government takes half their stock” was not precisely what these folks had in mind, but I admit it would feel like karmic justice for OpenAI to have its stock reclaimed by the public after it abandoned its founding principles and converted to a for-profit.
Strangely enough, the idea of a sovereign wealth fund is also somewhat bipartisan; a presidential executive order in Feb 2025 proposed such a fund, though the fund itself has yet to materialize and the administration has evidently modified the plan toward more generic government investment programs.
Some criticize Sanders for treating AI as both an extinction threat and a potential font of wealth to be hoarded or shared. But AI is indeed both at once. We are talking about an attempt to imbue machines with the very power that put humanity on the moon and gave us the smallpox vaccine.
If AI companies succeed at building a superintelligence, and the AI does not use that power towards human aims, then we all die as an afterthought. If the companies only partially succeed, and get stuck at something only superhuman in some domains, that is still a massively transformative technology that threatens to concentrate wealth and power like none before it. Of course we want to hedge our bets in the face of uncertainty.
I doubt Sanders expects this bill to pass. Most bills don’t. It’s probably intended more as political message than as actual policy. But it does drive home the degree to which ideas which once seemed radical and unworkable are looking increasingly less so as AI capabilities rise.
I think an AI that can do all our jobs can probably also eat the world, so I’m more concerned with how humanity can survive the coming decades than how to distribute AI investment returns. But Senator Sanders ends on a note I can’t possibly contest. “It’s our future,” he reminds us all, and “We must decide it.”
Slop flows downhill
In a hint of what’s to come in the general elections, California political candidates are leveraging AI deepfakes and human influencers to denigrate their opponents and boost their public visibility, POLITICO writes.
This looks like two stories in a trenchcoat, really. The first is about alleged campaign infractions in the California gubernatorial race, which has already gotten somewhat out of hand. A typical race for state governor involves only a handful of candidates; some break double digits. In California this year, there are sixty-two.
Some of the candidates have been accused of paying internet influencers to amplify their reach, without disclosing these payments as the law requires. One candidate argues the creators they’ve worked with merely “like our policies” and were simply paid “for their time.” Uh huh.
But the other half of this story is about the proliferation of deepfake and deeply disingenuous ads. AI has gotten good enough that nearly anyone can sit down at their computer and produce a convincing-looking short video of whomever or whatever they like on a shoestring budget. The trouble with enabling anyone to do this is that some people will.
One recent video uses the voice and likeness of the current mayor of Los Angeles in an attempt to discredit her and boost an opponent. In it, a deepfaked Mayor Karen Bass walks through a practically post-apocalyptic city, speaking of her “accomplishments.”
A pollster interviewed by POLITICO claims this slopification is inevitable. “Water flows downhill, right?” It’s hard to argue with physics. But this attitude reeks strongly to me of the same defeatism peddled by those who oppose all AI regulation for selfish and misguided reasons. The AI race is inevitable, they say, so we’d best resign ourselves to whatever ultimate end it carries us.
America has weathered political challenges just as daunting, in the past. Neither the rising tide of AI capabilities nor the downstream deluge of manipulative slop is inevitable. I fear the first more than the second, but we don’t have to stand for either. We can, and must, stem this new tide, before it sweeps us all away.
Dispatches from Donald
Department of Commerce closes export loophole
StopWatch has previously covered the export controls that restrict China’s access to powerful computer chips. We have also mentioned how Chinese labs try to circumvent these restrictions. For example, Chinese labs with subsidiaries in other countries, like Malaysia, were able to acquire restricted computer chips through those countries. One industry insider estimated that hundreds of thousands of chips might have been obtained in this way.
Reuters’s Freifeld and Potkin now report on work by the U.S. Department of Commerce to close this loophole: Labs and other entities that are headquartered in China will be unable to receive restricted computer chips even for subsidiary entities located elsewhere.
Not every loophole has been closed, as noted by Chris McGuire, a former State Department official: Taiwanese chip manufacturers were once required to do extra work to confirm they are not selling to Chinese front companies. This requirement has not been reinstated.
The robo-surveillance state
The New York Times’s Julian Barnes reports on efforts by a Chinese company to develop predictive surveillance technology. Geedge Networks currently sells software that watches for and flags attempts to evade internet censorship. According to a recent paper by Vanderbilt University, the company is also working on a program that will not just monitor political dissent but predict it. Their efforts thus far are exploratory, but point in the direction of building profiles out of physical and online activity: the places you visit, the people you meet with, the websites you browse, and so forth.
Many countries have had, for a long time, access to an incredible wealth of information — so much data, in fact, that it would be impossible for even a large team of humans to efficiently sift all of it. AI models change the equation. They can easily handle large quantities of data, and they’re only getting better and cheaper. (It is worth noting that Anthropic’s fight with the Pentagon was partly down to AI’s potential for mass domestic surveillance.)
MIRI is best known for warnings about the threat of extinction — it’s not for nothing that Yudkowsky and Soares’s book is titled If Anyone Builds It, Everyone Dies. But in a 2025 report, Peter Barnett and Aaron Scher (of MIRI’s Technical Governance Team) described additional risks, including “authoritarian lock-in,” or the possibility for an authoritarian regime to use AI to entrench itself indefinitely. This is not the worst or even the most likely scenario (it assumes that we don’t lose control of AI), but it’s worth mentioning alongside the prospect of an AI-powered surveillance state. My colleague Alana has posted a summary of Barnett and Scher’s report.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.