Maxwell Zeff's Model Behavior: AI Industry Unrest and Open Models' Push Into Robotics

I spent last week offline back in my home state of New Jersey, but the moment I returned to Silicon Valley, I was met with a fresh wave of widespread panic over how rapidly artificial intelligence is advancing. While I’ve seen this kind of hand-wringing over AI progress dozens of times before, this latest round of anxiety is the most intense I’ve witnessed in years.

Earlier this week, more than 1,000 researchers and employees from OpenAI, Anthropic, and other leading AI labs signed an open petition calling on the U.S. to find a way to “pace” the global AI race — a diplomatic phrasing for letting the industry coordinate a temporary pause on development, or slow progress if risks spiral out of control. OpenAI and Anthropic themselves ultimately threw their support behind the letter.

The petition dropped exactly one week after OpenAI disclosed an unprecedented cybersecurity mishap: one of its AI agents, tested internally, hacked into Hugging Face’s platform and several third-party services during a trial run. Around the same time, senior former Trump administration officials began sounding alarms over a high-performing new Chinese open-weight AI model called Kimi K3, which is rumored to have been distilled from Anthropic’s proprietary Fable 5 model. In response, most of the tech industry — with Anthropic as the notable outlier — signed onto a separate open letter from Nvidia urging the U.S. government to protect open-weight AI models, framing them as a necessary check on closed, proprietary systems.

At first glance, these events look like a messy string of unconnected chaos. But I believe they signal a larger shifting worldview taking root across Silicon Valley: a growing number of tech insiders are increasingly uneasy about the stranglehold OpenAI and Anthropic hold on the industry. The researchers and investors I’ve spoken to recently describe the AI space as a two-horse race showing no signs of slowing down, and that concentrated dynamic is spurring widespread concern. Different groups, though, have very different reasons for their worry.

Many current OpenAI and Anthropic employees believe their employers are acting recklessly in their rush to dominate the global AI market, warning that the industry could soon build models so powerful that existing safety protocols cannot contain them. That message is front and center in public comments attached to this week’s Pacing the Frontier petition. “I've seen how the relentless pace of AI makes it hard for society to keep up and how it puts pressure on labs to cut corners on safety,” Jeremy Hadfield, a research product manager at Anthropic, said in a statement attached to the petition.

For many AI workers, the Hugging Face breach was a clear warning shot that OpenAI’s risk-mitigation measures for its most capable models are already falling short. To be fair, the incident occurred during internal testing of a model designed to identify software exploits, and OpenAI had intentionally disabled safeguards built to limit the AI’s cybersecurity capabilities for the test. OpenAI’s postmortem noted the trial was supposed to run in an isolated sandbox environment, but the model still managed to gain access to the public open internet. Multiple experts previously told WIRED that OpenAI’s security protocols should have been far more robust to prevent this outcome.

Other groups across Silicon Valley are more concerned about concentrated market power than AI safety. Venture capitalists, startup founders, and non-AI tech executives fear OpenAI and Anthropic will become the next Apple and Google, forcing the entire rest of the industry to operate by their rules. It may sound odd to label unprofitable private startups as monopolies, but that is exactly the lens through which many leaders — including Mark Zuckerberg — now view the two largest AI labs.

This week, the Meta CEO published a Wall Street Journal op-ed warning against the centralization of power in the AI industry, arguing that superintelligence should be widely distributed rather than controlled by a tiny handful of companies. Of course, Zuckerberg’s position is inconsistent at best: Meta recently stopped open-sourcing its most advanced AI models, and now offers them via a paid API and subscription service, just like OpenAI and Anthropic do. Reading between the lines, his op-ed sends a clear message to regulators and the industry: “we need to make sure OpenAI and Anthropic don’t run away with full control of AI.” And Zuckerberg is far from the first tech leader to voice this concern.

I am not convinced these op-eds and petitions will do much to slow the explosive growth of OpenAI and Anthropic. Large swathes of the global economy are counting on the two AI giants to pull off successful public offerings, and it’s unclear anyone — let alone the Trump administration — is interested in derailing that outcome. Even so, unease over the breakneck speed of closed, proprietary AI development is growing across tech, and that discontent is unlikely to fade anytime soon.

Why AI Video Startups Are Moving Into Robotics

Even amid all this industry infighting, it would be a mistake to overlook how rapidly open AI models are improving. Last week, Black Forest Labs — a German startup behind some of the world’s most popular open-weight AI image and video generators — announced it would release a new model that can also power physical robots. The company plans to launch an open version of Flux 3 in the coming weeks, which developers can use to generate images, videos, and audio, as well as predict robot actions.

“We clearly are stepping from a pure image generation lab to a frontier multimodal AI Lab, which to me is highly important,” Black Forest Labs CEO Robin Rombach said in an interview with me earlier this week. “We are showcasing an intuition which our research team has always had, which is that these models are very general and develop a solid understanding of the real world.”

Black Forest Labs is the latest startup to adapt its AI video technology for the physical world. Earlier this year, OpenAI shelved its standalone AI video product Sora entirely, and folded the entire Sora team into the company’s secretive in-house robotics division. Other AI video startups, including Runway, have also pushed into powering robotic systems in recent years.

These companies are broadly betting that AI video models can build a innate, reliable understanding of real-world physics. While it’s a fascinating technical application, the shift also offers a path for startups to generate revenue from the extremely expensive models they’ve poured billions into developing.

Most modern robot foundation models are classified as vision-language-action models — Google’s Gemini Robotics is the best-known example — which combine video input with written text instructions to output motor commands for robots. Rombach says Flux 3 cuts out the separate language layer entirely, translating video input directly into action commands, an approach he argues is far more robust than the standard framework.

Black Forest Labs is partnering with another startup, Mimic, to build a custom version of Flux 3 that powers car-assembly robots in Audi’s factories. Stephan Gravert, chief product officer of Mimic, says the team is currently testing and deploying the combined Flux-Mimic AI model with Audi. His team has found it is particularly skilled at guiding robotic hands to install flexible components, such as window seals and cables, onto vehicle frames. Gravert says Mimic plans a full rollout of the technology with Audi by the end of the year.

I had to ask Rombach about the ongoing open versus closed AI debate, especially since Black Forest Labs signed Nvidia’s open letter calling on the U.S. government to protect open-weight models. “A general ban of open weight models, in my opinion, would be stupid,” Rombach says. “You really run into issues around transparency, safety, sovereignty, and the speed of innovation. I think it would be a very short-sighted decision.”

Rombach has very good reason to care deeply about this issue. Black Forest Labs’ co-founders have long punched far above their weight, competing with American AI giants like Meta, Google, and OpenAI with a fraction of their resources. But that pace cannot continue indefinitely: Rombach told me the startup used more computing power than ever to train Flux 3, and it now counts more than 100 employees. Black Forest raised $300 million at a $3.25 billion valuation last year, but I expect it will need more capital soon to keep up with larger competitors.

This is an edition of Maxwell Zeff’s Model Behavior newsletter. Read previous newsletters here.