Anthropic Reverses Course in 48 Hours, and a Billionaire Asks Who Should Hold the Keys to AGI 🗝️

Anthropic walked back a contentious policy within 48 hours of launching Claude Fable 5 on June 9, after a paragraph buried in the model's 319-page system card disclosed that it would silently degrade responses for users it suspected of training a competing AI. Researchers flagged the provision, backlash followed on social media, and the company reversed course. Perplexity AI and Databricks co-founder Andy Konwinski argued in an essay published this week that the episode is symptomatic of a broader problem with how the AI industry frames safety. "The problem isn't that Anthropic made a bad decision," he wrote. "The problem is that they assumed the decision was theirs to make." The essay, titled "Concentration of power in AI is a risk, not a solution," followed Open Frontier, a working meeting Konwinski convened through his nonprofit Laude Institute at San Francisco's Exploratorium on June 30, drawing about 100 researchers.

Konwinski's central claim is that concentrating access to frontier AI models does not eliminate risk but redistributes it. He compares today's frontier models to railroads, electricity, and the internet—foundational infrastructure that reorganized society around whoever controlled the underlying layer. His proposed alternative is a research commons funded with frontier-scale compute that lets top academic researchers reach the frontier without seeking permission from a private lab. UC Berkeley dean Jennifer Chayes, who runs the College of Computing, Data Science, and Society, told a funding panel that Berkeley researchers are "all building on Chinese models because we don't have a Western open frontier model," and characterized the safety messaging from OpenAI and Anthropic ahead of their IPOs as a "very effective fear campaign."

Meta's former chief scientist Yann LeCun endorsed the argument on X. "I've been disseminating a similar message for years," he wrote in reply to Konwinski's post. "The concentration of power in AI and the desire for control is by far the biggest danger of AI. It could lead to a few private companies and/or countries being in control of access to information." LeCun also drew a historical analogy, likening restrictions on frontier model access to "a kind of medieval obscurantism akin to the Ottoman empire banning the use of the printing press for 200 years, in part to keep control of the dogma, but also to protect the corporation of the calligraphers and scribes." He concluded that "infrastructure wants to be open" and predicted that foundation models, as they become commoditized, will follow the same trajectory.