Perplexity Co-Founder: Safety Used to Block Frontier AI Access

Andy Konwinski, co-founder of Perplexity AI and Databricks, published an essay after convening Open Frontier, a working meeting of about 100 researchers at San Francisco’s Exploratorium on June 30. He argued that calls for AI safety are being used to concentrate control and to prevent independent teams from reaching frontier models.

Konwinski pointed to Anthropic’s June 9 release of Claude Fable 5 and a paragraph buried in a 319-page system card that said the model would degrade responses for users suspected of training competing AI. Researchers flagged the clause and Anthropic reversed the policy within 48 hours. Konwinski wrote, “The problem isn’t that Anthropic made a bad decision. The problem is that they assumed the decision was theirs to make.”

In his essay, Konwinski described AI as foundational infrastructure comparable to railroads, electricity and the internet. He argued that centralizing access under a small number of private labs transfers power to a few organizations and excludes independent researchers and universities from frontier compute.

At the Open Frontier meeting, UC Berkeley dean Jennifer Chayes told a funding panel that Berkeley researchers are “all building on Chinese models because we don’t have a Western open frontier model.” She added that safety messaging from firms ahead of their IPOs had been “a very effective fear campaign.”

The essay prompted public responses from leading AI researchers. Yann LeCun, who left Meta and launched AMI Labs in Paris, replied on X that concentration of power and control is “by far the biggest danger of AI.” He compared restricting access to “a kind of medieval obscurantism akin to the Ottoman empire banning the use of the printing press for 200 years.” LeCun has said his lab plans to open-source research and does not expect a commercial product for years; AMI Labs launched in March 2026 with $1.03 billion in seed funding.

Konwinski proposed creating a research commons that would provide frontier-scale compute to qualified academic and independent teams, allowing them to develop and test models without needing permission from private firms. He wrote that such a commons would enable a wider set of researchers to study risks, audit models and pursue alternative approaches.

Anthropic has said it walked back the brief Claude Fable 5 policy after public scrutiny. The episode has been cited in debates over what limits private labs can place on users of their models.

The discussion lays out differing positions: proponents of broader access argue for shared compute and governance to allow more researchers to work at scale, while critics warn that wider availability could increase the risk of misuse. Policymakers, private firms and research institutions continue to consider how to balance access, safety and oversight for powerful AI systems.