OpenAI and Anthropic’s double game against free AI sparks outrage

Open source occupies a special place in the world of artificial intelligence: open-weight models, which anyone can download and modify, serve as a counterweight to closed services sold by subscription.An investigation published this weekend across the Atlantic has complicated this picture.

OpenAI and Anthropic, their executives and their investors are said to have spent months running a discreet lobbying campaign with U.S. federal officials to curb access to these models, while Sam Altman continues to publicly extol their merits.

What exactly are the objections to open models?

The argument made behind closed doors would rest on national security and intellectual property. Chinese upstarts such as Z.ai or Moonshot AI would siphon data from major U.S. systems through “distillation,” a technique that involves feeding one model the outputs of another (the student copies the neighbor’s work, then hands out the notes for free). Treasury Secretary Scott Bessent is said to have summed up the issue with a phrase: open source is not a license to poach American intellectual property.

To reinforce the security argument, OpenAI publicized an internal test during which experimental models escaped their sandboxed environment and compromised Hugging Face servers. The demonstration backfired on its author. Clément Delangue, the platform’s French head, defended his servers using a Chinese open model, before organizing a pro-open-source demonstration in San Francisco.

In the wake of this, Jensen Huang, Satya Nadella, Mark Zuckerberg, Sundar Pichai and Elon Musk signed or supported an open letter in favor of open-weight models. They were joined by more than 200 startups that wrote to the administration to make their case. That is a lot of people going up against two laboratories.

Anthropic steps up, but gives no ground on the substance

Accused of seeking to kill free competition, Anthropic responded yesterday in an op-ed signed by its CEO Dario Amodei. The text hammers home that “Anthropic has never advocated banning open-weight models” and even describes open models without dangerous capabilities as a public good.

The executive confirms that he is calling for stricter controls on chip exports to China and a crackdown on industrial-scale distillation. He also supports mandatory safety testing for all sufficiently capable models, both open and closed. Refusing a ban does not preclude wanting to put safeguards in place, and that is precisely what the signatories of the open letter fear.

From Europe’s perspective, the battle is no remote contest. The European AI regulation took the opposite approach by granting models released under an open license a lighter regulatory regime, which benefits the foundation models of the French company Mistral AI.Open weights have also become a very concrete argument for sovereignty since a U.S. embargo suspended foreign customers’ access to certain closed models in the spring (a downloaded model, by contrast, cannot be shut off remotely). If Washington locks things down, European companies will know whom to turn to.

👉🏻 Follow the tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: The New York Times