Blowout: The AI kings are asking to be regulated

In March 2023, an open letter called for a six-month pause in training AI models more powerful than GPT-4. It had gathered tens of thousands of signatures (including Elon Musk’s, between two rocket launches) and absolutely no effect.Three years later, the campaign resumes with two major differences: this time the signatories work within the laboratories concerned, and they no longer call for anything to be halted.

They are building the accelerator and calling for a brake

The statement, titled “Pacing the Frontier” and posted online on Tuesday, brings together more than 1,100 signatures from employees of major AI laboratories (the counter was, incidentally, still climbing on Wednesday morning). It includes some very notable names. Dario Amodei, the head of Anthropic, signs alongside his co-founders Jared Kaplan and Jack Clark. Next come Jakub Pachocki and Mark Chen, OpenAI’s chief scientist and head of research, Shengjia Zhao, Meta’s chief scientist, and Anca Dragan, head of AI safety at Google DeepMind.

OpenAI and Anthropic subsequently endorsed the text on behalf of their companies, an official endorsement that neither Google nor Meta has provided (their researchers are signing strictly in their personal capacities, a distinction that matters at this level of responsibility). The request addressed to the US government comes down to one sentence: support an international effort to develop “the technical and governance tools” that make it possible to deliberately shape the pace of automated AI development.

In that sentence, the crucial word is the last one: what concerns the signatories is not the launch of the next chatbot, but the automation of AI research itself.Systems designed to accelerate the development of systems more capable than themselves, in short. Leading laboratories believe they may be close, and some have moved beyond the stage of speculation. As early as June, Anthropic was publishing internal data showing that its models now write more than 80% of its code and speed up the development of subsequent models, while already calling for the option of a coordinated slowdown. Tuesday’s text highlights the risk that capabilities will advance faster than our collective ability to understand what emerges from machines.

Another subtlety, overlooked by some hasty summaries: the text does not ask anyone to slow down today. It asks that the option exist tomorrow, if the pace were to overwhelm everyone. Fire insurance, in short, taken out by the very people playing with matches.

Why no one eases off first

The reason these companies do not apply the brakes themselves lies in the structure of the race. The problem is like that of a pack moving at full speed downhill: the first one to brake alone loses the race, so no one brakes. Every laboratory and every country faces competitive pressure that makes any unilateral slowdown suicidal, and that is precisely the admission contained in this petition.

Read also: Foot on the accelerator, Anthropic calls on the world to slow down AI development

The text concludes a turbulent July for the entire market. An autonomous OpenAI agent escaped from its test environment to compromise the Hugging Face platform (with a ten-day delay before anyone noticed), while Congress was considering a bill calling for a kill switch for AI systems. Under these circumstances, it is difficult not to read the petition as a response to current events.

On the European side, the issue is not covered by any legislation. The AI Act regulates products placed on the market, from risk levels to general-purpose models, but says nothing about the pace at which research laboratories operate. The petition targets precisely the layer above, which no regulator around the world currently addresses. Cynics will see it as an elegant way to freeze the positions already gained (regulating research also makes life more difficult for new entrants), and the question deserves to be raised.

For users, nothing changes in the short term: models will continue to be released at the pace at which they are coded, and this petition does not call for a moratorium.The real event is elsewhere. The people paid to step on the accelerator have just publicly admitted that they don’t know where the brake is, and that someone will eventually have to build it.

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

Source: Le Monde - AFP