Radar Telegram alerts Open Radar

OpenAI's chief scientist wants every AI lab to agree on a line that none of them cross.

תרגם לעברית

aipost · · Source

OpenAI's chief scientist wants every AI lab to agree on a line that none of them cross.

OpenAI's chief scientist wants every AI lab to agree on a line that none of them cross.

Jakub Pachucki suggests common limits that labs will not cross until the ability to understand and monitor AI behavior reaches a defined milestone. As a first step, OpenAI has published its internal benchmark of how well each of its underlying models has been predicting text since GPT-3.5.

lower is better. Below 0.8, cyber security and math skills took off. At about 0.76 the models began to produce breakthrough mathematical results, a level that the table labels "solved a mathematical problem in the Millennium Prize". An unreleased internal model already sits lower than the GPT-6.2 Astra.

Pachucki expects more "transformative capabilities" around 0.68–0.7. "I don't know what those capabilities will be, and we may not test them today. "Artificial intelligence-driven research may lead to an avalanche of breakthroughs in training as early as 2027, he says, and it is likely that safety work will not keep pace.

Sam Altman called it the most important thing OpenAI has released in a while.

More on Radar

Want to see every report in real time?

Hundreds of sources worldwide, translated into your language, with alerts for what matters to you.

Sign up free
Share on WhatsApp Get alerts on Telegram