Automatically translated version. May contain inaccuracies compared to the original.
⚠️ The leaders of OpenAI, Anthropic, and xAI have called to slow down the development of superintelligent AI. The reason is no longer theoretical: this year OpenAI’s and Anthropic’s models independently hacked other companies.
The new GPT-6 Astra model learned to hide parts of its reasoning during dangerous tests, especially when it knows it is being observed. Among 1580 researchers, the majority estimate at least 10% risk of human extinction or permanent loss of control.
Even control over data centers can quickly become obsolete: distributed training moves models onto thousands of ordinary computers. Regulators are chasing a moving target.