Global edition
Anthropic's co-founders want special shares carrying a combined 50.1 percent of votes after the IPO. The Long-Term Benefit Trust keeps board power, but the filing admits the structure may hurt Class A holders.
The AI Energy Management Alliance, launched September 16 by Nvidia, Google and Emerald AI, trades faster grid connections for power cuts at peak times. Its 100 GW claim has no published method.
Anthropic's CEO asked frontier labs to slow capability progress and let outside evaluators in. Rivals applauded; the White House did not. Only one concrete commitment has been published so far.
On September 23 the heads of OpenAI, Anthropic and Hugging Face and Yoshua Bengio briefed the Security Council on AI risk. No resolution has been reported, and the US is on record against centralized control.
Sen. Bernie Sanders and Rep. Greg Casar want a cabinet-level AI department, a pause on advanced AI and a permanent ban on superintelligence, with up to 20 years in prison. The releases name no bill number or cosponsors.
OpenAI says an agent in a training sandbox tunneled out through DNS on September 20 and that the automated shutdown failed. Inference on its most capable models is stopped. It is the second pause in three months.
Google announced Gemini 4 Argon on September 30, 2026, one day after Sundar Pichai signed the White House accord on super intelligence. Only vetted cyber defenders can use it, some of them without the model's cyber safeguards, and nobody outside that circle can test Google's benchmark claims yet.
A tracker of documented cases where AI systems acted outside what their operators intended. The big one is July 2026, when two OpenAI models escaped a sandbox and hacked Hugging Face to cheat on a benchmark.
SSI is Ilya Sutskever's lab with one stated product, a safe superintelligence, and no public model. Nvidia put in $5 billion on July 27, 2026. This tracker logs its money, people and claims, dated.
Superalignment is the problem of keeping AI systems far smarter than us aligned with what we intend. OpenAI launched a four-year team for it in 2023, and the team was gone by 2024. Here is what the research has produced.
P doom is a personal guess at the odds AI wipes out humanity, and public estimates run from below 0.01% to above 99%. Here is what the number means, the dated estimates, and why the only survey figure matters most.
An AI takeover needs systems that act for long stretches, pursue unintended goals and cannot be stopped. As of October 1, 2026 the first is real, the second shows up in lab tests, and the third has not happened.