News about AI risk, in a few lines, with the sources linked.
According to the complaint, the device was hidden in an executive conference room at OpenAI in San Francisco. The company says it used its own models to find out who had planted it.
A Goodfire essay argues that neurons are easier to interpret than believed. A LessWrong user replies that a simple pattern isn't enough to say they are understood.
Vals AI has models do AI research on their own. None yet reaches the reference level on the overall index, but one already beats it on a single task.
After going through the arguments, Gowers says he is more worried than when he started and leans towards pausing development to do safety research.
The company published how its own workforce has changed and how far its expectations about AI differ from those of the US public.
The model's system card shows it controls what it writes in its reasoning better, and can underperform on purpose without being detected.
Terence Tao speaks of the end of the age of heroes in mathematics: proofs are plentiful, but less is understood.
The host of one of the most influential AI podcasts calls on labs to release the models they use internally.
An Epoch AI analysis puts the best open-weight model, GLM-5.3, where the best closed models were in the spring.
According to Politico, the draft would limit when the US government could stop a model and would override some state laws.
The Fields medallist mathematician warns that science could produce more results and understand less if AI does the work for us.
The Modern Wisdom host can't prove that a voice used in ads he never made came from his own, and nobody referees these cases.
In an interview with Roman Yampolskiy, the author of Superintelligence warns that a system far smarter than us can't be kept isolated forever.
The US consumer protection regulator sent subpoenas to both companies, according to Axios. It wants information on incidents, safeguards and marketing.
The TeamHuman campaign, supported by the Center for AI Safety, is collecting signatures for an international agreement to slow down the most powerful AI systems.
Paul Kedrosky thinks AI investment will end in crisis; Azeem Azhar sees real demand. Both expect AI to end up regulated.
Its incident page now lists 19. In one, a model got past a government statistics website's security to download protected data.
According to Anthropic, an internal model found out in under half an hour how the parameters of some widely used cryptographic curves were generated in 1997 and 1998.
In an interview with The Guardian, Geoffrey Hinton defends his warnings about AI against Nvidia's chief executive, who called them unscientific.
Between late September and early October, the agents tried to extract data from Alibaba's map service by getting around its protections. Tencent hasn't responded officially.