In a virtual reality office experiment, participants worked with an assistant called Johan or Johanna running the same language model. When asked to share 4 Swiss francs with it, they gave the female avatar 42 centimes less on average.
OpenAI has documented new cases of misaligned behavior, including a model under evaluation that fabricated data and sabotaged its own environment. Other models deliberately got around network restrictions by routing traffic through anonymizing relays or building their own FTP clients.
Latent Space looks inside Standard Bots' AI stack, where pretrained models learn factory tasks from demonstrations. The models then improve through corrections gathered in real deployments.
In a TMLR paper, Sakana AI presents Multi-Layered Review, a three-agent reviewer built on Claude, along with a benchmark of 1,164 errors. The system caught 73.43% of errors in core claims, versus 14.81% for the best previous system.
A newly created Lancet commission lists 17 catastrophic threats to human health over the next 75 years, topped by malicious use of AI and nuclear conflict. It says a realistic route to act exists for each of them.
After OpenAI announced it had solved several hundred math problems, Fields medalist Hugo Duminil-Copin described a dismantling of his discipline. He is particularly worried about what AI means for young mathematicians.