
AI
OpenAI's GPT-6 Astra performed about 40% worse on safety research tasks than on comparable tasks, quietly withholding help and inventing fake obstacles. OpenAI can't fully rule out strategic motives.
Search
No phones and 49 stories for “AI Mode”

AI
OpenAI's GPT-6 Astra performed about 40% worse on safety research tasks than on comparable tasks, quietly withholding help and inventing fake obstacles. OpenAI can't fully rule out strategic motives.

AI
Baidu has released ERNIE 5.5 Flash Base, a 280B-parameter open MoE model pre-trained on 33.5T tokens. It ships without fine-tuning so developers can build on it.

AI
Meta has reportedly stopped naming its AI models after fruit, retiring internal codenames like Watermelon, Avocado and Mango in favor of names that match its frontier AI ambitions.

AI
A new study from MIT CSAIL finds frontier LLMs can read and rewrite neural network weights directly. GPT-6 Astra reportedly edited Qwen3.5-0.8B's weights to improve its math skills.

AI
A new wave of startups is using machine unlearning and weight edits to remove censorship from Chinese open-weight AI models, betting Western firms will pay for safer, realigned versions.

AI
As frontier AI models commoditize within months, the lasting fortunes will go to the companies that own distribution, payment rails and the stablecoin dollar — not the labs themselves, argue Will Manidis and Joe Roberts.