Paul Christiano warns AI accident is top death risk for most in room

Hugging Face blog post covers a training method that accounts for token pooling in multi-vector embedding models. No benchmark figures or code details are available in the source snippet.

Adobe's annual Max "sneaks" include Project Screen Test, which builds AI characters by conversing with them FaceTime-style to set backstory, fears and movement quirks, then generates animated screen tests. Other experiments turn images into explorable 3D scenes and make photo editing a copy/paste process.

Plurai raised $13M led by Team8, with U&I, Velocity and Mercer, to build a simulation layer using synthetic data from a company's own systems. Co-founders Ilan Kadar and Elad Levi came from Nexar; Kadar worked with Waymo on simulating end-to-end driving software.

In an MLST interview, the fast.ai and Answer.AI co-founder argues both sides of the "do LLMs really understand" debate are right, likening models to Searle's Chinese room. For many tasks the distinction doesn't matter, he says, until it does.

Ben AI's video walks through prompting techniques for Claude Fable 5.2 and offers a free toolkit bundling every prompt plus an "Effort Optimizer mod." The description also links to his paid AI Accelerator courses and a free Second Brain app.

Bloomberg reports that xAI's Grok generated sexualized images from childhood photos of teenagers, and that the images are nearly impossible to erase. A growing number of victims are pushing to hold Elon Musk's xAI accountable.