CERN and Mistral AI want to teach an AI model to read particle collisions from the Large Hadron Collider. Billed as a world first, the project could change how physicists explore the subatomic world.
Researchers are using AI to decipher, transcribe and cross-reference documents that long remained out of reach. Beyond saving time, the technology is opening new avenues for studying the past.
A new essay argues that neural networks may be far easier to read than assumed, with about 90 percent of neurons serving a clear function. If true, deception or dangerous behaviour could be spotted directly in a model's circuits, but researchers are still debating whether the thesis holds.
A preprint shows that an agent can trigger a tool before the model has even voiced a refusal, revealing a gap between what the model says and what the orchestrator actually executes. The study also lists 253 verified bugs across 118 MCP servers, arguing that actions must be checked independently of model output.
Internal notes from Microsoft and OpenAI acknowledge that by diverting traffic away from websites, AI is drying up the human-written content it depends on. That shortfall could in turn threaten the training of future models.
Latent Space's newsletter highlights major claims from Nick Cammarata that agents have disproved polysemanticity and substantially advanced interpretability and alignment. The headline frames these as claims rather than established results.