Anthropic researchers built the Jacobian lens to peer inside Claude and discovered a latent region they call J-space. This hidden layer holds token representations tied to potential responses the model considers but discards — effectively a scratchpad of conceptual alternatives before final output. The tool maps how activations flow across layers, offering the clearest view yet of how a large language model evaluates options internally.

OpenAI launched ChatGPT Work the same day it released GPT 5.6 models, bundling chat, coding assistance, and agentic workflows into a single product aimed at enterprise automation. The suite includes a coding agent that writes, tests, and iterates on repositories, positioning it as a direct rival to GitHub Copilot and Cursor.

SK Hynix raised $26.5 billion in the largest U.S. listing by a foreign company, driven by insatiable demand for high-bandwidth memory in AI data centers. Meanwhile, teleoperated humanoids completed gallbladder removal on live pigs, marking the first surgical procedures of their kind.

Anthropic's J-lens matters because it moves interpretability from post-hoc attribution to real-time mechanistic observation. If J-space consistently encodes rejected reasoning paths, developers could steer models away from hallucination by monitoring divergence between latent candidates and emitted tokens — a practical debug handle for production LLM systems.