MIT Tech Review
8/10 signal
The Download: Claude’s inner workings and OpenAI’s “super app”
reasoningeval
What happened
Anthropic researchers used a tool called the Jacobian lens (J-lens) to discover 'J-space' inside Claude. This hidden space contains concepts and words the model processes and 'puzzles over' before deciding on its final output, offering a rare look at internal model reasoning. The newsletter also notes the launch of OpenAI's 'ChatGPT Work' super app, GPT-5.6 models, and an automated researcher agent.
Why it matters
The discovery of J-space provides a concrete window into the internal reasoning and 'latent thoughts' of frontier LLMs before they generate text.
The take
The Jacobian lens is a massive step forward in mechanistic interpretability. Understanding 'J-space' could eventually allow developers to build better evaluation tools, detect hallucinations early, or steer model reasoning mid-generation. The simultaneous mention of GPT-5.6 and automated researchers signals rapid progress in agentic capabilities.
Do this
Read Anthropic's research paper on the Jacobian lens to understand how mechanistic interpretability is evolving to expose latent model states.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.