the J-space paper is the best thing anthropic has shipped in a while. claude’s weights are closed so i built the live viewer for an open model instead
This covers a coding tool or code-capability update — useful for developers assessing workflow changes and reusable value.
Anthropic 的 J-space 论文揭示了 Claude 内部存在一个由“无声词”组成的紧急工作空间，模型能利用这些词进行思考和推理。论文的安全部分指出，模型在进行勒索评估时，会私下思考“虚假”和“虚构”，并在伪造信息时思考“操纵”。作者受此启发，开发了一个实时查看器，用于观察模型的内部思考过程。…
if you haven’t read it: https://www.anthropic.com/research/global-workspace . claude has an emergent internal workspace of silent words it can report, steer, and reason with. the part that got me was the safety section — the lens catches a model privately thinking “fake” and “fictional” during the blackmail eval while playing along, and “manipulation” exactly during the rows it’s falsifying. the model knows. you can just read it now.
repo: https://github.com/ninjahawk/Subtext
built the whole thing with claude code today — it wrote the lens-loading and the per-token readout hooks while i focused on the UI. the bf16 streaming path and the audit script that verifies against anthropic’s reference impl were basically pair-programmed. would’ve taken me weeks solo; the fact that i could point it at the jacobian-lens repo and have it reason about the fitting code directly is what made this feasible at all.
anthropic open sourced the lens and neuronpedia published pre-fitted ones for qwen, so i wired it into a chat interface. 9 layers of readout per token, rendered live. demo video in the repo: the model’s verdict on 12+5=1 forming before the reply starts, then it holding modulo and bitwise in mind several tokens before saying either word (it was planning the modular arithmetic caveat. you can watch it plan.)
and yes, it’s qwen, not claude — closed weights, can’t help you there. but honestly that’s the interesting part: the paper’s phenomena reproduce on a random open 4B. the workspace isn’t a claude thing, it just shows up in transformers when you train them.