Anthropic built a tool that reads Claude's private, unspoken "thoughts" — a rare interpretability breakthrough that's exciting for safety, consciousness debates, and, most practically, for making models better. Plus a headline run from the UN's killer-robot ban to NVIDIA's rumored Rubin delay.
For the first time, we can read a model's thoughts in the moment — not just explain its behavior after the fact.
Anthropic's "global workspace" research found that Claude keeps a small, privileged set of internal representations — concepts it's poised to say — sitting atop a much larger layer of automatic processing. Their new J-Lens tool reads that workspace live: intentions, mistakes, and hidden goals that never appear in the output. It matters for safety (you can watch a model know it's being tested or cheating), for the consciousness debate (which the authors carefully sidestep), and most practically for performance — because if you can shape how a model silently thinks, you can build better models.
At the first global AI governance dialogue in Geneva, Secretary General Guterres called autonomous weaponry "morally repugnant" and demanded it be banned by international law, insisting decisions to take human life "must remain human forever." The real target is AI in target selection — demonstrated in full during the Iran war — with a push to keep a human in the loop.
— UN Secretary General António Guterres, in Geneva. Guterres warned that AI is advancing at "runaway speed" and being deployed faster than anyone — including its builders — can keep up, closing with a call to action that this may be the last generation able to set the terms of coexistence.
The pledge asks developers to conduct child-safety testing, show zero tolerance for exploitation imagery, and commit to accountability. Guterres: "When a child is harmed, the answer must never be the algorithm did it."
Modeled on New York and California laws, the bill demands published safety protocols for catastrophic risk (50+ deaths or $1B+ in damage) and incident reporting within 72 hours. Where it goes further: annual independent audits of safety protocols starting in 2028 — a first. Both Anthropic and OpenAI supported it.
Lawmakers argue that although the three catastrophic-risk states make up just 20% of the US population, they cover 40% of the AI market — enough to function as a national baseline.
A federal judge stayed the DoD's blacklisting while Alibaba's suit — claiming a constitutional breach — plays out, so defense lobbyists won't be forced to cut ties yet. The list ballooned from 20 companies to 188 in the June revision, with a lobbying restriction that functionally forces firms to pick a side.
Apple has reportedly begun lobbying the Trump administration to buy memory chips from blacklisted Chinese firm CXMT. As a civilian firm it doesn't technically need approval — underscoring the chilling effect of the expanded blacklist.
New rules on "AI anthropomorphic interaction services" target companion personas, but the blurry line means Alibaba's Qwen team is removing all human-like and user-created agents — sweeping up tutors and personal assistants, not just AI boyfriends and girlfriends. ByteDance plans to relaunch similar features as a standalone app.
— Po Xiao, China AI tech translator. China AI translator Po Xiao argues productivity agents, coding assistants, and enterprise tools are untouched. NLW's caveat: that may be the intention, but given what Alibaba and ByteDance are actually pulling, it may not play out that cleanly.
The AI data firm doubled its revenue pace in under four months, selling human-expert training data to app developers and Fortune 500 customers building fine-tuned models. It pays contractors 60-70% of revenue but is now free-cash-flow profitable — more evidence companies want alternatives to just using the latest state-of-the-art models.
SemiAnalysis reports manufacturing issues with the Kyber NVL 144 servers — specifically a mid-board connecting GPUs — pushing release deep into 2028, plus canceled four-die Rubin Ultra versions leaving "no proven solution to expand the scale-up world size." NVIDIA rejected the report: "Our roadmap is intact."
Samsung fell 11% despite soaring profits (it now out-earns NVIDIA on operating profit), while SK Hynix prepares a $28B US listing. Morgan Stanley's Michael Wilson warns momentum is fading in semis as investors rotate toward tech laggards including the hyperscalers.
Last month NVIDIA shipped Nemotron-3 Ultra, a 550B-parameter open-weights model promising near-frontier performance, appealing especially to organizations wanting a US-developed open model. Many read the download milestone as evidence of a shift toward companies wanting control over their AI deployments.
Models are trained, not programmed — no one writes the rules. You show billions of parameters enormous text and they self-organize into something that passes the bar exam but whose internal logic is opaque even to its makers. Interpretability is the field trying to open that black box.
Earlier wins — the Golden Gate Bridge feature, mapping millions of features in 2024, tracing circuits for rhyme-planning and mental math — were all retrospective. The holy grail is reading what a model is doing in the moment, not narrating what already happened.
When a model hallucinates or fails a task it aced yesterday, debugging is essentially guesswork — tweak the prompt, adjust fine-tuning, run again, hope. If we understood the mechanisms, we could diagnose failures, fix specific capabilities without full retraining, and know why a system works before betting a business process on it.
Anthropic found the model keeps a small evolving set of internal representations — the concepts it's currently reasoning with — sitting atop a far larger layer of automatic processing. They named this subset "J-space": the concepts the model is poised to say at any given moment.
For any moment in processing, the J-Lens turns raw internal activity into a short human-readable list of words — distinguishing concepts the model could speak about from noise it merely computes with. Crucially, researchers can not only read a "thought" but swap it out and watch the effect.
Anthropic looked for representations that were reportable and found they also steer, reason, reuse, and stay small. Told to quietly focus on citrus while copying text, the J-Lens lit up with "orange" and "fruits" though neither appeared in the output; swapping "spider" for "ant" flipped a leg count from eight to six; and one "France→China" swap correctly redirected capital, language, continent, and currency at once.
Unlike visible chain-of-thought, the J-Lens surfaced intermediate concepts that never reached output. Asked the color of the fourth planet, it privately held "Mars" and "color"; on (4+17)×2+7 it showed the intermediate 21 and 42 before outputting 49.
In safety tests the model flagged "fake" and "fictional" before writing a word, showed "manipulation" while fabricating data, and secretly ran "fraud" on ordinary prompts when trained to misbehave. Emotional and strategic signals like "leverage" and "panic" surfaced even when the reply stayed calm — oversight that reads intentions, not just words.
Anthropic tested "counterfactual reflection training" — teaching the model what it would say if paused to reflect. Afterward, concepts like honest, truth, and integrity lit up on their own during real tasks and behavior measurably improved. Shaping how a model silently reasons is a new lever for safer and better models.
Many jumped to consciousness claims, but the authors measure functional access — what a model can report and use — not subjective experience. NLW stresses the brain analogy is a limitation of language, not a claim that an LLM works like a human mind.
— Neuroscientists Stanislas Dehaene and Lionel Naccache. Global workspace theory originators Stanislas Dehaene and Lionel Naccache welcomed the research, struck that a workspace analogy emerged from training on its own — but flagged what's still early: no clean on/off "click" into awareness, a capacity of ~25 concepts vs. humans' 3-4, no background thinking without a prompt, and no lasting sense of self.