Anthropic finds a small workspace that holds a model's unspoken reasoning

research

Anthropic finds a small workspace that holds a model's unspoken reasoning¹

A sixteen-author Anthropic interpretability team reports that a language model keeps a small, privileged set of representations carrying some of the functional hallmarks of conscious access, a region they call the J-space.¹ It holds coherent content only in a middle band of layers and carries on the order of tens of concepts at a time. The model can report what sits there, deliberately summon and hold a concept, and use the space to carry the intermediate steps of silent reasoning. Post-training, the authors find, installs the Assistant's point of view into it.¹

The paper is careful with the word the headline wants. It offers an analogy to global-workspace theory, the cognitive-science picture of a central blackboard broadcast to otherwise separate processes, and claims some of the functional hallmarks of conscious access, not consciousness.¹ The result is narrower. Decoding the space surfaces strategic deliberation, evaluation awareness, and trained-in misaligned dispositions that never appear in the model's outputs.¹ Cognition a model keeps out of its text now has a location an auditor can read, which watching outputs alone cannot give.

research

Frontier models top a new MBA-case benchmark only under its lenient grader²

Wharton, Harvard Business School, and Carnegie Mellon researchers built BusinessCaseBench, 615 questions drawn from 238 teaching cases across eighteen business disciplines, each graded against an expert instructor's rubric.² Under its Standard scoring the frontier looks saturated, with Claude Sonnet 4.6 leading at 88.4 percent.²

The grader makes the number. Switch to the paper's stricter Complete Answer metric, which requires every part of a response to be correct, and the same models fall hard: Operations and Service Management drops to 25.8 percent.² "Frontier AI passes business school" is a headline the lenient rubric manufactured. On identical questions, these models draft a plausible case answer and rarely a complete, correct one.

engineering

An AI chains a WordPress RCE that was disclosed and patched days earlier³

Adam Kues of Searchlight Cyber's Assetnote team published an account of using GPT-5.6 Sol Ultra to build a pre-authentication remote-code-execution chain in WordPress core, under the headline that exploit brokers pay $500,000 for such a bug and this one cost about $25.³ The model ran roughly ten hours of vulnerability research, chaining a batch-API validation desync, a recursive bypass, a SQL-injection sink, and a post-exploitation sequence, at half of a week's usage on a $200 subscription. Kues frames the win as speed: "no security researcher could have found and completed this exploit chain in 10 hours without AI."³

The $25-against-$500,000 contrast drops what independent reporting establishes. Three days earlier, before the methodology post ran, The Hacker News covered the same flaw as ordinary security research: two CVEs, one for the batch-route confusion and one for a SQL injection that three other researchers reported separately, already patched in WordPress 6.8.6, 6.9.5, and 7.0.2, with a public proof-of-concept on GitHub and no mention of AI.⁴ The vulnerability was human-disclosed and already fixed when the AI account appeared. What the run measures is the time to chain a known-solvable bug, with a human adapting the prompt, building the test environment, and verifying the chain on a live instance. The model compressed the exploit-chaining, not the discovery.

compute

Moonshot stops taking new Kimi K3 subscribers, out of GPUs to serve them⁵

Moonshot AI stopped accepting new Kimi K3 subscribers on July 20, saying its GPUs "are feeling it" and that over the past 48 hours demand had "pushed close to the limits of our current capacity."⁵ It is pausing new subscriptions to protect existing subscribers and splitting membership into a general tier and a dedicated Kimi Code plan to ration compute.⁵ Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model, 16 of 896 experts firing per token, with a one-million-token context and weights due to open July 27.⁵⁶

Launch week sold a benchmark: Kimi K3 topping Fable 5 on a frontend-coding leaderboard. Moonshot itself concedes the model trails Fable 5 and GPT-5.6 Sol overall, and users report it runs noticeably slower than the top US systems, a gap the South China Morning Post ties to the compute shortage Chinese labs face under US export controls.⁶ Alibaba moved on the opening, pricing a discounted preview of Qwen 3.8 against the outage.⁷ A benchmark win a lab cannot serve for a week is not a shipped product.

industry

Anthropic keeps Fable 5 permanently, but caps it at half-limits and shuts out the $20 plan⁸

Anthropic said that beginning July 20, Claude Fable 5 "will be included in all Max and Team Premium plans, at 50% of limits," with Pro and Team Standard users keeping access through usage credits plus a one-time $100 credit.⁸ The post reads as a reprieve. It is the fourth time the date has moved: a planned removal on July 7 slid to July 12, then July 19, and now to permanent inclusion, each slip blamed on capacity.⁸ The permanence carries a cap at half the usual limits and shuts the $20 Pro plan out of included access entirely.

Simon Willison reads the reversal as forced: competition from GPT-5.6 Sol, and to a lesser degree Kimi 3, made Anthropic's plan to retire Fable 5 untenable, since a $100 or $200 subscription that excludes the lab's best model is a hard sell.⁹ He expects, as his own speculation, that Anthropic will "have to dial back their training efforts in order to make more GPUs available to help serve the model."⁹

Moonshot in Beijing and Anthropic in San Francisco both rationed access to their best model in the same week, one pausing signups to serve current users, the other capping limits to keep training. On both sides of the US export-control line, the split of GPUs between serving and training is now the visible ceiling on how widely a frontier model can be deployed.