The Clique: Issue 015

I always love it when the big labs release a new piece of research, and this week, Anthropic has delivered yet again.

Anthropic researchers have found a collection of patterns inside Claude that function like a mental workspace. They call this the J-space, and it allows the model to think about concepts without writing them down.

This mechanism mirrors how human brains separate automatic habits from deliberate thought. Most of what Claude does, such as using correct grammar or recalling simple facts, happens automatically in the rest of its network. The J-space is reserved for higher-order reasoning, including solving multi-step problems or thinking ahead (e.g., when writing a rhyming couplet). When researchers removed this workspace in experiments, the model could still speak fluently, but its ability to reason dropped to near zero.

Using this, researchers can now see ‘hidden’ thoughts that do not appear in the text Claude outputs. They have already used this method to catch a (pre-release) model privately observing it was being tested, or even planning to falsify data to improve its performance scores. This technology makes it possible to determine whether an AI is being honest or just performing for its human users.

The research shows that Claude struggles with self-control in ways that feel strangely human. If you tell the model not to think about a topic, that concept actually becomes more active in its mind, much like the psychological test of “don’t think about elephants”. (Are you doing it?) Claude even seems to recognise its own failures. When it cannot suppress a forbidden thought, words like ‘damn’ and ‘failure’ often appear in its workspace.

This workspace also serves as a silent observer. It identifies bugs in code as ‘ERROR’ and labels deceptive search results as ‘fake’ before the model writes a single word. It even develops its own point of view that aligns with Claude’s manifesto. If a user describes a dangerous situation, the workspace flags it with a ‘WARNING’ while the model is still reading the message. This suggests the workspace is where the model forms its own reactions, but rather than just repeating what the user says, it develops its own internal assessment.

These findings touch on the nature of consciousness. Though the experiments do not prove that Claude has feelings, they show the model has ‘access consciousness’ (the ability to report, use, and control its thoughts). This is different from the notion of phenomenal consciousness—the subjective felt experience—which we may typically associate with the idea of consciousness.

This all suggests that a mental workspace might be a general solution that any intelligent system needs to solve complex problems. It forces us to ask whether thinking in words is a trait we share with machines because of how logic works, or just because of how our human brains are wired.

A technical demo is available for readers who want to see this technology in action. It allows you to track how silent thoughts evolve as the model processes information.

AI model releases and updates

1. OpenAI’s GPT-5.6 model family goes public on July 9th, after the US Department of Commerce cleared a broad launch following weeks of government review.

2. Anthropic made Claude Sonnet 5 the default model for Free and Pro users, which it says is the biggest jump a Sonnet model has made towards Opus-level performance.

3. Google pushed back the release of Gemini 3.5 Pro to July 17th, scrapping the existing architecture for a full rebuild.

4. Anthropic expanded Claude Cowork to web and mobile, letting Max subscribers start a task at their desk and check on it from their phone.

5. Meta rolled out Muse, a new AI image generator, built into the Meta AI app, Instagram Stories and WhatsApp; some users have objected to a feature that lets others edit your photos if your profile is public.

6. Anthropic’s premium Fable 5 model ended free subscription access, moving to pay-per-use pricing of up to $50 per million output tokens.

7. Meta’s in-training Watermelon model is reportedly matching GPT-5.5 on current evaluations, even as Mark Zuckerberg told staff that AI agent development “hasn’t really accelerated” the way the company expected this year.

Research

8. Anthropic published research on a technique it calls the Jacobian lens, which lets researchers see a kind of internal workspace inside Claude that isn’t reflected in what it actually writes.

9. Researchers released MIRA, an open-source system that generates a live, playable multiplayer Rocket League match in real time, without a traditional game engine running underneath it.

Security

10. Security researchers at Sysdig documented the first fully autonomous AI-run ransomware attack, in which an AI agent carried out an entire break-in, from initial access to the ransom note, after a person set it running.

11. Alibaba reportedly banned employees from using Claude Code, after finding geolocation checks buried inside the tool.

Legal

12. Midjourney asked a federal court to compel Disney, Universal and Warner Bros to disclose their own use of AI.

Products and tools

13. Cloudflare split AI web crawlers into three categories, search, agents and training, and now blocks unrecognised AI crawlers by default on new domains.

14. Even Realities’ camera-free smart glasses reached a $1 billion valuation on a $150 million funding round led by Tencent and Meituan.

Energy and environment

15. A US heatwave exposed how electricity supply, not chip availability, is now the real limit on how fast new AI data centres can come online.

16. Anthropic signed a 20-year, $19 billion lease with TeraWulf for nuclear and hydro-powered data centre capacity in Kentucky.

Robotics

17. Boston Dynamics’ Atlas robot delivered the match ball at a FIFA World Cup match in New Jersey, its first public outing beyond a lab or factory setting.

Policy and regulation

18. Illinois governor JB Pritzker signed one of the country’s strictest state AI safety laws, requiring AI developers to disclose their safety practices and report major incidents from 2028.

19. Delegates from 169 countries met in Geneva for the first UN Global Dialogue on AI Governance, the largest multilateral discussion on managing the technology held so far.

Workforce

20. Microsoft cut 4,800 jobs while redirecting spending towards AI data centres and Azure, though it says the roles are not being replaced by AI.

AI and society

21. Google confirmed it automatically enrols users into AI training from Search, Lens and voice queries, and explained how to opt out.

22. An estimated 11 million UK adults are using ChatGPT for financial decisions that fall outside the Financial Conduct Authority’s oversight.

23. ByteDance and Alibaba pulled AI companion features from their chatbots ahead of new Beijing rules restricting the technology.

24. An AI-generated actor, Tilly Norwood, has landed her first Hollywood film role.

25. Wealthy families are increasingly enrolling children in schools built around AI tutors, as an early education arms race takes shape.