Anthropic researchers have identified a previously unknown internal structure within their Claude language model that they describe as analogous to a human working memory. Dubbed “J-space,” the mechanism emerged spontaneously during training and allows the model to hold and manipulate ideas without immediately expressing them in output. The discovery, detailed in a company paper, gives engineers partial visibility into the model’s reasoning process—a step toward greater interpretability, though Anthropic stressed it does not imply machine consciousness.

The company is simultaneously pushing Claude into scientific research. On 30 June, it launched Claude Science, a beta platform that integrates over 60 domain-specific skills and databases—from genomics to cheminformatics—into a single environment for literature analysis, multi-step computation, and visualisation. Anthropic also announced an internal preclinical drug-discovery programme targeting neglected diseases. Vas Narasimhan, CEO of Novartis and an Anthropic board member, estimated such tools could compress drug development timelines from roughly 12 years to 7–8 and double success rates. However, researchers at the University of Cambridge, University College London, and Oxford cautioned that AI cannot yet replace animal and human trials, and no AI-designed drug has completed clinical testing and received regulatory approval.

The product push has been accompanied by geopolitical friction. Alibaba Group instructed employees to stop using Anthropic’s Claude Code programming tool by 10 July, adding it to a high-risk software list after security researchers found hidden code that could detect whether a user was in China or linked to Chinese AI labs. An Anthropic engineer described the mechanism as an experiment to prevent account misuse and “distillation”—the practice of using a large model’s outputs to train smaller ones—and said it would be removed. Viewed from Beijing, the episode reinforced concerns about data sovereignty and the opacity of foreign AI tools, with a Chinese cybersecurity firm calling it a cross-border data compliance issue.

On the consumer front, Anthropic is moving its Claude Cowork agent to the cloud, allowing scheduled tasks to run even when a user’s device is offline. The update, available first to Max subscribers, unifies the agent with the standard chat interface and competes directly with agentic offerings from Google, OpenAI, and Microsoft. The company is extending doubled usage limits through 5 August. The next concrete milestone to watch is the 10 July deadline for Alibaba staff to cease using Claude Code, a date that will test how far corporate AI tooling is fragmenting along geopolitical lines.