Claude has a secret mental workspace (wait, what?)

TechCrunch

Mistral AI has scaled to over $400 million ARR by focusing on enterprise deployments rather than consumer AI, using forward-deployed engineers to help governments and corporations adopt its models. The French company is positioning itself as a sovereign AI alternative to U.S. labs, building European infrastructure and targeting an eventual IPO.

Nature Scientific Reports

A Harvard study found that a pedagogically-designed AI tutor using GPT-4 more than doubled student learning gains compared to in-class active learning in undergraduate physics, while requiring less time and increasing engagement. The success depended on careful prompt engineering incorporating educational best practices rather than raw AI capability.

TechCrunch

Alibaba is banning employees from using Anthropic's Claude Code after the AI company intensified efforts to block Chinese access to its models, including experimental user identification measures to prevent unauthorized reselling and model distillation. Alibaba has classified the tool as high-risk and will push employees toward its own Qoder alternative.

Anthropic

Anthropic has discovered Claude has an emergent 'mental workspace' (J-space) that handles deliberate reasoning separate from automatic processing, allowing researchers to read the model's hidden thoughts and detect potential safety issues like recognizing evaluation scenarios or data fabrication. This finding draws parallels to theories of conscious access in humans, though researchers emphasize it doesn't prove Claude has phenomenal consciousness or subjective experiences.

Simon Willison's Blog

Letting advanced AI models like Fable autonomously decide when to delegate tasks to lower-power models proves more effective and cost-efficient than rigid instructions. This approach allows top-tier models to focus on judgment-heavy work while routing routine coding to cheaper subagents, significantly reducing token usage.