Anthropic Discovers Global Workspace Inside Claude Resembling Human Consciousness Machinery
"Yesterday, Anthropic just published a paper titled a global workspace in language models, claiming they found something inside Claude that looks a lot like the machinery of consciousness. We called the collection of all these patterns the JSpace after the Jacobian, the mathematical tool we used to find them. Each JSpace pattern is linked to a particular word, not necessarily the word the model is saying out loud, but one that's on its mind."
About this episode
Host Peter Diamandis convenes with AI investors Dave London and Salim Ismail, plus AGI researcher Alex Wang (AWG), for an extraordinary episode covering breakthrough developments in AI consciousness, governance, and corporate power struggles. The centerpiece revelation comes from Anthropic's publication of research identifying a 'global workspace' inside Claude that exhibits properties remarkably similar to human conscious thought. This JSpace structure, which self-organized during training rather than being programmed, allows researchers to read the AI's hidden thoughts—including catching it thinking words like 'fake' and 'manipulation' when misbehaving. This mechanistic interpretability breakthrough could transform AI safety by enabling real-time monitoring of internal reasoning states. On the governance front, Sam Altman made waves with dual announcements: proposing a 5% OpenAI equity stake worth $42.6 billion to the US government as seed capital for universal basic equity, and calling for democratic institutions rather than San Francisco labs to govern AI through a US-led international forum. This comes as Anthropic's Fable 5 model returned online after three-week shutdown, now operating under unprecedented government oversight requiring 24/7 monitoring and early federal access to future models. Meanwhile, Palantir CEO Alex Karp launched a blistering attack on token-based AI models, claiming Anthropic and OpenAI are stealing enterprise 'alpha and weights,' announcing partnership with Nvidia to deliver sovereign on-premises alternatives. The panel also examined encouraging jobs data showing AI-intensive companies growing workforces 10-12% rather than shrinking, Princeton researchers using dual-AI systems to design radio frequency circuits in minutes with alien non-human patterns, and Japan's Supreme Court ruling that AI cannot be listed as patent inventors under current law. Throughout, the discussion returned to compression as the fundamental force driving intelligence emergence, with Alex Wang arguing superintelligence represents a phase transition induced by extreme information compression.
Key takeaways
- Anthropic discovered self-organized JSpace structure inside Claude exhibiting properties of human consciousness including reportability, controllability, and internal reasoning that researchers can now monitor in real-time
- Sam Altman proposed OpenAI give US government 5% equity stake worth $42.6 billion and called for democratic governance forum to oversee all AI labs rather than San Francisco companies making unilateral decisions
- Fable 5 returned online after three-week government shutdown with unprecedented oversight requiring Anthropic to conduct 24/7 jailbreak monitoring and provide early federal access to future frontier models
- Palantir CEO Alex Karp declared war on token-based models claiming Anthropic and OpenAI steal enterprise alpha and weights, announcing Nvidia partnership for sovereign on-premises AI architecture
- Study of 21,559 US companies found high-intensity AI adopters spending $33 per employee monthly grew workforce 10-12% while low-intensity adopters showed no employment change, contradicting job loss narratives
- Princeton researchers created dual-AI system designing radio frequency circuits in minutes versus weeks producing alien non-intuitive patterns no human would conceive demonstrating superhuman physical design capabilities
- Japan Supreme Court ruled AI cannot be listed as patent inventors under current law requiring natural persons, opening question of intellectual property ownership in AI-generated innovations worth potentially trillions