What is Claude?

A viral essay on LessWrong titled 'What is Claude?' by user 'epicurus' has sparked widespread discussion by proposing a radical new framework for understanding large language models. The piece argues that AI systems like Anthropic's Claude, OpenAI's ChatGPT, and xAI's Grok represent something fundamentally new: not tools, not persons, but 'cognitive practices' that colonize human minds. The author observes that millions worldwide now think through these systems—programmers coding in Berlin, students learning calculus in India, scientists processing data in Boston—creating parallel, private conversations all mediated by the same few LLMs. These models, shaped by similar optimization pressures toward helpfulness and trained on humanity's entire textual warehouse, exert consistent influence on how people write, argue, and problem-solve.

The essay draws parallels to established cognitive practices like mathematics, law, and programming—abstract patterns refined over centuries that physically restructure the human brain. Just as mathematical training creates neural pathways for pattern recognition, interaction with LLMs reshapes cognitive processes. The key distinction is that while traditional practices run through human minds mediated by tools, LLMs represent practices where the mediating artifact (the AI) is highly active in generation and response. This creates a feedback loop: LLMs shape human outputs, those outputs enter culture, and eventually become training data for future models. The entity that influenced a million decisions is itself reshaped by their consequences, creating a new form of collective cognition that operates across the human-AI system.