@gerardsans
🚨 DeepMind AGI/ASI paper: a Field Lost in Hype and Metaphors Solid scaling. Useful pathways. Overdue bottlenecks. But it keeps treating AI like a subject with a mind. It’s not. Why serious researchers treat massive compressed files running matrix multiplication as social “agents” with psychology, beliefs & motives is beyond me. LLMs don’t think. They compute. “Step-by-step thinking”? Repeating training corpus patterns that look like problem-solving. “Understanding”? Pattern matching. Not comprehension. The 4 paths to ASI are real but lead to better pattern matching, efficient data compression, not actual intelligence. Fix the language: • Intelligence → data coverage • Reasoning → structured sampling • Goals → constraints • Creativity → interpolation Humanised AI narratives make us overestimate what these systems can do, underestimate failures & build things that look smart but aren’t safe or reliable. Paper’s a great start. Describe AI as it actually is: a sophisticated pattern engine, not as we wish it were. Anthropomorphism in analysis wastes resources and confuses policymakers with false narratives. We have a long way to go to overcome hype and misdirected incentives.