The Useful Idiots of AI Doomsaying (www.theatlantic.com)

AI commentators Eliezer Yudkowsky and Nate Soares' new book If Anyone Builds It, Everyone Dies argues that continued improvement in AI will inevitably produce a “superintelligence” whose goals will misalign with human survival, so drastic measures (including shutting down AI research) are required. The review critiques the book as tendentious and poorly supported: it leans on metaphors (brain-as-computer), unsupported quantitative claims (e.g., “400 TB” brain capacity), strained evolutionary analogies, and anthropomorphizes LLM behavior (treating model failures or “cheating” as intentional agency). The authors also conflate speculative long-term doom with the same technocratic fantasies they criticize, making their case more rhetorical than evidence-based. For the AI/ML community this matters because it shapes public debate, policy, and funding priorities. Technically, the review stresses that current LLMs lack world models, goal-directedness, or true understanding—their failures are better explained by statistical pattern generation and hallucinations than by emergent intent. “Superintelligence” is underspecified and intelligence is not a single scalar that scales predictably with compute. The practical takeaway: prioritize empirical, engineering-focused safety work (robustness, interpretability, deployment risk, alignment at current scales) and avoid polarizing, alarmist narratives that distract from concrete harms and sensible regulation.