The AI-collapse pre-mortem (berthub.eu)

The author offers an “AI-collapse pre-mortem”: even if an investment or hype bubble around large language models (LLMs) bursts, that won’t erase the genuine technical breakthroughs already achieved. LLMs are neither generally intelligent nor reliably useful in many production settings—they hallucinate, miscount, and often aren’t a clear path to profit—but they are astonishingly capable in specific tasks. The “chess‑playing pigeon” metaphor captures this: LLMs can perform miracles (grammar fixes, code hints, tense shifts) while still losing most of the time. The piece warns against equating current overhype or economic failure with the end of the technology. Technically, the author contrasts brittle LLMs with pattern‑matching AI that is demonstrably transformative: Whisper’s multilingual, jargon‑robust speech‑to‑text; neural models that predict splice junctions; and protein‑folding advances (cited via the AlphaFold Nobel) that solve problems previously inaccessible to pure human intuition. These systems show superhuman performance in narrow domains and already run on modest hardware or independent implementations (e.g., Karpathy’s nanochat). Implication: a market correction would force reevaluation, but won’t negate AI’s power in prediction and transformation tasks—so practitioners should stay pragmatic, focus on domain‑appropriate applications, and not discard the field because of hype or short‑term economics.