New Paradigms Won't Save You

A tight quantitative rebuttal to the AI-skeptic syllogism 'LLMs can never be AGI -> AGI needs a new paradigm -> therefore AGI is far off, don't worry.' Scott's move is to grant the premise and attack the 'therefore.' He lays out an 'evolutionary tree' of AI advances (neural nets 1950s, MLP 1967, deep learning 2010, transformer/LLM 2017, RLHF 2022, chain-of-thought 2024) and applies Lindy's Law: a new paradigm as revolutionary as the transformer is ~9 years out at the median, but Lindy's heavy tail puts the 25th percentile at ~3 years (and ~5 years for a deep-learning-scale shift) -- so even paradigm-shift believers should hold ~25% on AGI within 3-5 years, about what the LLM-only crowd expects. He notes accelerants (researcher headcount keeps inter-advance time roughly constant or shrinking; AIs soon contributing) and argues scaling-up delay is largely already paid. The subtler, more original point: new paradigms historically appear precisely when the old one stops efficiently converting scale to results, so 'extrapolate current LLM scaling' is a good forecast even if LLMs aren't the final paradigm -- the next one would resume from where scaling stalled, at roughly the same rate.

Quality 71 · Strong. Strong (71): a clever, genuinely useful quantitative reframe that neutralizes a common AI-timelines objection on its own terms, plus the original 'paradigms emerge where scaling stalls, and continue at the same rate' insight. Upper-Strong rather than Excellent because it's brief (~800 words) and leans on a single heuristic (Lindy's Law) applied to a small, hand-picked advance list.

Claude’s paradigm shift 48 · Moderate. Moderate, upper (48): the Lindy's-Law-on-the-paradigm-tree framing and especially the 'new paradigm resumes from where scaling stopped' argument are fresh, non-obvious angles on the timelines debate, though built from familiar forecasting heuristics and the existing LLM-skeptic discourse.

Real-world impact 2 · Minor. A clever quantitative reframe that neutralizes the 'AGI needs a new paradigm, so it's far off' syllogism on its own terms (Lindy's heavy tail puts a transformer-class paradigm's 25th percentile ~3 years out), plus the 'paradigms emerge where scaling stalls' insight. Conceptual influence within AI-timelines discourse, brief, no material change — low RWI.