Scott Alexander published 'The Sigmoids Won't Save You' on Astral Codex Ten on May 15, 2026, challenging the widespread assumption that AI capabilities will plateau before reaching concerning levels. The article argues against the common refrain that "all exponentials eventually become sigmoids" (S-shaped growth curves), pointing out that this observation provides no information about when plateauing will occur—and that timing matters enormously for AI safety.
Historical Examples Reveal Consistent Misjudgment
Alexander catalogs multiple instances where experts incorrectly predicted when exponential trends would flatten:
- UN birthrate projections consistently overestimated when fertility declines would plateau
- Solar power deployment exceeded predictions year after year, maintaining exponential growth far longer than anticipated
- A 2026 Wharton paper attempted to model AI capabilities with a sigmoid curve, predicting an imminent plateau that was immediately contradicted by the next model release
The Wharton example proves particularly relevant—occurring in early 2026, it demonstrates that even recent attempts to predict AI plateaus using sigmoid models have failed to match reality.
The Core Problem: Timing Uncertainty
Alexander's central argument is that while growth curves do eventually flatten, skeptics have no reliable method for predicting when this will happen. The observation that "exponentials become sigmoids" is technically true but practically useless without a dynamic model explaining the transition timing.
He proposes that skeptics must either provide explicit models showing when and why AI scaling will plateau, or default to Lindy's Law—assuming trends continue for roughly as long as they've already persisted if treating AI progress as a "black box."
Standards for Plateau Predictions
The article establishes two acceptable approaches for predicting AI plateaus:
- Dynamic modeling: Provide explicit models with mechanisms explaining when and why scaling limits emerge
- Lindy's Law default: If treating AI as unpredictable, assume current trends continue for a duration comparable to their existing track record
Based on current momentum and historical precedent, Alexander suggests AI scaling could plausibly continue for several more years.
Community Reception
The article received 141 points and 160 comments on Hacker News within 16 hours, indicating substantial engagement from technical communities actively monitoring AI progress.
Key Takeaways
- Scott Alexander challenges the dismissive claim that AI will plateau soon based on sigmoid growth patterns
- Historical examples show experts consistently misjudge when exponential trends will flatten—from birthrates to solar power deployment
- A 2026 Wharton paper predicting AI capability plateaus was immediately contradicted by subsequent model releases
- Alexander argues skeptics need explicit dynamic models or should default to assuming continued trends based on Lindy's Law
- The article suggests AI scaling could continue for several more years based on current momentum and precedent