Empiricists vs. Extrapolators

The essay argues that people keep underestimating the speed of AI progress and that this kind of mistake comes from exponential blindness. It contrasts empiricists, who wait for more proof and often fall into skepticism, with extrapolators, who rely on scaling laws and structural models to make forecasts about what AI systems will be able to do.

Reading notes

#

• The author says ChatGPT was an early sign of a fast-approaching AI wave, even though many people dismissed language models at the time.
• A recent UPenn preprint questioned METR’s forecast on long-horizon tasks, but new OpenAI and Anthropic models appeared to confirm the exponential trend on the same day.
• The essay divides AI commentators into empiricists and extrapolators.
• Empiricists are described as trusting hard data but often settling for “I’ll believe it when I see it,” which can drift into Humean skepticism.
• Extrapolators are contrasted with shallow curve fitters and linked to deeper structural models that make out-of-distribution bets.
• The author says extrapolators have had a strong record in AI because they grounded their forecasts in physics, biology, neuroscience, and statistical mechanics.
• Anthropic’s strategy is presented as a high-conviction bet that scaling laws are genuinely law-like.
• The essay says scaling laws have been measured across many orders of magnitude and likely generalize across different modalities.
• Once a training method begins to work, the author argues, it is reasonable to expect continued progress toward the summit.
• Video generation is used as an example, with early distorted outputs treated as evidence that life-like video models were coming.
• The author says complex systems can still be forecast when stable invariants dominate, and gives average temperature versus weather as the comparison.
• AI capability forecasting is compared to thermodynamics rather than the three-body problem.
• The conclusion is that forecasting AI is necessary because institutions move much slower than AI progress, so preparation has to happen before the capabilities fully arrive.