The Sigmoids Won't Save You
7 hours ago
- #Sigmoid Curves
- #Lindy's Law
- #AI Forecasting
- AI capabilities have been improving dramatically since at least GPT-1 in 2017, with many extrapolating exponential growth, but technically all exponential processes eventually become sigmoids due to physical or practical limits, as seen in epidemics or airspeed records.
- Forecasting when an exponential trend will transition to a sigmoid is challenging; examples like UN birthrate projections, solar power deployment predictions, and AI capability models often misidentify the timing, assuming a flattening too early.
- In conditions of ignorance about a process, Lindy's Law suggests assuming it will continue about as long as it has already; applying this to AI, if scaling began around 2019, the default expectation might be another seven years of similar progress.
- To argue that AI capabilities won't reach a 'high' level, critics must either provide a detailed model of AI dynamics, including factors like data center growth and algorithmic progress, or justify why Lindy's Law doesn't apply as a black-box expectation.