How Much of the World Is It Possible to Model?

Abstraction: Limits and nature of mathematical models from climate to LLMs

Key points:

• Models must replicate the known while generalizing into the unknown; reliability scales with closeness to immutable physical laws
• Earth System Models (ESMs) validated against 2,000 years of historical temperature data with thousands of coupled equations; IPCC 2022 report shows remarkable fit
• S.I.R. epidemic models (Susceptible-Infected-Recovered) captured COVID-19 curve shape but broke down during rapid mutations and lockdown policy shifts; CDC shut down its COVID forecast project Dec 2021 citing "low reliability"
• LLMs described as "vertiginous curve fitting": hundreds of thousands of neurons with trillions of parameters capturing statistical correlations in vast text corpora — traceable to McCulloch-Pitts 1943 logical neuron paper
• Key danger: conflating task performance with understanding of the underlying phenomenon (thought); interpretability/XAI addresses this gap
• Elegance is a trap — models too simple may give useful heuristics (flatten the curve) but fail at specific predictions; some phenomena may require "baroque" models

Connections: Ipcc · Mathematical Modeling · Large Language Models · Simulation · AI Interpretability

Source: https://www.newyorker.com/culture/annals-of-inquiry/how-much-of-the-world-is-it-possible-to-model