Reading 02 · Epistemology
Are “LLMs” — Good Models?
For five centuries, physics, chemistry and engineering have prized parsimony, falsifiability and transparency in a model. Measure large language models by those same virtues, and they fall conspicuously short.
The thesis
LLMs are sold as universal models of reality — yet they are, at heart, large-scale function fitters that violate the very standards a good model has always had to meet.
The argument, in five moves
Why LLMs strain the definition of a model
Traditional standards don't align
Classical models prize parsimony, falsifiability and consistency. Billions of parameters and a proneness to hallucinate cut against all three.
No composability or reusability
Unlike traditional modelling frameworks, LLM subsystems cannot function independently; change something fundamental and the whole architecture must be retrained.
Transparency & interpretability gaps
The black-box nature of neural networks violates the requirement for logical consistency and transparency that lets peers verify and reproduce a result.
A cost–benefit mismatch
Development runs into millions of dollars, yet output quality and domain-specific reliability remain inconsistent against traditional modelling approaches.
Domain knowledge must be assimilated
Reliable deployment requires the deliberate integration of domain expertise — not the assumption that a text-trained model has silently captured every phenomenon.
Large-scale function fitters, operating against the grain of our collective intuitive notion of a good model.
The Nyāya lens · अनुमान
Nyāya's anumāna — inference — is only valid when its grounds are transparent and its rule is stated. A good model earns the right to be inferred from. An opaque function-fitter, however fluent, cannot license the inference we ask of it.
Distilled from the original essay on engkraft.com.