General AI
The original purpose of science was to find truths that are true in the future as well as in the past. This allows us to model.
It took me ages to realise that the only difference between your local LLM and humans is that LLMs don’t have any maths and can’t do any modelling of the future.
(Unless they just happen to have access to some python code that is fit for purpose).
Nor do most humans by the way which is why the Turing test is so stupid.
I can’t see any reason why a neural network couldn’t specifically be trained on scientific truths alone.
You ignore scientific papers and train only on the classic text books in all fields of all the sciences, medicine, maths, and engineering. Maybe even psychology.
You figure out how to embed the equations, figures and text into vector space. Not just the text.
Embed the rules of reality first, not just the words about reality. Query; hard coding versus correlation?
Then you have one neural network that can model the future and the unknown.
That, working in concert with a general purpose wordsoup LLM, and voila, general AI.
—
How Humans Model Reality (and Why It Works)
Decomposability: The universe (fortunately) has structure that can be broken into subsystems (e.g. you can study fluid mechanics without understanding general relativity; in fact you’re better off without it).
Specialisation: No human scientist needs to know everything, just their local domain very well.
Communication standards: Units, mathematics, and experimental methodology standardise knowledge exchange across fields.
Error correction: Peer review, replication, and scientific debate correct errors over time (imperfectly, but better than random evolution).
Canonization: text books and software act to create standardised generally accepted models of the truth that can be used by anyone that is interested.
—
What This Implies for AI
You don’t build one general modeller. You build a society of tiny specialist models, each with:
A sharply bounded domain it knows deeply (e.g., thermodynamics at moderate temperatures, or Navier-Stokes equations for incompressible fluids).
Common communication protocols (e.g., math libraries, symbolic standards, constraints on units/dimensions).
Negotiation and handoff mechanisms when interacting with other specialist models (like APIs or graph edges).
Self-monitoring rules (e.g., “Am I extrapolating beyond validated conditions?”).
In other words: recreate a distributed society of scientific specialist agents.
offshore westerly 