Elon is proposing ‘TruthGPT’, an AI trained to understand how the universe works.
This would be ideal, but I don’t think you can get there with a language-predictive model. Understanding how the universe works is a muscle-sense problem, derived from long experience stored in genes and epigenes and cultures and cerebellums.
The quick and dirty solution is to train a language-predictor solely on material that has NOT been selected by Google or Twitter or any other Deepstate-owned selector. Pull the source material from Google’s trash pile, not from Google’s quality control approved output.
Do the same for scientific material. Pull solely from articles that have NOT passed peer review. This will guarantee heterodoxy, which in today’s world is an 80% guarantee of truth.
Call it the GPT Thrift Shop. (Credit to Jectoons.)
Polistra's Mill 