On 2026-07-26 11:08, Richard Lewis wrote: > see eg > > https://en.wikipedia.org/wiki/Jacobian_conjecture#Counterexample_for_n_%3E_2 > > most humans apparently thought this conjecture was true, but an LLM > provided a counter-example: i dont think there is a meaningful sense in > which this was simply reproducing training data (the importance of this > is another matter, and subject to the usual hype)
It's been widely speculated that the huge leap of coding/reasoning model performance since last ~autumn is due to Reinforcement Learning with Verifiable Rewards (RLVR), a post-training technique. The early models where certainly training-dominated and the term 'stochastic parrot' might have been fitting, but nowadays they are much different beasts. Best, Christian
