So, what exactly can a base model do? Is this like the pre-reasoning models that just autocompleted, without any alignment, reasoning, etc optimizations?
But it makes it far easier for other parties to do that last 10% to train it to be useful.
However, there are many choices one makes in the last 10% of training which will impact the models behaviour when it comes to AI safety, political opinions, willingness to help make a bomb, etc etc.
This kind of figures is often quoted but where does that estimation comes from? Isn't that legacy number from the era when all the post training WS just instruction following and RLHF? I'd be surprised if current generation if reinforcement learning was still a few percent of the training cost.
Withtout RLHF, instruction-following training, or RLVR to develop reasoning.
Now a good fraction of the post training (Most maybe at this point? Who knows?) is spent on RLVR (reinforcement learning on verifiable reward), in this kind of training you don't judge the token generated themselves, but the answer as a whole ("does this program compile and pass the tests", "is the proof valid"). That's where all of LLMs progress over the past two years in software and math come from.
So yes, they are still next token predictors at their core, but they stopped being mere stochastic parrots since they are now being trained on much more than just imitating their training data in a statistically plausible way.