Security is another problem, and should be tackled systematically. Artificially making dependency inclusion hard is not it and is detrimental to the more casual use cases.
When you go travelling and do not want to carry around a backpack, and 1 day of heavy video recording, watching youtube on train plus 1 year of lithium battery degradation. That's when I want larger battery.
Law should be considered to be artificial rules optimized for the collective good of society.
What's the worst that can happen if we allow unregulated AI training on existing music? Musician as a job won't exist anymore lest for the greatest artists. But it makes creating music much more accessible to billions of people. Are they good music? Let the market decide. And people still make music because the creative process is enjoyable.
The animus towards AI generated music deeply stems from job security. I work in software and I see it is more likely that AI can be eventually able to replace software devs. I may lose my job if that happens. But I don't care. Find another career. Humanity needs to progress instead of stagnating for the sake of a few interest groups.
AlphaGo seems more like an automated process to me because you can start from nothing except the algorithm and the rules. Since a Go game only has 2 outcomes most of the time, and the model can play with itself, it is guaranteed to learn something during self-play.
In the LLM case you have to have an already capable model to do RL. Also I feel like the problem selection part is important to make sure it's not too hard. So there's still much labor involved.
Andrej's video is great but the explanation on the RL part is a bit vague to me. How exactly do we train on the right answers? Do we collect the reasoning traces and train on them like supervised learning or do we compute some scores and use them as a loss function ? Isn't the reward then very sparse? What if LLMs can't generate any right answers cause the problems are too hard?
Also how can the training of LLMs be parallelized when updating parameters are sequential? Sure we can train on several samples simultaneously, but the parameter updates are with respect to the first step.
It's questionable a mere algorithm would get us there without a fundamental change in computer architecture. (in terms of Intelligence / W)