> LLMs have been able to train new AI for a while now.
I admittedly am not an expert, but I don’t think this is true? At least not in the “AI is building itself” sense, which is how I interpreted it.
I admittedly am not an expert, but I don’t think this is true? At least not in the “AI is building itself” sense, which is how I interpreted it.
I've had LLMs write code which performs a genetic algorithm for balancing video game AI*. There's a very soft line in quickly moving metaphorical sand between what it can do and what it can't, and right now I'm not sure which side of that line you'd find a user saying they want a computer vision system to turn a robot arm into a barista given LLMs can search for, download, read, and seemingly do a decent job of implementing research papers e.g. sim2real.
Just to be explicit though: I'm not saying the state of AI-led research is sufficient to render unimportant researchers who find better ways to train AI, just me as a software engineer is no longer a blocker to productising that research.
* The same LLM had also written the game AI, but while I'm fine pointing to a genetic algorithm balancing it, I don't think that the game AI in itself counts because "game AI" is GOFAI.