I thought training new models on AI generated output leads to model collapse?

aytigra 9 minutes ago | parent | next [-]

I think it case of coding it may not be as bad, because new training data (AI generated code) is always empirically validated by tooling and by consumer. It may not be good but it mostly works, otherwise it is discarded or patched, so it has a bottom bar of "it works".

▲

Chance-Device 6 hours ago | parent | prev [-]

My point is more that they won’t need new training data to extrapolate to new problems, especially if that new problem is just a new syntax or API. Put the whole shebang into the context. Done. Yes, you need super long context for this, or just pretty good search over it, but I think this is likely to become a solved problem.

	▲	MattPalmer1086 5 hours ago \| parent [-]
		Agreed if its just a new API using a current language using common patterns, that is mostly just having a big context. However, the ability to reason generally on novel problems is AGI and we aren't there yet. Eventually in the absence of AGI, we will have to train models on them, and that will require data.