Does this also include the non-coding time required to train any neural networks involved, or to test the new system?

@NLeseul Can't speak for what we will decide, but I would say no; the relevant criteria are whether this is indicative of the singularity/doom being near. Requiring more time to train or test does not sound like a significant safeguard.

@amaurylorin Bear in mind that the actual code to execute current LLMs is very, very straightforward and boilerplate; ChatGPT could probably spit out something reasonable for you right now if you wanted. Actually improving an LLM is entirely a matter of assembling a better training data set, throwing enough GPUs at it to train it in a reasonable amount of time, and tweaking its performance based on human evaluation. So those factors will be the main obstacle to anyone trying to build an improved LLM-based AI.

@NLeseul Yes, I think it's likely (>.6) that an AI-assisted team will be able to do that in a day by next year. Most of my unlikelihood is in the efficiency criterion.

@NLeseul in that case, how would the new one be better? I think this question is supposed to be something very different than "can a team back together a weird but functional assistant in a hurry".