▲ | moffkalast 17 hours ago | |||||||||||||||||||||||||
ML models have the good property of only requiring investment once and can then be used till the end of history or until something better replaces them. Granted the initial investment is immense, and the results are not guaranteed which makes it risky, but it's like building a dam or a bridge. Being in the age where bridge technology evolves massively on a weekly basis is a recipe for being wasteful if you keep starting a new megaproject every other month though. The R&D phase for just about anything always results in a lot of waste. The Apollo programme wasn't profitable either, but without it we wouldn't have the knowledge for modern launch vehicles to be either. Or to even exist. I'm pretty sure one day we'll have an LLM/LMM/VLA/etc. that's so good that pretraining a new one will seem pointless, and that'll finally be the time we get to (as a society) reap the benefits of our collective investment in the tech. The profitability of a single technology demonstrator model (which is what all current models are) is immaterial from that standpoint. | ||||||||||||||||||||||||||
▲ | wincy 16 hours ago | parent [-] | |||||||||||||||||||||||||
Nah, if TSMC got exploded and there was a world war, in 20 years all the LLMs would bit rot. | ||||||||||||||||||||||||||
|