@Technop54777070@taosx@__tinygrad__ and i think in an agentic world, with lora and other things, a more GPU - refreshed latest model can steer the super fast etched models on their execution or change the system prompt enough on the underlying workloads over time.
a model is a lot less of a hard coded software.
@AngelicaOung lol, saying this the day jeff dean leaves google after the deepseek-kimi onslaught.
theres nothing magical about american firms other than the blue water navy magic.
@NahBrohall@Duckyduck_7 Syama doesnt always been black, it also refers to deep blue/ deep green (emerald) so maybe to present a humanish complexion, fair is used.
@NahBrohall@Duckyduck_7 the question is did it happen, given that Shri Parshuram bhagwan is known to be a dambhi, and no clear verse saying so, better stick to the source material.
@neural_avb EOY hoping a bunch of inference hardware moonshots to materialise,Taalas Deepseek R1 inference cluster at 12,000 tps being one of them.
Society is about to solve information as well as atom manipulation in a flash by conventional standards.
@rajivej_@datta_arvind his point seems to be that leveraged bets like software with their sky high PEs will come back to ground and high cash businesses will benefit in a high interest environment, but hey, who knows.
@atomtanstudio@TheAhmadOsman we will have good consumer inference hardware for sure.
as far as ever bigger models are concerned, yes, but for increasingly scarce users, most likely nation states, and giant corpos.
@Vinay45428476@ShaatirBanda you dont attack the dragon until you are ready.
letting elephant think he is still in command is better, a gradual takedown is recommended.
Pakis know this and so does he.