[dogma1138]

Again that may reasonable but it’s a completely different argument. Whether there is a bubble or not and whether NVDA is overvalued is irrelevant to the subject at hand.

If it’s cheaper to train models it means far more customers that will try their luck.

If you reduce training requirement from a 100,000 GPUs to a 1000 you’ve now opened the market to 1000’s and 1000’s of potential players instead of like the 10 that can afford dumping so much money into a compute cluster.

[fragmede]

the holy grail is to not have a separate train and inference steps. when the model can be updated while it is inferencing is where we're headed. deepseek only accelerates the need for more compute, not less

[Jlagreen]

THIS is the only correct statement in all of this.

The goal for AGI and ASI MUST BE to train, inference, train, inference and so on and that all on the fly in fractions of a second from every token produced.

Now good luck calculating the compute and hard work in algorithms to get there.

Not possible? Then AGI won't ever work because how can AGI beat a human if it can't learn on the fly? Not to mention ASI lol.