2 ms·
Sensible with the caveat that deepseek R1 still took millions of dollars off compute time, so you're not training the next one on the box in your basement with
by fragmede 2y ago
Sensible with the caveat that deepseek R1 still took millions of dollars off compute time, so you're not training the next one on the box in your basement with a pair of 3090s (though you could certainly fine-tune a shared quantized model). you can't run the full sized model on anything cheap, so. basement researcher still need access to a decent amount of funding, which likely requires outside help.