3 ms·

You'd still need those giant data centers for training new frontier models. These Taalas chips, if they work, seem to do the job of inference well, but training

You'd still need those giant data centers for training new frontier models. These Taalas chips, if they work, seem to do the job of inference well, but training will still require general purpose GPU compute

bonoboTP 8mo ago
Next up: wire up a specialized chip to run the training loop of a specific architecture.
amelius 8mo ago
Yeah but you need even bigger factories to fabricate those inference chips, so what is the point?