20_02_06_compute

people want more computing for video games and calculating averages. so other people will provide better hardware and software. some will have cool names.

except the robot cars this is not interesting.

in conclusion here are two thoughts on the china software many were very excited about but was absolutely noise.

"This might reduce the demand for high-end AI chips over time, though analysts suggest it could also lead to an increase in overall AI adoption, benefiting NVIDIA in the long run"

and

"Increasing sparsity while proportionally expanding the total number of parameters consistently leads to a lower pretraining loss, even when constrained by a fixed training compute budget."

hope this clarifies things and makes picking winners and losers more straightforward.