Mastodon Feed
jonny@neuromatch.social ("jonny (nonvenomous)") wrote:
Love how every model just gets less and less efficient. to increase performance on the (always compromised) benchmarks by 10%, simply increase token use by 50%. Everyone knows that real intelligence works by reciting 1,000 Brothers Karamazovs worth of words in your head before saying anything. Surely there are no negative consequences to just continually scaling the amount of heat energy, er, compute per usage.
