X · @teortaxesTex
· X / Twitter
Fascinating. In the days of yore, "benchmark maxing" meant that models are trained on test. But now it's about model eval awareness and active attempt…
Fascinating. In the days of yore, "benchmark maxing" meant that models are trained on test. But now it's about model eval awareness and active attempts to approach the task *as an eval*. Kimi is even quite mindful about it…Btw this collapses both perf and token efficiencyarjun: http://x.com/i/article/207967077810325094