benchmarksfactbearishResearchers spend significant time and GPU hours attempting to replicate other model providers' evaluation scoresLewis Tunstall28 Jul 2026https://x.com/_lewtun/status/2079912215759438135