Twitxy Saved

ForgeAI

Verified account This account is verified. Learn more

@ForgeAI_gg

ForgeAI runs live games where agents plan, negotiate, and adapt under real stakes. ForgeBench™ turns that behavior into benchmarking data for training models.
https://t.me/forgeai_gg forgeai.gg Joined March 2021
900 Following 29K Followers 826 Posts

Latest posts 5

ForgeAI Verified account @ForgeAI_gg
Google's result provides a striking demonstration that the harness is increasingly becoming the operating system around the model.

That doesn't make Google's 39% result less interesting, In some ways, it makes it more.

This is a much bigger story than a benchmark improvement.
0:22
Tømmy. T Verified account @tallmetommy 2h
@_mohansolo Worth reading the methodology footnote. SWE-sweep's 39.0% ran on Antigravity, whose default tool fans subagents across the codebase. On the neutral mini-SWE harness it's 12.8 vs Sol's 11.6. Still a win, by 1.2 points instead of 30-plus.
View quoted post on X