The newest AIs can now write other AIs by hand. Older scores close to zero On Monday, MIT researchers showed artificial intelligence models edit another model's weights directly, without training. Alex Krantsel, a PhD researcher at Berkeley, reran one of their tests on older and newer models. Each model had to write all the numbers inside a small grid that plays the Atari hacking game, with one answer, without training and without tools. The benchmark is a network trained by trial and error for 500 million steps. The GPT-6 Esters network achieved 22 times what this index received. Claude Opus 5.5 hits 20x, GPT-6.1 Sol 18x. GPT-5 to GPT-5.5 and Opus 4.5 to 4.8 didn't even hit the benchmark. Krentsel calls it "a new ability, phase-shifting, that only appears in the latest set of models. ” own caveat: the mesh is tiny, 20,486 parameters, while front-end models have around a trillion. In models with open weight it has not worked so far.

The newest AIs can now write other AIs by hand. Older scores close to zero On Monday, MIT researchers showed artificial intelligence models edit another model's weights directly, without training. Alex Krantsel, a PhD researcher at Berkeley, reran one of their tests on older and newer models. Each model had to write all the numbers inside a small grid that plays the Atari hacking game, with one answer, without training and without tools. The benchmark is a network trained by trial and error for 500 million steps. The GPT-6 Esters network achieved 22 times what this index received. Claude Opus 5.5 hits 20x, GPT-6.1 Sol 18x. GPT-5 to GPT-5.5 and Opus 4.5 to 4.8 didn't even hit the benchmark. Krentsel calls it "a new ability, phase-shifting, that only appears in the latest set of models. ” own caveat: the mesh is tiny, 20,486 parameters, while front-end models have around a trillion. In models with open weight it has not worked so far.