What else happened on Oct 28th
On October 28 frontier models did poorly on simple physical tasks - on Robocurve's DataCenterBench Tier 1, none of the nine models managed to plug a charger into a power strip. The models moved two robot arms, guided by images from three cameras. And DeepMind's new model, according to DeepMind itself, let no indirect prompt injection through on one benchmark. Google DeepMind wrote in its Gemini 4 Carbon launch post that on the Gray Swan benchmark none of the attacks tried on the model got through.