Rohan Anil Says Rethink Gradient Descent for More Progress

Rohan Anil, co-founder of CoreAutoAI and former Distinguished Engineer at Google DeepMind on Gemini pretraining, posted that there is a lot more progress if gradient descent is rethought for training. The tweet presents the point as a repeated assertion from the author and includes his professional background. The packet contains only this single source line with no additional claims, examples, or corroboration from other accounts.