DeepMind demos model reasoning in blocks to self-correct

mtizard · x · 2026-08-17

Rumor claims Google DeepMind's VP of Research criticized autoregressive models for their one-word-at-a-time limitation. DeepMind reportedly showcased a model that generates text in blocks, allowing it to spot and fix mistakes before output. It claims to solve Sudoku by seeing the whole puzzle, hinting at a shift from prompt engineering to 'harness engineering'.

Original post →

More from Models

Models channel →