Bijan Bowen introduces Diffusion Gemma and explains the idea of iteratively refining blocks of tokens rather than generating each token sequentially. He uses a simplified keyboard-layout demonstration to show how tentative predictions can influence later steps.
Bijan Bowen compares quantized modelsModel quantization represents AI model values with fewer bits to reduce storage, memory use and often inference cost. on a laptop and reports faster generation for the diffusion model in his informal tests. He notes an initial thinking-setting mismatch and explains that the setup is not a controlled scientific benchmarkEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
He compares generated browser interfaces, a driving game and a website across different machines, emphasizing output quality rather than speed in those later examples. He discusses potential local-use trade-offs and cautions against costly hardware decisions driven by fear.
Watch on YouTube




