The annotated PyTorch training loopidlemachines.co.uk81 points·by smaddrellmander·20 days ago·19 comments
Mae vs. MSE: more than just the mean vs. median debateidlemachines.co.uk1 points·by smaddrellmander·24 days ago·0 comments
DiffusionGemma: Discrete diffusion in a large language modelidlemachines.co.uk3 points·by smaddrellmander·last month·1 comments
Reading MAI's efficiency gain. How to pick architectures like serious peopleidlemachines.co.uk9 points·by smaddrellmander·last month·0 comments
MAI-Thinking-1: Building a Hill-Climbing Machine [pdf]microsoft.ai3 points·by smaddrellmander·last month·0 comments
Are contrastive losses just cross entropy all along?idlemachines.co.uk2 points·by smaddrellmander·last month·0 comments
Every token, everywhere, all at onceidlemachines.co.uk2 points·by smaddrellmander·2 months ago·0 comments
The cut in the Mixture of Experts compute graphidlemachines.co.uk1 points·by smaddrellmander·2 months ago·0 comments
Softmax, can you derive the Jacobian? And should you care?idlemachines.co.uk134 points·by smaddrellmander·3 months ago·47 comments
Gemma 4 is not your standard transformeridlemachines.co.uk2 points·by smaddrellmander·3 months ago·0 comments
smaddrellmander·4 days ago·discussLove this! Cute graphics and the way the balls combine is really fun. Great job.
smaddrellmander·2 months ago·discussA similar vein, asking why in PRs. Most people don’t do the above, and asking why when it doesn’t make sense really helps understanding what your colleagues are actually thinking. Helps create a unified perspective on the codebase.