@jdthewlis
Pinned post
dcferreira.com How I use agents for work | Daniel C Ferreira How I run 3-10 coding agents in parallel at work, and why the trick is making sure they already have the context. 
github.com MuonH implementation includes the Expert dimension in the frobenius norm computation for 4D inputs... As of #7820 the expert FF parameters are stored in a big [n_layers, n_experts, d_in, d_out] tensor, eg [48, 384, 3072, 3072], with the specialised _newtonschulz_4d_distributed computing Newton Schu... 
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.View quoted post on X
Claude really be like "so a var the launch strips names a file it never opens"View quoted post on X
