Open training run: EDA beats GDN-2, auto expert failover survives 7 host crashes
jon_durbin · x · 2026-10-08
Developer jondurbin shares an update on his open model training run:
- EDA outperforms GDN-2 and has so far entirely avoided the prior failure mechanism; the new run (lambda) is blowing past the old one (kappa).
- The run survived 7 host failures thanks to automatic expert ownership failover.
- He also tried adapting triadic linear attention to EDA for better long-context ability, but saw no clear benefit (possibly a skill issue); the experiment code is public.
More from Infra
- Glimmer model details: full training pipeline and 50 tok/s on a MacBook via 4-bit quant + DFlash — TimDarcet · 2026-10-08
- Qwen3.8 27B on two used 3090s beats Claude Haiku 5.5 for free in a game-building bake-off — iamfakhrealam · 2026-10-08
- llama.cpp PR #30100 fixes major accuracy issue with Clef models on MacOS/Metal — yuicebox · 2026-10-08
- Cisco's AI infra business hits $9.3B in orders in two years, CPO Jeetu Patel tells YC — ycombinator · 2026-10-08
- GLM 5.3 Flash spotted hitting 227 tok/s with 4 concurrent streams on DGX Spark — natesiggard · 2026-10-08
- AWS CEO Matt Garman on $220B capex, GPUs for startups, and rebuilding the cloud for agents — a16z · 2026-10-08