Google's Sleep framework lets LLMs keep learning without forgetting
The paper claims near-perfect recall to 10M tokens; also today: SkyPilot's rack-aware GPU scheduler and Mistral Large 4's open weights going live.
Copy markdown
Models that learn after deployment
Google Research's Sleep framework alternates a "wake" phase with an offline "sleep" phase that folds recent context into the weights, then generates synthetic "dream" data to fine-tune on — and reports near-zero forgetting across sequential tasks plus near-perfect recall out to 10M tokens on BABILong, ahead of GPT-4 and Titans. It's a paper, not a release — no code or weights yet — so read it as a signal that post-deployment continual learning is getting real, not a library to install this weekend.
Up to 27% more usable GPUs, same racks
SkyPilot makes the case that topology-aware scheduling — pinning each multi-node slice to one NVL72 rack so traffic stays on NVLink, holding two hot-swap spares per rack, and defragging topology-blocked jobs — lifts a 128-rack GB300 cluster from ~1,602 to ~2,037 usable nodes at 2% node failure (a capacity model, not a measured speedup). If you run your own cluster: size multi-node jobs to fit one rack (their example is 64 GPUs / 16 nodes), leave spares rather than packing all 18 nodes, and checkpoint so jobs can be preempted and resumed.
Update: Mistral Large 4 weights are live
The ~1T-parameter (≈52B active), natively multimodal MoE we flagged as "coming this week" is now on Hugging Face under Mistral's Ecosystem Open-Weight (modified-MIT) license — free to use, fine-tune and run commercially, with a separate license only if you resell it as a hosted service. Serve it with vLLM, SGLang, NVIDIA NIM or TensorRT-LLM, or hit the API at $0.70/$2.10 per million input/output tokens if you'd rather not run the weights yourself.