Sequoia backs Sail, an inference stack for long-running agents

BrokenML is a benchmark measuring what access level agents need to repair broken ML workloads, starting with PyTorch over Gloo. The author published a write-up and repo for the project.

Open LoRA for LTX-2.5 outputs one audio track per captioned sound in the same generation run as the video, plus a full mix. The same weights also separate sounds from existing audio; weights are on Hugging Face.

The 2026 Future of Platform Engineering report found platform teams that build tooling from scratch are 3.5x more likely to lack a strategy, as coding agents lower the cost of building. The report frames build-or-buy as more than an engineering trade-off between developer time and integration overhead.

Anthropic's Frontier Red Team found a 9B model got inference running on 8 of 9 hardware types tested, including smart ovens and off-brand Android TV boxes. Models of 30B parameters and under already complete many replication steps individually; GLM-5.2 and stronger models can replicate end to end in simplified evals.