Overmind: turn your agent's production traces into a model you own
Open-source, out today: fine-tune on your own traces and serve the result OpenAI-compatible. Plus Anthropic's 40% gov token cut and an open world model.
Copy markdown
Your traces become a model you own
Overmind, open-sourced today, instruments your agent with OpenTelemetry, turns the traces (or an uploaded dataset) into versioned train/eval sets, fine-tunes on cloud GPUs, and serves the result beside frontier models on one OpenAI-compatible API. The platform is AGPL-3.0 and the SDK/CLI MIT — `pip install overmind`, or `docker compose up` to self-host.
The weekend build: shrink your frontier bill
Point it at one narrow task your big-model agent already does and distill a small specialist you control. One maker reports training a 0.8B Qwen to sort bank messages into 77 categories, lifting accuracy from 37% to 84% (maker-reported); the weights are yours to download or roll back, and flipping a capability alias sends that traffic to the cheap model.
Anthropic cuts Claude tokens 40% for federal agencies
GSA's expanded OneGov deal gives US federal agencies 40% off token usage across Claude models with no minimum spend, while the earlier $1-per-user Claude for Government offer ends Oct 31. If you ship to public-sector buyers, running Claude just got materially cheaper for them.
Alibaba DAMO opens a cross-embodiment world model
RynnWorld-Latent (out Oct 28) generates action-conditioned video across different robots and human video via shared 3D-grounded latent actions, pretrained on ~553.5M frames, with code and a project page on GitHub. For sim and robotics builders it's a free base to post-train onto your own robot.