A week when open-weight models arrived all at once

Open-weight models came out in a cluster within a single week, along with the claim that a 27B-class model runs in the 200-tokens-per-second range on a single consumer GPU while coming close to top commercial models. In particular, the release of not just models but an agent harness, which lets tools, session logs, the agent loop and sub-agents be swapped out like plugins, bears directly on anyone trying to run agents on an internal network cut off from the outside internet. On the other side came, one after another, a case in which conversation logs were collected, classified and handed over to investigators, a decision to embed probability-bias watermarks in output text, and a feature that records all on-screen computer activity, which translate into three strands of internal-policy issues: log retention, output traceability and employee monitoring. On the robotics side, a report said a scaling law spanning six orders of magnitude was observed in a model pretrained on 1 million hours of recorded human behavior. However, the figures and benchmarks were all passed along without sources, so it is hard to use any of this as more than a watch list.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.