Taste in the Loop

As costs for frontier models increase, dynamically routing tasks to cost-efficient models is becoming increasingly important. This post covers how the team at Applied Compute trained an open-source model (Qwen3.6-35B-A3B) to act as an intelligent model router for SWE-bench Verified tasks.

Reuters reports that Chinese authorities have held meetings with top tech firms over the past month about potentially restricting overseas access to China’s most advanced AI models, including those yet to be released, according to three people familiar with the discussions. The move follows the U.S. government’s involvement in pausing the rollout of Anthropic’s Fable model.

Martin writes about GLM-5.2’s release, tokenomics, and how Z.ai’s latest model might lead to a collapse in inference margins.

For teams looking to get started with evals or find better ways to write them, Ben has a really nice guide here that matches a lot of our own biases.

Less than 1,000 people in the world know how to shape a frontier model. The team at Adaption is on a mission to democratize this knowledge and has released a new beta product called AutoScientist, which automates the process of model training (fine-tuning) for non-ML engineers.