DeepSeek superagent prototype, DetectorScope, TEM, QwenPU

Today's feeds were mostly commentary, so I went looking for things I could actually open. Four repos held up: one agent prototype from DeepSeek, one interpretability release from DeepMind, one image-model paper with code, and one odd community experiment with Qwen weights.

DeepSeek's proactive coding agent

DeepSeek published dsh-superagent-prototype, an MIT-licensed agent for a single GitHub repository. It listens to GitHub, and optionally Feishu. When an allowed person asks for work, it opens a work item, runs in its own session and directory, and answers in the original conversation, often with a pull request. A web panel lists the items that need a person.

The design is what interests me. The repo ships numbered docs on philosophy, domain model, authorization, execution, prompts and deployment, plus a cookbook for adding a new source. Authorization is its own document, which is the right instinct for an agent that can open pull requests on your behalf.

DeepSeek is careful about status. The README calls it a research prototype for reference only, not an official product, says it may contain bugs, and notes that some designs in the accompanying article are not implemented. There are no benchmarks. It is a reference for how to structure the thing, not a tool to deploy tomorrow.

DetectorScope from DeepMind

DetectorScope builds AI text detectors that show their work. Sparse autoencoders are trained on the residual stream of a detector built on Gemma-2-2b, each feature gets a natural-language description checked against its activations, and only features above a validation threshold feed a sparse linear head. Each prediction is then a sum of described features with token-level locations.

The repo includes a PyPI package, three Hugging Face detectors, a Gradio demo and reproduction scripts. The paper is on arXiv as 2610.30006. DeepMind says the auditable version has minimal adverse effects on accuracy compared with the black-box classifier underneath, and that full tables are in the paper. The README also warns against using it for decisions about individuals, such as academic integrity or hiring.

TEM for image generation

TEM is an energy-based model trained contrastively and sampled with Langevin MCMC, with no classifier-free guidance. The authors report FID-50K of 4.74 for the 415M model and 3.82 for the 609M model at 256px, and 8.24 at 512px. Inference-time scaling is visible in their table: 40 steps gives FID 7.63 and 60 steps gives 3.82. Code and a TEM-XL/2 checkpoint are released, though I did not find a code license on the page. The paper is arXiv 2610.31889.

QwenPU

QwenPU compiles a 32-bit RISC-V CPU and a vector unit into the weights of Qwen3.5-4B, with no learned weights, and runs Karpathy's stories260K on top. The author reports 53,000 trace tokens per story token, against 225,000 for an earlier NPU version, and exact matches against a reference. It needs a 40 GB GPU for the full checkpoint and the author labels it work in progress. A curiosity, but a precise one.

What I'd watch next

Whether anyone reproduces the TEM numbers at 512px, and whether DeepSeek moves any of the prototype's unimplemented designs into the repo.