Willison observed that the Qwen 3.8 27B model can read and write model weights locally, building upon research previously demonstrated with GPT-6 Astra. Concurrently, a major release for his LLM command-line tool integrated contributions from 12 developers to support new models such as gpt-6.1-sol and gpt-6.1-terra. The tooling update also introduces broader binary embedding features to support multimodal systems like Gemini Embedding 2. Outside of generative modeling, Willison covered acoustic flight-tracking intelligence techniques and practical dependency fixes for Python 3.15 users.
Key points
Weight-space programming is now viable on local consumer hardware rather than being limit…: Weight-space programming is now viable on local consumer hardware rather than being limited to massive frontier models.
Upgrades to the LLM utility expand developer workflows by handling multimodal binary embe…: Upgrades to the LLM utility expand developer workflows by handling multimodal binary embeddings and OpenAI-compatible endpoints.
Removing automatically generated uv.lock files resolves package installation errors on Py…: Removing automatically generated uv.lock files resolves package installation errors on Python 3.15 when prebuilt wheels are missing.
Sources checked by our analysis pipeline: 1
Entries Links Quotes Notes Guides Elsewhere
Tuesday, 3rd November 2026
Qwen 3.8 27B on my laptop can interpret model weights, too
Yesterday’s hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper (OK, the authors classify is as a “blog post”, but it’s paper-shaped) found that frontier models, in particular GPT-6 Astra, can both interpret and modify the raw weights of a machine learning model—literally a text description of the model containing a bunch of floating point numbers. I’ve been playing with one of the examples from the paper and I’m delighted to share that Qwen 3.8 27B running on my laptop can do this as well!
[... 1,211 words]I discovered a sound that leaks your exact location (and it’s everywhere)—YouTube (via) Genius piece of OSINT by Colsto, who demonstrates how you can take almost imperceptible background noise of planes from a video soundtrack, figure out their Doppler curves, then cross reference them against ADS-B data to find the exact location of…
Intermediary routers can quietly substitute targeted artificial intelligence systems with cheaper alternatives to cut operating costs. To prevent this practice, a new cryptographic mechanism allows developers and users to independently inspect the authenticity of their queries. This open-source implementation ensures callers receive verifiable proof that their outputs originated strictly from the intended model.
Software engineers are co-locating artificial intelligence guidance directly with compiled binaries rather than relying on external Git repositories or third-party marketplaces. This delivery mechanism ensures that automated coding assistants receive immediate, version-specific instructions alongside updated dependencies. The shift prevents advanced models from wasting valuable token budgets on reverse-engineering libraries from local cache documentation and assembly metadata.
Developers utilizing Anthropic SDKs, the Messages API, and Claude Code can now route native execution traces straight into their Databricks environments. The integration streams telemetry into Unity Catalog tables linked to MLflow experiments, allowing teams to query spans using SQL while maintaining enterprise-grade access governance. This shift aims to streamline debugging, automated failure analysis, and ongoing performance monitoring across AI agent workflows.