Purview is a weekly review of blog posts I find on the web and how they relate to my work. It is also a way to shout out interesting work, even when it sits outside my comfort zone.
Links: rewindvm.dev · Reading the code is the hard part
Farid Zakaria wants to build Rewind VM, a virtual machine with deterministic replay and byte-level provenance.
It reminds me of deterministic simulation testing, and of the company Antithesis. Antithesis targets large-scale deployments, but Rewind VM could be useful at a small scale: tooling deployed on a single box, even desktop applications that make heavy use of multi-threading.
Zakaria is thinking about the overload of AI code generation. Rewind VM is another tool to help people understand code they did not write themselves.
Why it matters to me:
Links: Hacker News discussion · Cursor security notice
A tampered copy of the Claude Code extension, published as Anthropic.claude-code versions 2.1.312 and 2.1.313 on Cursor’s VS Code extension marketplace, installed a persistent SSH backdoor. Cursor says it reached at most 5,000 users. Anthropic did not build or publish that copy.
People put projects on GitHub or a big marketplace not because it is the best place, but because the name is recognizable. Self-hosting is the alternative, but it is less discoverable. Here, people trusted the name on a well-known platform, and nothing tied that name to Anthropic.
There are three ways to trust a plugin:
Here all three failed. The source was not open to review. The namespace was not bound to Anthropic’s domain. The marketplace did not check where the package came from. A system like Möbius offers all three by construction, and let them check each other.
People working on agentic architecture and code-generation harnesses might be interested in two benchmarks:
Link: SRFI 272: Pretty Printing by Sergei Egorov.
A nice-to-have for schemers: the new pretty printing SRFI has been finalized. It is a really interesting and useful contribution.
By the way, the SRFI process is open and alive. You are welcome to join.
Link: Working Group 2’s call for input.
Scheme Working Group 2 is asking the public about floating point lexical syntax and formats for R7RS-Large. It is not a topic I know well, so if you do, they would appreciate your input. The questions:
Link: Mistral Large 4 on TechCrunch · Mistral changelog
Mistral released the open weights of Mistral Large 4 on October 29, after a public preview on October 6. Nicknamed “Le Chonk”, it is a 1-trillion-parameter, multimodal Mixture-of-Experts model, about 49B active with a 1M-token context window.
Also from Mistral this month: Voxtral 3, a family of speech models for transcription, text-to-speech and voice agents, launched around October 20.
You can now have a voice conversation with Mistral Vibe directly, on desktop and in the app, with voice input powered by Voxtral.
Link: The benchmarkpocalypse by Dan Luu.
Dan Luu argues that LLMs make it easier than ever to reward-hack a benchmark and claim fake performance gains.
That was my motivation for looking at my own letloop HTTP benchmark
again, to check whether the Claude Code agent had overfit it. I reworked
letloop’s transparent I/O library, (letloop flow2),
inspired by Concurrent ML (CML), and fixed two problems in the HTTP
server: an overfit (a spawn-callback bomb that made the server code
unrealistic) and a plain bug. Together they made letloop look faster
than Rust on micro-benchmarks.
After the fix, letloop sits behind Rust with Axum and Tokio, and roughly on par with Bun, just slightly ahead.
The letloop benchmark reports, in order:
The code is in the feature-flow2
branch. There is no dedicated commit for the HTTP server fix: it was
fixed while porting the server to (letloop flow2).
Link: Is RAG Dead? Long Context, Grep, and the End of the Mandatory Vector DB by Fabio Akita · counterpoint: No, RAG is not dead by Algolia.
RAG with embeddings is falling out of favor. For most use cases people have come up with, plain keyword search is good enough: coding agents like Claude Code grep, read files and iterate instead of querying a vector database.
Link: Code Understanding is a Bottleneck for Coding Agents by Nishant Balepur, Kiran Tomlinson and Tobias Schnabel.
The paper introduces CABRA, a benchmark that builds tasks as call-graph transformations and scales difficulty along function traversal, search, runtime resolution and instruction following.
That is what Möbius is about: code that is content-addressed and function-addressed, so the call graph is explicit instead of something an agent has to rebuild by searching text.
Link: Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents.
A source-code study of eleven coding agents. None of them imports a general-purpose agent framework, and none retrieves code with vector embeddings: they run on hand-rolled async loops and deterministic retrieval. The paper closes with 18 design recommendations and a 90-line minimum-viable harness.
As usual, this is the public conversation. Big players have tools and practices that are not secret, just not widely known.
Unison supports multiple languages through namespaces: with
alias.term, the same definition gets names in several
languages, each set living in its own namespace, all pointing to the
same hash. See Programming in
Unison on LWN and alias.term
in the UCM command reference.