Robert Knight

@robertknight2.bsky.social

Machine learning engineer - https://github.com/robertknight.

Version 0.25 of the RTen ONNX runtime has been released. This includes support for several widely-used non-standard "contrib" ops, attention-related operators and bug fixes/compatibility improvements found by AI-assisted testing of over a thousand ONNX models from Hugging Face.

Agents are very useful for large-scale verification of a new implementation of something against existing real-world data. I used Claude to pull hundreds of ONNX models from Hugging Face and run them against RTen. This has found panics, spec edge cases and helped prioritize unimplemented features.

xd009642.github.io/2026/07/21/o... ohno 6 days later and I'm posting again. A one line change at work lead to an ~84% cloud cost saving and lead to me learning some slightly surprising things about how onnxruntime works. Plus a shoutout to rten an alternative in Rust that also solves the issue.

Onnxruntime and the Pit of Failure

Software engineer doing stuff in Rust, recovering C++ programmer. Using this as a place to talk about my current projects including the Rust code coverage tool tarpaulin

xd009642.github.io

Of all the manual or micro-managing corrections I find myself needing to make to Claude-generated code nowadays, the vast majority are to docstrings, comments and commit messages. On the surface this seems like a superficial issue, just a quirk that it tends to be verbose.

We are THRILLED to announce that our Ecma standards will now use UUIDs instead of integers. ECMA-262 is now ECMA-39c506d7-731f-4220-8ffd-a68970945a65. Far more scalable. No two specs will ever collide again. Perfect system. Shortened forms like ECMA-39c506d7 are acceptable in casual conversation.