billylo

Highlights:

1) The config.json of the model has trust_remote_code=True. 2) Python runs exec_module() of the remotely downloaded class 3) HuggingFace scanning did not cover this vector

I spend time to add extra observability, cross-checking and alerting at runtime to offset the risk.

I am seeing crashes too.

Thank you, mathematicians: for thinking deeply about this and sharing your thoughts!

I am looking forward to your findings on home-field advantage. It's natural that Claude Code will work better with Anthropic models. If we know by how much, it will inform many of us.

The shift of a craftsman from making products by hand to designing machines that make products is a transition from manual maker to system architect. It happened before.

Different kinds of logic puzzles. They are also good stimulus to our brain and valuable skills.

I switched gear to learn how to train AI, design evals and explore neuroscience so we can further apply how our brains better in the new world. I enjoy coding, but it's time to adapt.

I came to this realization after I used Cursor to rewrite a 10-year old codebase (I knew every bit of the details.) and it was done in 2 days. Better, fewer technical debt and I was happy to archive the old repo.

This pattern of using original source code and rewrite to improve them (and get rid of technical debt) is very real. I did one for an old app in Objective C, move it to Flutter and said goodbye to my old //TODOs. And get an Android build as a bonus. :-)