Originally published on LinkedIn. View the original post ↗

Last night, Dave Sifry’ gave an Ignite talk and made a good case for trying a modest model for AI code review.

In his core test, GLM-5.3-Flash found 71 of 147 test-verified bugs across six pull requests. Opus 5 found 74 using the same review process. The average cost was under 3 cents per review for Flash, versus $2.97 for Opus.

No code review is perfect, and I’ll want to give this my own eval. But getting that close at that price is thought provoking. I hadn't assumed I needed to bring Fable/Astra to every code review, but I definitely hadn't tested giving it the small ones.

The ignite talks were definitely a highlight of The AI Conference for me.

Dave's results:
https://lnkd.in/eJqKmNsw

AI-assisted developmentCode reviewEvaluation
Alexander Thomas
CTO. Architect. Builder.

Keep thinking.

2 min read
Engineering leadershipLinkedIn post

Your test strategy determines your speed limit

I've never played Go. I've played enough chess to appreciate its complexity, but Go is orders of magnitude more complex by possible positions. More than the number of atoms in the universe.