Andrew Siah
@theandrewsiah
Evals are all you need.
Claude/Codex knows "how" to do anything.
The gap is knowing "what" to do.
Coding agents fail when the "what" is underspecified and unverifiable.
Evals bridge that gap of "what".
I'm hosting a series of Reading Groups on Evals in South Park.
Starting
Claude/Codex knows "how" to do anything.
The gap is knowing "what" to do.
Coding agents fail when the "what" is underspecified and unverifiable.
Evals bridge that gap of "what".
I'm hosting a series of Reading Groups on Evals in South Park.
Starting
0:22
Brendan (can/do) 3 22