Cthos @cthos.dev · Jan 1 I cannot believe I am doing this on NYE, but a few things: 1) Opus 4.5 doesn't appear in the paper PDF. 2) The JSON they're publishing on that page reveals that the low confidence is 1.8 hours all the way up to 20 hours. The json also has a USD field which I assume is cost, and, it's high.
4 likes 2 replies
?
Replies
Yes. The PDF is from March. They've updated the graph since then but not the research methodology paper. And yes. Good call out on very large error bars.
Augh, s/JSON/YAML. I'm tired. Anyhow, I don't think you can confidently claim that that is what they're claiming based on that graph alone.