points by Kim_Bruning 17 hours ago | 0 comments
The opposite is more interesting, (and the author points it out too). SOTA agents can help run small experiments and help you do the writeup. Should you publish it? I'd personally want to keep a human at least partially in the loop, or have very solid checkers. But for one off personal research it's at the same point as one off personal code: make sure you have a way to test what's valid; but if you do, it often Just Works(tm) Now.
[–] devmor 16 hours ago
[–] Kim_Bruning 16 hours ago
You're not wrong. Up till recently I'd have agreed without reservation even.
Except in recent months, I've seen modern agents verify things more rigorously than I do now.
Could be a me thing ofc, but I've found that a human's eyes tend to glaze over while the agent just keeps on checking in more and more detail.
[–] devmor 9 hours ago
With the caveat that I do not use AI for research, I regularly notice it catch things I would not in code reviews. I also regularly catch a lot of things that the AI would not.
Most importantly, it’s bad at catching things that are not normally part of documented processes or openly reasoned about. I think that this specific weakness is extremely important and exactly how you end up letting something go by that would be trivial to catch for a human reviewer.
[–] suddenlybananas 16 hours ago
Quis custodiet ipsos custodes?
If you don’t rigorously verify everything by hand before then, you are telling the world that they must do what you should have done for you. Or more likely, you are telling the world that that you expect free intellectual labor and will be ignored and chastised- and cause more scrutiny to fall upon others who did the work properly.