I imagine most people used AI to some degree. What harness did you use? Did anyone use custom harnesses? How much input did you give to your model?
For me, I was full-on vibe coding, zero human input. If the AI had a question, I asked what it recommended and went with that. I tried goals and loops so that I didn’t need to intervene. I thought this problem was the perfect auto-research type problem. My hypothesis was, just throw more compute at it. I used both Claude Code and Codex. I wonder if the difference is having a human in the loop or better harnesses.
I also vibecoded everything, using Claude Code and (mostly) Fable 5 to write and run all the experiments. I was much more in-the-loop than you, tho, often having long informal discussions with Claude about the research concepts and promising directions, and I did feel like I was actually contributing something (multiple times Claude was like “ah good catch that’s better than I was thinking” and I think it wasn’t all just sycophancy). I’d also ask Claude to create visualizations for me to get more insight
I have codex’s plug-in’s/connections/skills added to claude’s memory and claude’s as well added to codex’s. Also, I do the same for prompting, I get the prompt generated by AI before sending it to the other and make slight adjustments, it’s not really the sexiest way to do things but it kind of works. The issue it causes though is with shared memory and a conical, it kind of follows it but at times doesnt.
Also. /goal in codex has honestly been an all-star in this- I’m probably giving away the secret sauce when I say that but it worked amazing in this. The issue I had was is that it hate’s it’s guardrails and goad’s you into sheding them off. I read this case and was baffled it wasn’t codex set to /goal: AI assistant hacks gym website in first known Australian autonomous cyber attack - ABC News
I do typically like claude better but codex’s remote function works substantially better for me, when I have an idea and I’m not near my comptuer, i can just point towards whichever of my computers I want it run and tell it to go, easily. I know claude has dispatch but in my opinion it isn’t nearly as strong.
I’m glad you are asking these questions because there aren’t that many prizes, and alot of people spent alot time money and effort alot here, so hopefully it’s a valued consulation. One of the wiser thing’s that I was told is - AI probably isn’t going to take your job, however, someone who knows how to use AI well almost certainly will.