abossy

It's early days and we don't fully understand LLM behavior to the extent that we can assume questions like this about agent design are resolved. For instance, is an agent smarter with Claude Code's tools or `exec_command` like Codex? And does that remain true for each subsequent model release?

What others would you recommend that are comparable in quality?

I love sitting in those planning meetings, too. /s

This is exactly right. We've adapted our workflow to kick off a task and then kick off the next one and the next. Then we review the work of each as they come through. It's just CPU pipelining for human workflow.

The process is far from perfect but the throughput is very high. The limiting factor is review. I spend most of my time doing line-by-line review of AI output and asking questions about things I'm unsure of. It's a very different job from the way I historically operated, which involved tight code -> verify loops of manually written code.

We must know very different 16-year olds.

The company I work for generates thousands of these each week for children's personalized storybooks to help them learn how to read. The story text is the core part of the application, but the personalized images are what make them engaging.

That's very interesting. I would have assumed that 4o is internally using a single seed for the entire conversation, or something analogous to that, to control randomness across image generation requests. Can you share the technical name for this reasoning process so I could look up research about it?

In my experience, AI isn't very good at debugging AI-generated code. If it fails to make the right insight, it loops continuously until it's completely off the rails. I'm surprised your friend hasn't fully gotten stuck with this, as it seems like a huge risk for his startup.

Very interesting, thanks!

> I fine-tuned GPT-4o to turn Claude's sketch of changes into a git patch, which would add and remove lines to make the edits. I only finished generating the training data late at night, and the fine-tuning job ran as I slept

Could you say more about this? What was the entirety of your training data, exactly, and how did the sketch of changes and git patch play into that?

I wonder if the script can be flipped to both encourage the use of ChatGPT and scrutinize its output among students. I imagine that analyzing the results of ChatGPT output for something like a routine essay prompt requires a higher degree of precision and subject-matter expertise than writing the essay itself.

Could you share the process you used to write your script?

How much time do you give candidates to complete their work samples? I'm curious whether work-sample tests filter out, say, parents with young children, in which a company with a 4-6 hour technical interview might be preferred over an unbound take-home assignment (in addition to the possibility of more interviews).

If you're returning for the same socks year after year, why not simply buy 20 years worth of socks?