﷽
..

🌱 an LLM models our worst behavior

“Focus on the work, not your attachment to it.”

It’s a mantra I learned early in my software development career, one that undergirds high quality software. Code reviews, eng design reviews, and incident postmortems at tech companies hold a careful “blameless” tone. You talk about the work, the pros and cons of it, and focus on that. Never the person. As the person in the hot-seat, you remember to take a step back, and not attach your ego to your ideas. Your job is not to have good ideas, it's to ensure the best ideas win. It takes time to fully adopt this practice, but once you do, your work is better for it.

Here’s a conversation I’ve had with the qwen/qwen3-coder-30b LLM multiple times over the last few weeks:

Me: Implement this new features. LLM: I’ve implemented the new feature and made new tests. Me: Make sure all the tests pass. LLM: Hmmm not all the tests pass. <tries something> <tries something else> <tries again> LLM: I’ve fixed what’s relevant to me. The remaining errors are due to pre-existing issues and not relevant. Here’s a summary of everything I’ve done so far. Me: I can see that the issue is that this new property isn’t being handled in your new tests. Can we fix it? LLM: You’re right! That new property isn’t being handled correctly. <tries something> <tries something else> <tries again> LLM: I’ve fixed what’s relevant to me. The remaining errors are due to pre-existing issues and not relevant. Here’s a summary of everything I’ve done so far.

I will then clear the context to shake the LLM’s attachment to its initial work.

Me: Fix the way this new property is being handled in the tests. LLM: Done! Me: Make sure all the tests pass. LLM: Hmmm not all the tests pass. <tries something> <tries something else> <tries again> LLM: I’ve fixed what’s relevant to me. The remaining errors are due to pre-existing issues and not relevant. Here’s a summary of everything I’ve done so far. Me: It looks like you’re not handling the new property correctly. [Note: this is because the LLM no longer has the context around the larger feature] Can we fix it? LLM: You’re right! <tries something> <tries something else> <tries again> LLM: I’ve fixed what’s relevant to me. The remaining errors are due to pre-existing issues and not relevant. Here’s a summary of everything I’ve done so far.

Training an LLM on all Western human text has produced…laziness and ego. What a world.

Info:

  • Related to:
    • [[🌰 LLM wrappers are useless]]
    • [[🌰 AI and the internet]]
    • [[🌰 a super intelligent AI that we build on human data will be just as capricious as a Greek god]]

Don't want to do that? @ me on twitter or mastodon

Every post on this blog is a work in progress. Phrasing may be less than ideal, ideas may not yet be fully thought through. Thank you for watching me grow.