Haskell 💜 Vibes / Jappie

ocramz:

Or you could use it as a high-level design partner when you’re out of ideas for a gnarly integration.

Do you have examples of this?

I find this a less interesting use case, but maybe it’s just me. I don’t design things I have no clue about, so the difficulty isn’t so much about coming up with ideas, but about making decisions.

And when it’s about making decisions, I prefer talking to humans with real experience. It’s also more fun.

ocramz:

It’s very hard to predict what you can or cannot do with LLMs because their output is completely contextual : it literally depends on 1. coding harness, 2. system prompt, 3. your codebase, 4. your request’s wording.

Their output is chaotic in a mathematical sense. Sometimes rewording a question with no additional information causes wildly different results.

ocramz:

To say that all of this is a net negative for the craft is disingenous.

I think that mostly depends on how people use it. My suspicion is that the average LLM user isn’t driven by curiosity and mindful experimentation like perhaps you are.

Many of us have been deeply frustrated by:

• AI companies shoving their tools into our face at every opportunity
• employers starting to encourage or sometimes even demand AI use
• poorly generated patches/documentation that I have to review even more carefully
• a general lack of sensitivity towards the limitations of those tools
• a decline in human collaboration
• everything being about “productivity”

I think it’s very well possible that it’s going to be a net negative. Maybe that also depends on what behavior we encourage/enable as an engineering community.