Post

Thoughts on the MIT study and what long-term LLM usage might do to the expertise of developers and their ability to validate its outputs?

Using LLMs effectively in a software engineering job is a skill we are all still developing. Some folks have really advanced techniques with multiple LLMs, and I don’t know what else to call this other than a new type of programming.

It's definitely a jump, but I'd be careful about the sideways movement. Two immediate things come to mind to keep LLM on it's lane -- you'll need an expert review of the code and you need to maintain an extensive test suite -- modified through human review -- to ensure the bot doesnt start drifting.

❤️ the avoidance of right angles for non-deterministic agents😁 Our (upcoming) platform embeds AI's non-deterministic creativity within a logically consistent and self-aware framework, functioning like a biological immune system, identifying and rejecting insecure patterns, anti-patterns, "AI slop".

Interesting introduction, but where will non-deterministic behaviours lead us in terms of programming ? Will we have to adopt / create defensive tools to keep results aligned ? Personally, I don’t find that exciting but more of a (unnecessary?) burden.

That’s a great observation imho the sideways is as important or more important than the up, we simply haven’t figured out what to do with it yet

"can't just store my prompts in git and know that I'll get the same behavior each time...how to live with this non-determinism" Developer checks code generated by AI and thus bears full responsibility for its behavior and outcome. Can't just store prompt because AI does not guarantee correct output.

Non-determinism in software development isn't a good thing. It might be a leap off the page, but it's leaping into art rather than an industry leap ahead. This is why so many commentators are confused and think themselves increasingly irrelevant, whereas they should just be patting LLMs on the head.