Rob Napier - TIL:AI. Thoughts on AI

I use AI a lot for work, pretty much all day every day. I use coding assistants and custom agents I’ve built. I use AI to help code review changes, dig into bugs, and keep track of my projects. I’ve found lots of things it’s very helpful with, and lots of things it’s terrible at. If there’s one thing I have definitely learned: it does not work the way I imagined.

@cocoaphony These anecdotal “I wasn’t sure about AI so I tried myself & I didn’t get cancer, so the anti-AI people are just paranoid” experiences are not as helpful as they seem.

@baldur has written about this in a way that is useful, I think.

https://toot.cafe/@baldur/114664474072967489

https://toot.cafe/@baldur/114654199537177678

Baldur Bjarnason (@[email protected])

Avoiding generative models is the rational and responsible thing to do – follow-up to “Trusting your own judgement on ‘AI’”: https://www.baldurbjarnason.com/2025/followup-on-trusting-your-own-judgement/

Toot Café
@cocoaphony really well-written, I’ve learned a few things. Thanks!
Hailey (@[email protected])

Attached: 1 image Remarkable chart in a new paper out of MIT looking at cognitive performance of LLM users in an essay writing task - https://arxiv.org/abs/2506.08872

hails.org
@cocoaphony This post was really timely for me - today I had to refactor around 20 call sites of an API where the refactor was a bit more complex than I could easily search and replace, so I had AI do it. I was unpleasantly surprised by a) how long it took and b) how it hallucinated a couple of the results, though most were correct. The idea of asking it to write a script to do the refactor and then execute the script is such a good one, and I’ll be sure to try it next time!
@cocoaphony Great article! Your experience matches my own, but you're definitely further down the learning curve than I am. Thanks for writing down your thoughts!
@cocoaphony Pretty spot on for me too! Love the Memento analogy
@cocoaphony Thank you for the nice summary and explanation! I already shared the article to some of my non-developing friends just to explain how the "magic" really works. It's way too much hype and ambiguity in news about AI right now and pragmatic view really helps to get sense out of it.
@cocoaphony I decided this week to give Copilot a chance on some small coding tasks. First was a fairly standard garbage JavaScript task I was going to throw out anyway. Second was a slightly larger, and more permanent, bit of React code that - importantly! - I had already implemented this pattern elsewhere, and was just looking to have it done for speed’s sake, so I could spot any shenanigans in the code. In both cases, I treated it exactly as you described.