Two AI-focused engineers with deep experience at places like X, Stripe, and MIT share how large language models helped them sketch prototypes faster yet left behind tangled, low-quality code that demanded far more cleanup than starting from scratch, ultimately erasing any sense of satisfaction from the work even after a notable technical win.
I can relate to this. These LLMs are really great at speeding up prototyping but they are so awfully bad at programming that it sucks out all of the joy..
@KamStaszewski No. I was using it to prototype stuff and then thought I'd clean it up by hand, but some of the stuff that crept in was so moronically bad that it took a whole day of semi-automated bullshitting to get it to normal messy code that I can now manually clean up.
Both engineers describe LLM output as quick to generate but so flawed that reaching usable code still required a full day of fixes, turning what should have been an enjoyable process into extended drudgery.
Suarez reports achieving a major breakthrough where the models contributed no useful ideas at all, leaving him with no enthusiasm for the weeks of effort despite the faster initial prototyping.
Many users say LLMs drain the joy of coding by producing messy code that lacks correctness and demands constant babysitting, while a few appreciate the faster prototyping and lower refactoring costs.
No Digg Deeper questions have been answered for this story yet.
@yacineMTB I have gotten absolutely 0 enjoyment out of my work in the last few weeks. Prototyping may have gone a little bit faster but, even though I made a major breakthrough and codex contributed precisely 0 useful ideas, I still have no enthusiasm for it
I can relate to this. These LLMs are really great at speeding up prototyping but they are so awfully bad at programming that it sucks out all of the joy..
@jsuarez Same..
@yacineMTB I have gotten absolutely 0 enjoyment out of my work in the last few weeks. Prototyping may have gone a little bit faster but, even though I made a major breakthrough and codex contributed precisely 0 useful ideas, I still have no enthusiasm for it
@yacineMTB joy, which is required to get good and move things forward
@yacineMTB Currently on my 10th iteration of a feature. "No. Bad robot."
@yacineMTB totally feel this
LLMs are great at scaffolding but when you need real correctness they fall short the joy of coding is in solving the problem yourself not
@yacineMTB I think using ai as planning + autocomplete is the best way for now,
better models are yet to come, as they cannot accommodate even somewhat complex projects
@yacineMTB What do you mean by « bad at programming »? Do you mean performance-driven? Complexity? And is your use case essentially ML applied to robotics? I’m genuinely interested, because at my level (backend with Python), it’s good, tbh.
@yacineMTB Yeah, it's great for getting started but as soon as it becomes serious you really gotta get in there and fix things. The hard part is intervening before it gets too unwieldy. For now, at least
@yacineMTB Where is all the 'u just suck at prompting' crowd 👀 ?
@yacineMTB Not my Fable. That beast knew how to program with the best of us.
@yacineMTB Your tweets about AI and programming confuse me. One day you are saying it 50x’s your productivity and the next you are saying some variation of this
@yacineMTB programming woo hoo!!!!!!!
@yacineMTB AI is like a credit card. Just pay off the debt frequently and it's fine. And the cost of refactoring is massively lowered.
@yacineMTB Suck at programming? 😅 maybe assembly
@yacineMTB it's the tail end of the distribution. if your code needs to be really really really good, then it just doesn't give you what you need unfortunately.
it's like, what is the slice of all code that it can nicely replace, and it hasn't crept that far yet.
@yacineMTB In your experience, what is the worst pattern you see from codex?
@yacineMTB I already miss the days of only hand coding. Now I just use AI for work/hustle stuff, but hand code games just for the love of it. I wonder how long it will be feasible for though
@denbvk @yacineMTB @jsuarez I think its a personality thing some people just liked handwriting code more and writing tests but the ppl who like working at a higher level clearly do better with claude code.
@jsuarez @yacineMTB The practical hope is code as a construction material that can be graded by quality and deployed according to fitness for purpose, much like particle board vs. solid wood. Raw LLMs provide a cornucopia of "particle board" code which is circumstantially useful but not a panacea.
@yacineMTB they’re not awful you retard