> The bottleneck isn’t code production, it is judgment.
It always surprises me that this isn't obvious to everyone. If AI wrote 100% of the code that I do at work, I wouldn't get any more work done because writing the code is usually the easy part.
It always surprises me that this isn't obvious to everyone. If AI wrote 100% of the code that I do at work, I wouldn't get any more work done because writing the code is usually the easy part.
A shift to not writing code (which is apparently sometimes possible now) and managing AI agents instead is a pretty major industry change.
It's like how every job requires math if you make it far enough.
Imperfectly fixing obvious problems in our processes could gain us 20%, easy.
Which one are we focusing on? AI. Duh.
All I had to do was a two line prompt, and accept the pull request. It probably took 10 minutes out of my day, which was mostly the people I was helping explaining what they thought was wrong. I think it might've taken me all day if I had to go through all the code and the documentation and fixed it. It might have taken me a couple of days because I probably would've made it less insane.
For other tasks, like when I'm working on embedded software using AI would slow me down significantly. Except when the specifications are in German.
All OSS has been ingested, and all the discussion in forum like this about it, and the personal blog posts and newsletters about it; and the bug tracking; and theh pull requests, and...
and training etc. is only going to get better and filtering out what is "best."
At best, what I find online are basic day 1 tutorials and proof on concept stuff. None of it could be used in production where we actually need to handle errors and possible failure situations.
There is barely anything that qualifies as documentation that they are willing to provide under NDA for lock-in reasons/laziness (ERPish sort of thing narrowly designed for the specific sector, and more or less in a duopoly).
The difficulty in developing solutions is 95% understanding business processes/requirements. I suspect this kind of thing becomes more common the further you get from a "software company” into specific industry niches.
If they could write exactly what I wanted but faster, I'd probably stop writing code any other way at all because that would just be a free win with no downside even though the win might be small! They don't write exactly what I want though, so the tradeoff is whether the amount of time they save me writing it is lost from the extra time debugging the code they wrote rather than my own. It's not clear to me that the code produced by an LLM right now is going to be close enough to correct enough of the time that this will be a net increase in efficiency for me. Most of the arguments I've seen for why I might want to consider investing more of my own time into learning these tools seem to be based on extrapolation of trends to up to this point, but it's still not clear to me that it's likely that they'll become good enough to reach a positive ROI for me any time soon. Maybe if the effort to actually start using them more heavily was lower I'd be willing to try it, but from what I can tell, it would take a decent amount of work for me to get the point where I'm even producing anything close to what I'm currently producing, and I don't really see the point of doing that if it's still an open question if it will ever close the remaining gap.
Never is a very strong word. I'm not a terribly fast typist but I intentionally trained to be faster because at times I wanted to whip out some stuff and the thought of typing it all out just annoyed me since it took too long. I think typing speed matters and saying it doesn't is a lie. At the very least if you have a faster baseline then typing stuff is more relaxing instead of just a chore.
How many hours per week did you spend coding on your most recent project? If you could do something else during that time, and the code still got written, what would you do?
Or are you saying that you believe you can't get that code written without spending an equivalent amount of time describing your judgments?
So reducing the part where I go from abstract system to concrete implementation only saves me time spent typing, while at the same time decoupling me from understanding whether the code actually implements the system I have in mind. To recover that coupling, I need to read the code and understand what it does, which is often slower than just typing it myself.
And to even express the system to the code generator in the first place still requires me to mentally bridge the gap between the goal and the system that will achieve that goal, so it doesn't save me any time there.
The exceptions are things where I literally don't care whether the outputs are actually correct, or they're things that I can rely on external tools to verify (e.g. generating conformance tests), or they're tiny boilerplate autocomplete snippets that aren't trying to do anything subtle or interesting.
Yes, there is artistry, craftsmanship, and "beautiful code" which shouldn't be overlooked. But I believe that beautiful code comes from solid ideas, and that ugly code comes from flawed ideas. So, as long as the (human-constructed) idea is good, the code (whether it is human-typed or AI-generated) should end up beautiful.
My judgement is built in to the time it takes me to code. I think I would be spending the same amount of time doing that while reviewing the AI code to make sure it isn't doing something silly (even if it does technically work.)
A friend of mine recently switched jobs from Amazon to a small AI startup where he uses AI heavily to write code. He says it's improved his productivity 5x, but I don't really think that's the AI. I think it's (mostly) the lack of bureaucracy in his small 2 or 3 person company.
I'm very dubious about claims that AI can improve productivity so much because that just hasn't been my experience. Maybe I'm just bad at using it.
Speed of typing code is not all that different than the speed of typing English, even accounting for the volume expansion of English -> <favorite programming language>. And then, of course, there is the new extra cost of then reading and understanding whatever code the AI wrote.
I guess we find out which software products just need to be 'good enough' and which need to match the vision precisely.
Okay, you've switched to English. The speed of typing the actual tokens is just about the same but...
The standard library is FUCKING HUGE!
Every concept that you have ever read about? Every professional term, every weird thing that gestures at a whole chunk of complexity/functionality ... Now, if I say something to my LLM like:
> Consider the dimensional twins problem -- how're we gonna differentiate torque from energy here?
I'm able to ... "from physics import Torque, Energy, dimensional_analysis" And that part of the stdlib was written in 1922 by Bridgman!
And extremely buggy, and impossible to debug, and does not accept or fix bug reports.
AI is like an extremely enthusiastic junior engineer that never learns or improves in any way based on your feedback.
I love working with junior engineers. One of the best parts about working with junior engineers is that they learn and become progressively more experienced as time goes on. AI doesn't.
And come on: AI definitely will become better as time goes on.
It’s sort of the opposite: You don’t get to the proper judgement without playing through the possibilities in your mind, part of which is accomplished by spending time coding.