I just want computer-assisted, approximate pattern recognition.
I would like to be able to say, show me in the text I've written what my tics are, what parts of this code I'm writing look similar to other code already written, where it looks like a common error, an old bug, like a frequently-used command I can alias, a habit I might want to break, a cliche to avoid.
I don't want the computer to create more almost-work repetitiion, I want it to help me see the work I've done clearly.
@grimmware @benjamingeer You joke, but one of the biggest differences between junior engineers and senior engineers is precisely how they feel about that.
Juniors who brag about how much code they're writing just do not impress senior engineers who talk about how much redundancy they can prune out or simplify in a codebase.
@benjamingeer @mhoye I think a lot of people who aren't technical think that more code = better/more capable, but really it's like more code = more shit that can break
"The engineer knows their work is finished not when there is nothing left to be added, but when there is nothing left to be taken away."
@mhoye saw this initially in isolation via boost and interpreted the "code" part of "generate more code faster" as "machine code", and i was like...
"but... but... faster codegen gives you more edit/compile iterations per hour, which means you get to experiment more, which helps you to explore more of the solution space, which..."
then i scrolled up and realized you meant "gen-a.i.-vomited source code", in which case: yes; agreed.
> "What you've created is an entropy spigot pointed at the proxy metric graph we're using because your management doesn't understand anything."
I want this translated into Latin and tattooed on my bum.
The flourish with which I shall despatch the opposition will make me the darling of the debating club.
"Id quod creavisti est puteus entropiae ad graphium mensurae substitutae directus, quem adhibemus quia administratio tua nihil intellegit."
The student asked the master, "how should one develop code?"
The master replied "the novice always create more code; the middling maintain code; the skillful eliminate code"
The student then asked "but what does the sage do?"
The master asked "with what?"
At that moment, the student was enlightened.
@mhoye This is the fundamental problem
Every previous advance in computing was an abstraction. You delegate some thinking and testing, and get to be responsible for less code.
LLMs increase the volume of code that you personally are responsible for. And, at present, it’s easier for them to help you generate stuff than help you understand what you’ve done.
There is a trap here but the tool providers must be aware of it. I’m not aware of any theoretical reason why they can’t improve.
@neilk @mhoye "the tool providers must be aware of it" Must they? I have seen no evidence of that yet. 🤷
"... not aware of any theoretical reason why they can't improve" This strikes me as true but unhelpful? Like, LLM code generation seems genuinely surprising to 2020-me; I personally am not optimistic that it will turn out to be genuinely useful for programmers, but that is a live debate.
But regardless, it isn't even *tackling* the "help you understand the codebase" problem.
@esnyder @mhoye I have worked for large corporations and also been on the outside making disruptive / open source tech
100% of the time, many if not most of the people at the big corp are fully aware of the problems with their product. 100%.
Unlike the rest of the world, the big corp has a product pipeline, complex internal politics, existing revenue streams, message discipline, paid PR, a sympathetic business press, and an investor story to maintain.
@esnyder @mhoye Consider what would happen if OpenAI declared tomorrow that they’re not really sure how to manage code gen and many customers are struggling with adoption
Also consider the case where they are racing like mad to solve this, and think they have an answer, but it isn’t ready yet. They not only can’t announce it, they can’t even let on they know it’s a problem
I'm wondering, from anyone in the industry, what percentage of the time programmers/coders spend generating new code for new problems vs percentage of time spent fixing, improving, implementing and testing?
This is an ancient complaint. IBM, near the beginning of coding, wondered how to assess productivity and decided to do so with "k-locs": thousands of lines of code. The people around the birth of Microsoft were properly scornful of this measure.
But it keeps coming back.
@mhoye 🤷Actually, I've been experimenting, and that's not exactly true.
Describing to an AI assistant what you want usually takes longer than writing the code.
It might result in better thought out code because you have to verbalize the requirements, and the idiot intern will almost certainly take every shortcut that you did not specify explicitly in your description.
But for stuff where I don't have to learn the language & environment at the same time as coding, I'm surprisingly faster solo.
@grimalkina We talked the other day about deliberate and structured practice, and I wonder basically all the time now what I could learn about myself if I could actually see my efforts the way the machine sees them, from a wider view than I can muster, the patterns I can't see because I'm too close or too far away from them, how I could share what I learned with other people also trying to learn.
Computers could be so good for us.
@mhoye I think there are many opportunities like this to make really clever improvements to existing tools with AI, especially around search. The problem is that you need to really understand the models and take them apart and put them back together again in very specific configurations.
That takes a lot more skill than just calling the ChatGPT API with "write this code for me".
@mhoye I think I disagree.
Untransparent training data is bad ethically, but these models are representation learners as well as generators. The text embeddings they provide are incredibly powerful, well-disentangled and pretty objective.
You could easily use them to do lookups for the things you described without ever sampling any text.
@mhoye A model like Comma (mostly) solves the data issue, and isn't too big.
@mhoye *lots* of what people describe AI-code-assistance as helping them with, I would have approached with code slushpiles/skeletons/personal comments/libraries. Efficiency, but increasing my understanding, maybe even my team's understanding or everyone's, if it got to "library".
My take is that that got outsourced to Stack Overflow just long enough to be forgotten.
@mhoye started researching something similar to that a couple of years ago
gave up due to the reception ("oh cool this will enable more AI-based code to-" ok we're done, not continuing this project)
mhoye
Benjamin Geer
grimmware
sidereal
FoolishOwl
Strider Uwe 🇺🇦🇨🇦🇲🇽
Joe ❌👑
James Widman
doboprobodyne
Georg Greve
Aaron
Brian Swetland
Eric McCorkle
Neil Kandalgaonkar
Emile Snyder
Ignas Kiela
Frang 
Emily C
WandreCanada
zenkat
JackPearse
The Frog
Rich Puchalsky ⩜⃝
Andreas K
lachlan slowly taming rust
BrazMogu (i.e. Bruno Guedes)
Cat Hicks
kit 🌬️🌦️
Peter Bloem
Dr David Mills
Jonty Wareing
clew
Alex Celeste, Princess Consort of Burnout