There are countless examples of people up-skilling using AI. Today, that might threaten the bottom of the market. Soon, it will put everyone’s jobs at risk for disruption.
[0] https://hbr.org/2023/08/ai-wont-replace-humans-but-humans-wi...
"AI"-based static analysis tools might be valuable. If I ever see "AI"-based code generation on a project I run before I get stuck cleaning up the mess.
LLMs don't just dash out hundreds of lines of code. They can take a concept and break it down into concrete steps or even a fully planned architecture for you.
That's literally the hard work you describe. AI can do it beautifully, it just needs a human capable of critical thought to steer it (human understanding of the subject at hand is merely a bonus, not a requirement).
The separation will very clearly be “legacy devs” vs “AI-powered devs”.
I’ve spent a tremendous amount of time thinking about how the problems should be solved on my current project — big picture, and small. Once the system design became clear, about 80% remains as “just write the damn code” type work. I outsource about 70% of that to GPT, while keeping my focus on the important parts; the right abstractions, modular and self-contained functions, good naming and function signatures, making sure it all works and makes sense in harmony.
I don’t care (too much) about the inner workings of every function, so long as the abstraction and usage of it is great (and well tested). Refactoring, if ever needed, is trivial then.
My overall work has never been better.
Anybody dismissing AI for any level of programming is doing themselves a huge disservice.
Could elaborate more on your workflow? Do you use Copilot, or just GPT4? Are you copying and pasting large blocks of existing code to hint at how things should fit together, or do you describe how the existing code is structured in English? Do you find yourself decomposing your work differently so as to fit your new AI workflow?
I talk GPT as if it’s a skilled colleague who’s got memory issues. When it forgets, I remind it with stuff like “this is our current code now, remember X, Y, Z”
I’ve become an even bigger fan of small, compostable functions that do one thing very well. GTP excels at that, both writing and testing them.
I don’t involve it much for architecture and high-level design atm (mostly because I got that part solved on my current project). I tend to have a design in mind, give GPT the overview of how it fits together with mock code, and ask it to review with me, propose other paths we could take, etc before proceeding to implement.
One function at the time, with tests. When refactoring happens, it’s usually isolated to a few hundred lines at most. When tests fail, directly or indirectly, I give it the full output and ask it to debug and fix. I find this helps GPT remember the code’s responsibilities when it gets lost/forgets. It also helps me avoid regressions when GPT returns functions that miss use cases we had covered before.
I keep long running chat threads — weeks at times, hundreds or even thousands of messages. The longer we go, the better it tends to perform (web app performance, even on an M2 Studio, does suffer after a while though)
At worst, GPT is a fantastic rubber duck. At best, it’ll help me see superior approaches and solutions I wouldn’t have considered, AND give me perfect code in seconds.
Once tooling gets really good, and AI can understand the whole code base/database/infra… we’ll probably be in real trouble.
You've really changed your workflow to adapt to these new tools. The amount of mental effort seems comparable to learning a new IDE, maybe a bit less.
Instead of the chat app, I wrote a python script that uses the GPT-4 completion API. I can just pop over to the terminal and type 'chat' and it's there. As far as I can tell, it's basically the same as the app.
We are starting to see AI tooling that can fit an entire 100k line code base in its context window. I still see myself having a job five years out. Ten years out, not so much. Luckily I'll be close to retirement.
I was primarily FE for a while, and used to agonize over the smallest details when building eg React components. After a few years, I concluded that the innards don’t really matter most of the time, at least not early on, so I limited my obsessions to the “public interface” (naming, types and prop design) and libraries of well abstracted, composable, low-level building blocks. If and when things needed to be rewritten or optimized, replace the internals.
ChatGPT just gets me there faster, and more often than not, gives me acceptable production-level innards while I get to stay focused on perfecting the exterior and cohesiveness of the overall solution.
I’ve found this approach usually pays dividend when reworking parts of the system, too — GPT picks up on the flow of everything much better when the code is “self documenting”. Same appears true if humans need to get directly involved.
Still blows my mind on a daily that we’re here already. I’m glad I’m not early in my career, I’d be very worried if I had another 40 years to go.
The PR I'm currently working on has about 1K diffs – I've literally touched NONE of that code directly. Freakin' wild!
I don't think there will be many people not using AI, if at all. Just like everyone learned to use Google and Stackoverflow, everyone will learn to prompt a chatbot. And since it's not really rocket science prompting a chatbot, we might see programming becoming a cheap commodity (and after that - pretty much most knowledge work)
Don’t get me wrong — I think this’ll be the end of MANY jobs within almost every field in the next decades, ours included. But I sure as shit can’t beat it, so may as well join it and get something out of it before I become just a commodity.
Yep that's the crux of the issue I agree. It's not just one repo, any well established org has millions of lines of code - can ChatGPT handle that kind of context size without any human directing it ? It's basically having an LLM trained specifically on your company data, and having it retrained constantly since the data keeps changing - what are the costs of that? We'll see.
GPT-4 and Copilot absolutly makes me more productive as a programmer, wiring all the tests, getting maybe 80% of them right and then I fix them up. I am not producing 10,000 lines of crud, but I AM producing 1000 lines of test-covered code, I'm doing the hard bits, and the AIs are doing the menial parts.
I am by far the most productive programmer on the team however you measure it (we have about 60 different metrics available to management) and a large part of this is be having enough energy to always be attacking the hard problems, because I am not wasting mind power and time writing the trivial bits that can be automated.
Does it get it right every time? Nope. Does it make hilariously wrong mistakes? Yep, about once or twice a day! Is it worth it? Absolutly.
Everyone (including management) is aware of all of the tropes and cliches about how metrics cease to be good when they become targets and how metrics can be gamed blah blah blah - we all know that - and have imbibed that, but the rationale is that more (but imperfect) visibility is better than no visibility or solely subjective interpretations.
As long as all metrics are assumed to be wrong until proven correct, with some digging, some thorough analysis and a judicious and abundant use of the benefit of the doubt, I find them really helpful.
1) Humans are terrible predictors of the future.
2) Linear progression isn’t really a thing in multi step, complex technology progressions
The impact of the steam engine was nuclear weapons. Direct line. Predictable?
Spot on analysis.