Coding, as we knew it, has forever changed

Back when I was terribly naïve

When I was younger, and definitely naïve, I was so looking forward to AI, which will help us write lots of good, reliable code faster. Well, principally me, not thinking what impact it will have industry-wide. Other more general concerns, like societal issues, role of humans in the future and so on were totally not on my radar.

At the same time, I didn’t expect this will actually happen. Even years later, things didn’t change dramatically. Even the first release of ChatGPT a few years back didn’t click for me, as the limitations were still significant.

Hints of serious change

The first hint of the change, for me, was when a few months ago (yes, behind the curve), I asked ChatGPT to re-explain a concept to me, and it just wrote a lot of words, but without a clear explanation. On a whim, I asked Grok—then recently launched, I think—to do the same. And for the first time, the explanation clicked and I felt I could have a conversation with it. Of course, now I forgot again that theoretical CS concept, but the first step was done: I can ask an LLM to explain something, and it will, and I can have a back and forth logical discussion, even if on some theoretical concept. Additionally, I learned that not all LLMs are the same, and that means there’s real competition and that leap frogging is possible.

Another topic on which I tried to adopt early and failed to get mileage out of it, was GitHub Copilot (in VSC). I tried, it helped, but didn’t feel any speed-up at all. Then more recently, in May, I asked Grok what’s the state of the art in AI-assisted coding. It said either Claude in a browser tab, or in VSC via continue.dev extension.

The continue.dev extension/tooling is a bit of a strange/interesting thing. It seems to want to be a middle-man between the user and actual LLM services, i.e. you pay a subscription to continue.dev, not to Anthropic itself, and they manage the keys/APIs, for whatever backend LLMs you want to use. The integration with Visual Studio Code is very nice, but I don’t know if long-term their business model will make sense. Well, not my problem.

Claude: reverse engineering my old code and teaching new concepts

So I installed the latter and subscribed, thinking 20 CHF for a month is good for testing. I skipped the tutorial model/assistant, created a new one from scratch, just enabled Claude 3.7 Sonnet, and started using it. And then, my mind was blown-not just by the LLM, but by the ecosystem. As said, I’ve used GitHub copilot before, but it didn’t seem effective. I don’t know if a threshold has been reached, or Claude (3.7 at that time) is just better than ChatGPT.

I didn’t use the AI to write (non-trivial) code for me, at most boilerplate snippets. But I used it both as partner for discussion - “I want to do x, what do you think, A or B?”, and as a teacher, especially for fronted topics, which I’m not familiar with.

Since May, in mostly fragmented sessions, I’ve achieved more than in the last two years. Migration from old school JS to ECMA modules, a webpacker (reducing bundle size by 50%), replacing an old Javascript library with hand written code using modern APIs, implementing the zoom feature together with all of keyboard, mouse, touchpad and touchscreen support, simplifying layout from manually computed to automatic layout, and finding a bug in webkit for which it also wrote a cool minimal test (cool, as in, way better than I’d have ever, ever written, because for me it didn’t matter that much). And more. Could I have done all this? Yes, definitely, nothing was especially tricky here. But hours and hours of reading MDN, scouring Stack Overflow and Reddit, and lots of trial and error. So doable, but much more toily.

This, to me, feels like cheating. 20 CHF per month to make me 3x more productive is free money—well, except that I don’t make money on my code which is written basically for myself. However, I don’t get stuck anymore searching hours in the web for guidance, I ask my question, and I get at least direction if not answer, and I’m finished way earlier. I can now actually juggle more hobbies, in the same amount of time, if my personal code takes less time or differently said, if I’m more efficient at it.

Not all is roses, of course. Once, it did write code with such an endearing error that it made me laugh. It was so blatantly obvious that you shouldn’t keep other state in the array that holds pointer status because that confuses the calculation of “how many pointers are down”, probably to itself too if I’d have asked. But I didn’t, since it felt a bit embarassing to point out such a dumb mistake. Yes, I’m anthropomorphising again, because this is the easiest way to deal with things.

In general, it does an OK-to-good-to-sometimes-awesome job, and the best thing is that it summarises documentation and all of Reddit and Stack Overflow. And gives links to those.

Now, I have no idea yet what this means for the job of a software engineer. If on open source code, my own code, it makes me 3x faster—reverse engineering my code from 10 years ago is no small feat—for working on large codebases, it should do at least the same, if not more.

As an example of how open-ended the assistance can be, at one point, I started implementing a new feature—threading a new attribute to a large number of call points. This is not complex at all, just add a new field to a Haskell record, and modifying everything to take it into account, populate it, merge it when merging the data structures, etc. The code is not complex, tending toward boilerplate a bit, and I was wondering on a few possible choices for implementation, so, with just a few lines of code written that were not even compiling, I asked “I want to add a new feature, should I do A or B if I want it to behave like this”, and the answer was something along the lines of “I see you want to add the specific feature I was working on, but the implementation is incomplete, you still need to to X, Y and Z”. My mind was blown at this point, as I thought, if the code doesn’t compile, surely the computer won’t be able to parse it, but this is not a program, this is an LLM, so of course it could read it kind of as a human would. Again, the code complexity is not great, but the fact that it was able to read a half-written patch, understand what I was working towards, and reason about, was mind-blowing, and scary. Like always.

Non-code writing

Now, after all this, while writing a recent blog post, I thought—this is going to be public anyway, so let me ask Claude what it thinks about it. And I was very surprised, again: gone was all the pain of rereading three times my post to catch typos (easy) or phrasing structure issues. It gave me very clearly points, and helped me cut 30-40% of the total time. So not only coding, but word smithing too is changed. If I were an author, I’d be delighted (and scared). Here is the overall reply it gave me:

• Spelling and grammar fixes, all of them on point except one mistake (I claimed I didn’t capitalize one word, but I did). To the level of a good grammar checker.
• Flow Suggestions, which was way beyond normal spelling and grammar. It felt like a teacher telling me to do better in my writing, i.e. nitpicking on things that actually were true even if they’d still work. I.e. lousy phrase structure, still understandable, but lousy nevertheless.
• Other notes: an overall summary. This was mostly just praising my post 😅. I wish LLMs were not so focused on “praise the user”.

So yeah, this speeds me up to about 2x on writing blog posts, too. It definitely feels not fair.

Wither the future?

After all this, I’m a bit flabbergasted. Gone are the 2000’s with code without unittests, gone are the 2010’s without CI/CD, and now, mid-2020’s, gone is the lone programmer that scours the internet to learn new things, alone?

What this all means for our skills in software development, I have no idea, except I know things have irreversibly changed (a butlerian jihad aside). Do I learn better with a dedicated tutor even if I don’t fight with the problem for so long? Or is struggling in finding good docs the main method of learning? I don’t know yet. I feel like I understand the topics I’m discussing with the AI, but who knows in reality what it will mean long term in terms of “stickiness” of learning. For the better, or for worse, things have changed. After all the advances over the last five centuries in mechanical sciences, it has now come to some aspects of the intellectual work.

Maybe this is the answer to the ever-growing complexity of tech stacks? I.e. a return of the lone programmer that builds things end-to-end, but with AI taming the complexity added in the last 25 years? I can dream, of course, but this also means that the industry overall will increase in complexity even more, because large companies tend to do that, so maybe a net effect of not much…

One thing I did learn so far is that my expectation that AI (at this level) will only help junior/beginner people, i.e. it would flatten the skills band, is not true. I think AI can speed up at least the middle band, likely the middle top band, I don’t know about the 10x programmers (I’m not one of them). So, my question about AI now is how to best use it, not to lament how all my learning (90% self learning, to be clear) is obsolete. No, it isn’t. AI helps me start and finish one migration (that I delayed for ages), then start the second, in the same day.

At the end of this—a bit rambling—reflection on the past month and a half, I still have many questions about AI and humanity. But one has been answered: yes, “AI”, quotes or no quotes, already has changed this field (producing software), and we’ve not seen the end of it, for sure.