How to not use AI to code for you

This is a bit circlekerky and the english is very broken

Bit ironic considering running it through a LLM would've fixed the 2nd part

It seems to mostly be about risk management - knowing how to contain risk? What is the cost of this piece of code failing to do what it is supposed to? Is it easy to replace?

If the answer is low and yes, then speed matters more than perfection and you don't have to always know exactly how it works. We already accept some risk while outsourcing parts of our problems when using dependencies. The risk management puzzle is different but still fundamentally the same?

If you don't know how the thing work in principle though, it's not going to be easy to replace.

It was a fateful day in one of the Indian Monsoon in the late 2000s. I rode the ever-reliable tuk-tuk in the city of Mumbai to reach a coveted Indian outsourcing company that has taken on a multi-year, multi-billion-dollar contract. They had inherited a project and discovered that one of the programmers was me, as indicated by comments inside the code, and they had reached out.

Fortunately for me, they are stuck in the specific styling of an enterprise app that has been in operation for a quarter and is now starting to reflect in their P&L account.

After all the security check-ins through the gates, and after being seated and coffee-ed, I walked to a desk and spent as much time just staring at the codes. I pretended to be in some deep trance code-thinking thing. Then, as the clock ticked by, I did the hand-clap thing that Stanley did in Swordfish. I wrote a few lines of code - a utility CSS class to center align a block element, preferably a DIV.

I walked out with people clamoring for my autograph and selfies with their Nokia N95. Later that evening, I sent an invoice that was ten times the monthly salary of the project lead working on that project.

/s

My interpretation of this is he feels that AI-assisted programming is okay if you still participate in the programming sometimes, but if you only prompt then you're not really a programmer. Or something.

He does say that you should provide more context rather than just repeating yourself.

There is a distinction as far as having some engineering skills and knowledge to give the AI hints, but it might be more of a symbolic one than people realize. And as AI continues to improve, that will be more so.

But it does bring up a valid point that if we don't do any problem solving or thinking for ourselves, we will be totally helpless.

I think this is something that all students should be taught pro-actively. Actually there should have already been a national standard reinforcing to kids what happens to your brain if you don't use it and how important it is to actually read, write and solve problems sometimes rather than feeding everything into an AI.

On the other hand, the robustness and general capabilities of ML models continues to improve. We have models that can roughly generate a video game frame-by-frame from a text prompt and input image. How long until we get this for application software? Or a diffusion LLM that just outputs a full WASM source/binary or update from a text prompt or series of images in 4 steps and 5 seconds.

I think it could be a really interesting experiment to create a diffusion LLM trained on thousands and thousands of 6502 games and programs, with the full manual (including images) as input and the output being assembly or machine code along with a bit of metadata for RetroArch or something. Or maybe also include recorded gameplay data/visuals.

I think that a big issue here is the chat UX implementations in most IDEs. In contrast, in zed, in the text-based chat threads, one can directly edit an LLM's answer, making actually fixing bugs oneself much less complicated in terms of UX. In a classical chat interface, where you cannot edit an LLM's answer, you have to either continue the thread, start a new thread with a different prompt asking it to avoid the issue, or fix the code and provide it as context again. If you do it in a new chat, you lose the context of the previous chat (you need to provided it again). If you continue the chat with the added code, you have dirty context (which makes llms perform worse). Moreover, they most probably require you to get in and out of the chat to do things, which is not ideal UX compared to just prompting again. The fact that I can just change the generated code in the chat itself, run it in the repl, and continue from there is the best UX imo for when such problems occur. I would like if more IDEs actually accepted something similar.

Of course, my workflow is very repl based and I assume different experiences exist for other contexts. Moreover, agent workflows have probably different pros and cons. It is great timewise that they can test code and fix bugs themselves, but essentially that's a bit like automating this prompt-reprompt loop. It saves time compared to that (at the cost of a lot more $$$ in tokens overall), but if the agent fails to solve the issues, yet again we are at square zero. What helps a lot is that work is done inside the codebase at least, which is supposed to be a good thing, but personally I have not understood how to use agents efficiently tbh.