michael-dean-k/

If you think LLMs suck at coding, you might have miscalibrated expectations. AI coding is really more of a two-way spatial language puzzle than a master-slave relationship. Yelling at it in ALL CAPS will not give you the right answer. Some thoughts:

1. It will always follow your commands, and so before you let it loose on your code, you should ask it to summarize its goal and to ask you follow-up question to clarify implementation.

2. Implementing even a basic feature is often a chain of 5-10 simpler events. The more ambitious the feature, the more likely it is to mess up. It will attempt any scope you give it, from layups to Hail Marys. It’s best to have it understand the larger feature, but to implement in micro steps.

3. AI will make mistakes if the context provided isn’t clear. Do you give it no reference, 1 file, all files, and handful of files? Depending on the situation, the answer is always different.

4. Even as AI gets better at coding, humans will benefit from understanding the larger architecture of their code base.

5. Take responsibility. When it gives you wrong code, it’s not because the LLM sucks, it’s because you haven’t given it the right directions. You need patience, the ability to troubleshoot and problem solve.

Considering it’s quite an active process, why do it? If you’re a code ninja, then you probably shouldn’t. If you’re not completely fluent, it makes sense. There’s a tradeoff, you get quick results, but you also lose the reps to build fluency. Is the ability to quickly write a file of code from scratch w/o AI important?

I think code has two parts to it: syntax and structure. I believe structure is the more valuable thing to learn (ie: how does a full-stack operate?). As LLMs get better, the syntax will be more accurate, enabling more people to build apps, but only if they understand the structure of the app (I think it’s harder for LLMs to grasp structure, especially as codebases get more complex).

What makes AI coding different from AI writing? In writing, the syntax IS the experience, so it’s not something you want to automate.

Now that anyone can code, it’s time or a “revenge of the idea guy.” Plus, AGI backwards is “idea guy alliance.”

The barrier to coding will continue to get lower and lower. But even if you always interact with code through “plan language,” I don’t know if the human can ever totally separate from the abstractions. There is a network of files, functions, and databases. To not understand their inner-workings is to lose a resolution of control and specificity. I imagine that a “visual scripting” system (like in Unreal Engine 4), will help non-coders visualize the architecture of their code.