Ciro Silvano: AI ate my learning curve

Why skipping the hard part is not always a good idea.

I am a DevOps engineer at Syslab.

A junior is not a slow senior. That may be obvious.

An average junior looks at flow of control in the code. A senior looks more at the code in a whole, or in bigger chunks.

CRUM: Computational Representational Understanding of the Mind (Thagard, 2005). So: look at our mind as a computer.

A senior will more quickly spot patterns.

How developers "talk" to AI coding assistants.

• Drifting. The "prompt pinball". You ask an LLM for direction, and blindly follow.
• Shepherding. You roughly have an idea of the goal you want to reach, but the AI keeps wandering away from it, and you constantly have to keep steering it back.
• Exploration. I don't know the answer, but I know how to choose one. You prompt the AI for alternatives, and you choose the best approach and implement it. The AI gives suggestions, and needs to convince you.
• Acceleration. I know the way to my goal, AI drives me there. AI does the actual coding, and you match patterns: check the code, refine it, or accept it.

The last two are preferable, but you need experience for that. How can a newbie gain experience?

Mitigation strategies:

• Skeletons. Learn how to create structure. Let AI generate a skeleton complete with justifications. Then brainstorm together with a senior, and look at the approaches AI proposes.
• Take the wheel. Choose a direction for one part, ask AI to implement that part, then check the response, improve it ot tell AI what to improve.
• Prune context. You can give the entire git repo, or millions of lines of logs, and ask AI to fix the bug. It does work, but what do you learn? Instead, filter this. Take care selecting which part of the code a bug is most likely in, and give that to AI.

Epistemic debt: trading your awareness in exchange for time. You have a starting point of knowledge, say you know 50% of a code base. That is fine. But one vibe coded feature later, you have no extra knowledge, but the code base has grown. So now you know and understand maybe onlyi 40% of the code.

Think about this a while and it will click: if epistemic debt was not a problem, you would have no job.

Fred: don't let AI trick you into going depth first. If it does not find a solution in a direction quickly enough, it should go a level higher and go breadth first.

Armin: there is not one AI. You can let multiple AIs talk with each other. Or multiple models with different prompts and perspectives.