Hacker News Reader

Comment by lazopm | Hacker News Reader

The way I see it you should calibrate the way you work with LLMs based on how confident you are on that specific area, and if it's your responsibility to own/understand it. Here's how it feels for me:

* Learning stage: 0.5x - 1x. I change my system prompt to teacher mode, taking the productivity hit for actually learning the system/tool pays off dividends later. I change my system prompt to "teacher mode" and slowly loosen it as I get more confident.

* Working-knowledge: 2x - 3x. Once I am ramped up enough I feel like I can get a decent productivity boost. Most of the time is spent at the planning stage. This is my mode for areas I don't really own or care about, just need to get work done.

* Mastered: 10x+ I have been doing web front end for 12+ years, I can quickly review plan/implementations and for my initial prompt I already know most of what I want built.

1x == my speed before AI

View on HN

> * Mastered: 10x+

This really needs to be calibrated to the type of work and complexity.

I can actually believe that LLMs would speed up basic web dev work in small, simple codebases 10X for simple requests.

These conversations usually turn into people talking past each other because they’re working on different things. For other less routine and more complex work, expecting a 10X productivity boost is not realistic at all. It doesn’t matter how good you get at writing prompts and reviewing plans. LLMs just don’t solve everything for you in a good way. Some times the true nature of the problem is revealed while implementing it and by deferring everything to an LLM you spend days throwing tokens at the wrong thing. There is a lot of work where the LLM speed up comes from helping you quickly search docs and codebases and double check your code, but handing the entire thing off to an LLM isn’t reasonable. These tasks aren’t going to reach this mythical 10X productivity boost that is genuinely achievable for much simpler work.

> Some times the true nature of the problem is revealed while implementing it and by deferring everything to an LLM you spend days throwing tokens at the wrong thing.

If you look at the code the LLM spat out (and you really, really should!) you will immediately notice that its shape is not what you thought it should be. You might not notice immediately if you're learning, but if you really "mastered" the domain, you will "just" see it. You will also recognize the problem with your assumptions, and immediately (or after some research) correct the prompt.

Looking at the diffs for everything the LLM does slows you down, of course. Not looking - or looking and not recognizing problems, for one reason or another - can be initially faster, but a single pathological case can eat both the time saving and tokens. For domains you truly "mastered", the current models can generate code as if they read your mind (because you can be that precise in the prompt, and quickly), so it's really glaring and very hard to miss when they somehow misread your mind.

It only works at the "mastered" / "unconscious competence" stages, and only in those narrow domains you truly mastered, but it does seem to work. Is it 10x? No idea, but there is a marked change in the speed boost when crossing from conscious to unconscious competence area, with everything else (harness, model) staying the same.

What they are saying is there are domains where there's no such thing as an expert who has the unconsious competence to immediately notice the LLM made glaring mistakes.

Some domains don't have the established regularity of architecture that it takes to achieve 10X efficiency. Two I can think of off the top of my head are embedded systems, with a variety of sensors, outputs, processing power, and memory, and novel protocols like ATProto, for which training data is thin on the ground.

Hey, could you share the prompt you're using for "teacher mode"?

Someone posted their AGENTS.md some time ago, that is used in practice to teach their students[1]. That might be useful to you in this context as well.

[1]: https://news.ycombinator.com/item?id=48359858

Reply on news.ycombinator.com