Comment by Scubabear68 | Hacker News Reader

Comment by Scubabear68 | Hacker News Reader

Parent Full thread

Who said I was "looking at walls of code [I] don't understand"? Really, you are reading a lot into this without context.

I've been working software development for close to 40 years now. I've worked in plenty of languages, software stacks, ecosystems, from assembler to VAXen to Unix (pre-dating Linux), to cloud stacks to...well, a lot of systems.

The software ecosystem has ballooned so tremendously in that time frame that no human can really be an expert in all of them, or even in a tiny subset of all of them. As I mentioned in another comment, I couldn't be bothered to keep up with the Javascript/Typescript/Web front end drama. I don't do rollercoasters, and that ride is enough to make even the staunchest stomach heave with distress from the endless loop-de-loops.

With LLMs, I can describe a little problem unique to my problem space, inspect the output, research it and even reverse engineer it, and see how well it fits (or doesn't). This is not unlike being a software engineering manager leading teams.

When I said being able to see the results was "cool", I did not mean or imply it was about me trying to "look cool". In fact, I really don't know where that is coming from. I do this for my day job and also on the side for a lot of different causes and personal projects, and no one could give a rats ass about how "cool" I am. Seriously.

People need to stop taking this all so personally. It's a tool. Maybe you all need to step back to my world in 1981 when I first started playing with code as a kid, and compare it to 2016. The advances from '81 to '16 are probably far more dramatic than what AI is doing now.

We miscommunicated. First, I think LLMs are a problem for newcomers in general, there is a lot of intuition that I saw that you build while tinkering at 3 A.M which makes programming easier in the long run, at least for me. For someone who invested enough time on that, such as you, can easily spot where LLMs fail and already have the knowledge to make the cogs turn. I clarified where I think LLMs are harmful at a sibling comment. This is where I think LLMs obstruct learning. 'Domain knowledge' is an important keystone for agentic coding that bypass this.

As for 'looking cool', I meant the code with its fancy formats and highlighting, no snark intended :)

LLMs are a great tool for building, especially for those who could conceivably be engineer team leads/project managers. For greenfield learning, its sub-par unless its guided well.

I appreciate the comment. I agree that LLMs can be a significant issue with new comers to the field, and their use in replace of foundational learning is causing real concerns for me and my company specifically.

I have found the LLMs work best where I have some knowledge already, and I can spot the gaffes and mistakes more quickly. It has to be rough on people flying them blind, as it were.

Unrelated but I have tried out LLMs on some legal research, and it has been appallingly bad. The LLMs get the gist of a lot of it, but falls down completely because the legal realm requires a certain amount of exactitude. Time and time again I will ask about laws in a certain U.S. State, and it will show me a law from Arizona or something which has no or limited applicability (you can try to cross reference other State laws in certain State courts, but usually that doesn't go very far unless it is a very novel argument).

You hit the nail on the head with the fancy outputs, I am guilty as charged there. The LLMs are frankly better at it then I am.

Reply on news.ycombinator.com