Hacker News Reader

Comment by roarcher | Hacker News Reader

I recently used Claude to help me understand a math-dense research paper. It was useful for answering general questions about the structure of the algorithm, where to find information in the paper, and gain a high level/intuitive understanding of how the algorithm worked. It was absolutely abysmal at implementing the code, and would regularly make things up when I probed it about subtleties in the math.

Overall, it sped up my learning greatly, but I had to verify everything it said and its code was a mess. It's a useful tool when used appropriately but it's not threatening my job anytime soon.

View on HN

I mostly used it for exploratory analysis of tools/libraries I am not familiar with, where it points me the parts of the API docs I need to lookup to verify the output. Then I manually adapt or rewrite it.

So in the end I use it fairly often when setting up new things (new infra, new files, new tools, new functions, etc). Although the time it saves is not coding time, but googling/boilerplating time. But in practice I work in a well established project where I rarely do this kind of thing (I don't think I even created a new file in the project last week).

If I am already familiar with the tool/library I almost always skip it (occasionally autocomplete is useful, but I could easily live without it). Occasionally I used for small self-contained snippets of code (usually no more than a single function). Last one I remember was some date formatting code.

Some project do have nice documentation that is a pleasure to use. Laravel's documentation, Postgres'one,... Even if I can get an answer faster with google search and now LLMs, I just open the main documentation site and spend an extra 5 minutes looking around.

I mean it more points me about stuff I don't know that I don't know.

For example, I use it quite a lot when setting up new terraform configuration on resources I am not super familiar with, it often points me out to resources and options I didn't know existed. Of course I always look up everything it spits out (it pretty much never outputs something that actually runs anyway and even when it does it doesn't work out of the bat).

But once the thing is set up, it is almost useless to use it to make small changes.

Maybe I lack a sense of rush, but I usually take the slow lane for these kind of work. Actually reading through the whole documentation and taking notes. Maybe do some experiment. I'm wary of things going wrong and me not being able to give a clear and short explanation on the cause. If something is important for me to do, I may as well do it well (or explain the tradeoffs that makes it a hacky solution instead).

It can be hard to explain to people that slowing down to make sure something is done right, can make the overall project go much faster.

I was on reddit/math and someone was talking about how they were trying to get AI to create there final year thesis from uni. They tried 4-5 different AI solutions and reckoned they got 40-60% of the work from AI but it couldnt conclude the project. A few others on the thread had said the same

Doesnt mean it wont get there - just that it isnt there yet

> I had to verify everything it said

Letting it condense something like a paper and checking it afterwards might be a good learning exercise.

Reply on news.ycombinator.com