The LLMs have been this moment for me - the toolkit, usage patterns, and strengths and weaknesses are so different from what I’m used to that I have a really hard time trusting my instincts or judgements on when to use them and where. They’re the first piece of technology that’s so far outside my experience set that I need to onboard a fully different paradigm to understand them as a tool.

To be fair, currently it looks like most people's "instinct" about LLMs is "let's just blindly trust what this thing says because it sounds authoritative". I'm waiting for someone to coin the term "AI natives", a generation which, it'll eventually turn out, will be just as "native" as Gen Z are "digital natives".

> I have a really hard time trusting my instincts or judgements on when to use them and where

Its easy, do you want to find something but you forgot the name and have a hard time with Google? Ask an LLM, they will likely find the name you need to find it on Google. This goes for both things you have forgotten and new concepts that you never knew the name for but you think there exists something.

The other main use case is translations between languages, LLMs are by far the best translators today.

For other things LLM doesn't really help, except help turn out crap you have to spend so much time checking that it doesn't really make you more productive unless your job is to create crap.

Or that isn't fair, if you need to write a letter with some emotions that you have a hard time with due to your fleshy nature not wanting to fake emotions, you can ask the computer to fake those emotions for you. Computers are great at faking emotions now. Think corporate emails etc where you have to use a tone that doesn't at all match your state etc.

I wouldn't worry too much. I'm not using any LLMs for my work and don't feel that I'm falling behind in any way.

This is exactly what people who are falling behind would say. They're often the last ones who are aware of the situation. (Just food for thought - not saying you are actually falling behind.)

I guess you're correct. Still, I think that LLMs are overhyped and have yet to see any real life productivity gains because of them.

Perhaps it's too obvious, but if you're not using them, you wouldn't see any real life productivity gains, would you?

They have easily improved my coding productivity by 40%

What programming language are you using and in which ways are LLMs helping you? How do you deal with subtle mistakes?

Python, not that it really matters.

We had to develop something internally regarding genlock and a ring buffer that Python wasn't going to work, so an engineer at our company picked up C++. Mind you, he's never written a line of C++ in his life and has close to no experience with C-style languages at all.

Within 2 days he had completed the prototype to synchronize multiple latency sensitive gear with leading/falling edge methods in C++. (On top of his other work.)

3 days prior to it he'd written no C++.

When asked how he did it, he said: "Claude is the best translator out there."

As for subtle mistakes, you still need to know how to code. Compilers make plenty of subtle inaccuracies and we still use them for most applications, don't we?

That's interesting! I wouldn't feel comfortable using a non memory safe programming language without any experience in it though. Claude could introduce memory handling bugs and I might not be able to spot them.

My personal experience has been that the “smarter autocomplete” in the IDE is the bulk of the value add for me. It handles a lot of the really rote stuff - “create a struct that matches this data structure”, “add an accessor for this property”, “make another unit test with a slight variation on this” and saves me an enormous amount of time typing.

I’ve had basically no luck getting code from the chat-style interfaces, though. The bugs often aren’t even subtle, and it’ll either argue with me that there’s not a bug or apologize and produce exactly the same code again.

> saves me an enormous amount of time typing

But typing is like less than 1% of the time spent programming, it doesn't take many seconds to type the things you mentioned and it takes significantly longer to figure out you want to make them.

In the moment, having to type anything more than half a line is still a distraction, as you can still make typos, or start watching as the starts to highlight issues as you type (which may or may not be real issues). Reducing distractions in the moment still can help keep flow going for some people.