Oddbean new post about | logout
 I understand that ChatGPT makes things harder for people who are retarded. But if you already know stuff it just makes you irrationally angry when it's obviously wrong.

https://media.gleasonator.dev/ipfs/QmVM7rM57vTXk6pVG1kfufVYVxDnHKBMFvDkmxSsv5NVDX 
 I honestly see no benefit using ChatGPT or any other LLM when coding. Just the fact that I have to "change brain mode" and try and explain something in English is enough for me to lose all the potential "productivity gains". 

Then even if you do that, the code it writes is so bad, it writes bad code with the confidence of a cocky 10 year old who has just learned how to code and thinks they know everything. But it keeps decent ish variable and function names, decent type usage and a consistent style, so some less skilled developers can't tell how horribly bad the code is.

Also, finally, I honestly think the only developers who can even benefit from an LLM in any way where they save time is either beginners or people that just do repetive and iterative things.

The only thing it's sometimes genuinely useful for use an LLM is if you're dealing with a new programming language or some badly documented api and you want a quick way to search for something (and you don't mind if its wrong 20% of the time)  
 As a novice coder, it has been invaluable. I don't have time to read through an entire documentation, just to figure out that the library can't do what I want. I have an unrelated day job. I just want to ask "how do I do X with Y library, or is there a better library I should be using?" 
 The thing it has helped me on most is probably SQL. I think it is the world's foremost leading expert on SQL. 
 Yeah it can be useful for that but don't learn to rely on it, it will honestly just bite you in the ass if you do.  
 It points me in the right direction so I can actually read the *relevant* docs. I am a novice coder, but I am very familiar with hallucinations. I've been running local models for a couple years now. I understand that it's just a next token predictor.

That said, you might find some benefit in allowing a model to autofill boilerplate code for you. I don't program enough for it to be worth my time to set that up (I also use NVChad and there aren't a lot of easy tutorials for that, like there are with something like sublime or vscode)