this post was submitted on 18 Sep 2023
495 points (97.0% liked)
Programmer Humor
32443 readers
1249 users here now
Post funny things about programming here! (Or just rant about your favourite programming language.)
Rules:
- Posts must be relevant to programming, programmers, or computer science.
- No NSFW content.
- Jokes must be in good taste. No hate speech, bigotry, etc.
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Any task that can be expressed as mostly translation is a good task to try with an LLM.
And you know what? Stakeholders tend to love LLMs, so have fun with your complicated problems while I build them by using the ancient technique of slapping some boilerplate together and combining it with the new ways of pasting error messages into chatgippity.
I don't think that was the point. The thing is, people replace calculators with that...
Thing is, they only get some results right and hallucinate others. And you're doing billions of matrix multiplications just to calculate 2+1.
Sure. You can go to a construction site with only your one favorite tool. And use it for everything. And it's impressive to open a glass bottle of beer with a hammer and such. But I can guarantee you, you'll be slower digging that hole than the guys using a proper tool like an excavator.
And: you don't solve any fundamental problems if you don't have the data for it. If the information isn't in your data, the network will start guessing and it will be horrible.
That's not a translation problem, so LLMs are terrible for it.
Always use the right tool for the job. If there are a lot of nails to be hammered, you need a guy with a hammer.
Yeah, that was kind of my point. I think the meme picture means people throw it at everything. No matter what. And the next logical thing would be to strip the computer scientist out of the picture. We have Github Copilot now ;) Let AI decide if AI is the proper tool.
Translation, but not categorization. Trying to get reliable, and more importantly, predictibly accurate, metadata from an LLM without serious training is a pain. ML algorithms are far better for this but certainly take more brainpower (in my experience so far).
There are a disproportionately large number of people who get one pretty demo and think LLMs are the solution to everything. Even for translations, I'd be interested to see how accurate the major models are in real world scenarios. We've been struggling hard to find any practical usage of LLMs that doesn't require the user to be able to verify the output themselves.