I already have stable diffusion on a local machine. I was trying to find motivation to install a LLM locally. You answered my question in a different response
use cases where customization helps while quality does matter much due to scale, i.e spam, then LLMs and related tools are amazing.
I want to work on my stuff in peace and in private without worrying about a company grabbing my stuff and using it for themselves and to give/sell it to other outfits, including the government. “If you have nothing to hide…” is bullshit and needs to die.
You seem unnecessarily hostile about this. If you don’t like LLM just move on.
This is exactly why this sub about technology is better off without business news. You’re just reacting to something you hate and directing that at others.
FWIW I did try a lot (LLMs, code, generative AI for images, 3D models) in a lot of ways (CLI, Web based, chat bot) both locally and using APIs.
I don’t use any on a daily basis. I find it exciting that we can theoretically do a lot “more” automatically but… so far the results have not been worth the efforts. Sadly some of the best use cases are exactly what you highlighted, i.e low effort engagement for spam. Overall I find that either working with a professional (script writer, 3D modeler, dev, designer, etc) is a lot more rewarding but also more efficient which itself makes it cheaper.
For use cases where customization helps while quality does matter much due to scale, i.e spam, then LLMs and related tools are amazing.
PS: I’d love to hear the opinion of a spammer actually, maybe they also think it’s not that efficient either.
I have personally found generative-text LLMs quite good for creating titles. As an example, I have a few hundred tweets that I’m trying to put into a file, and I’ll use an LLM to create a human-readable name for them. It’s much better than a lot of the other summarisation mechanisms (like BERT) I’ve tried with it, but it’s still not perfect, because the model tends to output the same thing in slightly different words each time, so repeat runs will often result in the same thing with a different title.
Here’s an idea, maybe start with curiosity about how someone is getting value out of it? It’s possible you don’t know everything about other people’s experiences.
Exactly. I see AI as a tool to automate the boring parts, if you try to automate the hard parts, you’re going to have a bad time.
Take the time to learn the tools you use thoroughly, and then you can turn to AI to make your use of those tools more efficient. If I’m learning woodworking, for example, I’m going to learn to use hand tools first before using power tools, but there’s no way I’m sticking to hand tools when producing a lot of things. Programming isn’t any different, I’ll learn the language and its idioms as deeply as I can, and only then will I turn to things like AI to spit out boilerplate to work from.
build something - for Go, this was a website, and for Rust it was a Tauri app (basically a website); it should be substantial enough to exercise the things I would normally do with the language, but not so big that I won’t finish
read through substantial portions of the standard library - if this is minimal (e.g. in Rust), read through some high profile projects
repeat 2 & 3 until I feel confident I understand the idioms of the language
I generally avoid setting up editor tooling until I’ve at least run through step 3, because things like code completion can distract from the learning process IMO.
Some books I’ve really enjoyed (i.e. where 1 doesn’t exist):
The C Programming Language - by Brian Kernighan and Dennis Richie
Programming in Lua - by Roberto Ierusalimschy
Learn You a Haskell for Great Good - by Miran Lipovača (available free online)
But regardless of the form it takes, I appreciate a really thorough introduction to the language, followed by some experimentation, and then topped off with some solid, practical code examples. I generally allow myself about 2 weeks before expecting to write anything resembling production code.
These days, I feel confident in a dozen or so programming languages (I really like learning new languages), and I find that thoroughly learning each has made me a better programmer.
Thanks for that, was quite interesting and I agree that completion too early (even… in general) can be distracting.
I did mean about AI though, how you manage to integrate it in your workflow to “automate the boring parts” as I’m curious which parts are “boring” for you and which tools you actual use, and how, to solve the problem. How in particular you are able to estimate if it can be automated with AI, how long it might take, how often you are correct about that bet, how you store and possibly share past attempts to automate, etc.
I honestly don’t use it much, but so far, the most productive uses are:
generate some common structure/algorithm - web app, CLI program, recursive function, etc
search documentation - I may not know what the function/type is, but I can describe it
generate documentation - list arguments, return types, etc
But honestly, the time I save there honestly isn’t worth fighting with the AI most of the time, so I’ll only do it if I’m starting up a big greenfield project and need something up and going quickly. That said, there are some things I refuse to use AI for:
testing - AI may be able to get high coverage, but I don’t think it can produce high quality tests
business logic - the devil is in the details, and I don’t trust AI with details
producing documentation - developers hate writing documentation, which is precisely why devs should be the ones to do it; if AI could do it, other devs could just use AI to generate it, but good docs will do far more than what AI can intuit
At the same time, the trouble with local LLMs is that they’re very resource heavy. Your average household computer isn’t going to be able to run one with much usability or speed.
Which, you know, is fine. Maybe if people had an idea of how much power is required to run them, they would think twice before using a gigawatt to output a poem about farts, and perhaps even wonder how OpenAI can offer that for free.
Btw, a 7b model should run ok on any PC with at least 16GB of RAM and a modern processor/GPU.
it’s a lot slower that chatgpt but on my integrated graphics i7 laptop it ran decent, def enough to be useable. Also there’s different models to play around with, some are faster but worse and some are smarter but slower
reminder, there are localy ran LLMs. Right now is a vital time for open source to fight against closed source in the AI arms race.
https://www.nomic.ai/gpt4all
Another good resource to help people find models https://llm.extractum.io
Or just straight up install https://ollama.com
I like Ollama, and recommend it to tinker, but I admit this “LLM Explorer” is quite neat thanks to sections like “LLMs Fit 16GB VRAM”
Ollama just works but it doesn’t help to pick which model best fits your needs.
What is the need I have to put the effort in to install all this locally. Websites win in terms of convenience.
I don’t think I understand your point, are you saying there is no benefit in running locally and that Websites or APIs are more convenient?
I already have stable diffusion on a local machine. I was trying to find motivation to install a LLM locally. You answered my question in a different response
I want to work on my stuff in peace and in private without worrying about a company grabbing my stuff and using it for themselves and to give/sell it to other outfits, including the government. “If you have nothing to hide…” is bullshit and needs to die.
Good point. Everything you feed into chatgpt is stored for future reference.
Removed by mod
You seem unnecessarily hostile about this. If you don’t like LLM just move on.
This is exactly why this sub about technology is better off without business news. You’re just reacting to something you hate and directing that at others.
Removed by mod
FWIW I did try a lot (LLMs, code, generative AI for images, 3D models) in a lot of ways (CLI, Web based, chat bot) both locally and using APIs.
I don’t use any on a daily basis. I find it exciting that we can theoretically do a lot “more” automatically but… so far the results have not been worth the efforts. Sadly some of the best use cases are exactly what you highlighted, i.e low effort engagement for spam. Overall I find that either working with a professional (script writer, 3D modeler, dev, designer, etc) is a lot more rewarding but also more efficient which itself makes it cheaper.
For use cases where customization helps while quality does matter much due to scale, i.e spam, then LLMs and related tools are amazing.
PS: I’d love to hear the opinion of a spammer actually, maybe they also think it’s not that efficient either.
I have personally found generative-text LLMs quite good for creating titles. As an example, I have a few hundred tweets that I’m trying to put into a file, and I’ll use an LLM to create a human-readable name for them. It’s much better than a lot of the other summarisation mechanisms (like BERT) I’ve tried with it, but it’s still not perfect, because the model tends to output the same thing in slightly different words each time, so repeat runs will often result in the same thing with a different title.
But, that is also a fairly limited use case.
It definitely improves my experience coding in unfamiliar languages. So there’s your counter example.
Removed by mod
Here’s an idea, maybe start with curiosity about how someone is getting value out of it? It’s possible you don’t know everything about other people’s experiences.
Removed by mod
Alan Perlis said “A programming language that doesn’t change the way you think is not worth learning.”
So… if you code in another language without actually “getting it”, solely having a usable result, what is actually the point of changing languages?
I have a job to do. And I understand the other language conceptually, I am just rusty on the syntax.
Also the chat feature is invaluable. I can highlight a piece of code and ask what it does, and copilot explains it.
Exactly. I see AI as a tool to automate the boring parts, if you try to automate the hard parts, you’re going to have a bad time.
Take the time to learn the tools you use thoroughly, and then you can turn to AI to make your use of those tools more efficient. If I’m learning woodworking, for example, I’m going to learn to use hand tools first before using power tools, but there’s no way I’m sticking to hand tools when producing a lot of things. Programming isn’t any different, I’ll learn the language and its idioms as deeply as I can, and only then will I turn to things like AI to spit out boilerplate to work from.
Mind explaining a bit your workflow at the moment?
I’m not sure how to succinctly do that.
When I learn a new language, I:
I generally avoid setting up editor tooling until I’ve at least run through step 3, because things like code completion can distract from the learning process IMO.
Some books I’ve really enjoyed (i.e. where 1 doesn’t exist):
But regardless of the form it takes, I appreciate a really thorough introduction to the language, followed by some experimentation, and then topped off with some solid, practical code examples. I generally allow myself about 2 weeks before expecting to write anything resembling production code.
These days, I feel confident in a dozen or so programming languages (I really like learning new languages), and I find that thoroughly learning each has made me a better programmer.
Thanks for that, was quite interesting and I agree that completion too early (even… in general) can be distracting.
I did mean about AI though, how you manage to integrate it in your workflow to “automate the boring parts” as I’m curious which parts are “boring” for you and which tools you actual use, and how, to solve the problem. How in particular you are able to estimate if it can be automated with AI, how long it might take, how often you are correct about that bet, how you store and possibly share past attempts to automate, etc.
I honestly don’t use it much, but so far, the most productive uses are:
But honestly, the time I save there honestly isn’t worth fighting with the AI most of the time, so I’ll only do it if I’m starting up a big greenfield project and need something up and going quickly. That said, there are some things I refuse to use AI for:
At the same time, the trouble with local LLMs is that they’re very resource heavy. Your average household computer isn’t going to be able to run one with much usability or speed.
Which, you know, is fine. Maybe if people had an idea of how much power is required to run them, they would think twice before using a gigawatt to output a poem about farts, and perhaps even wonder how OpenAI can offer that for free. Btw, a 7b model should run ok on any PC with at least 16GB of RAM and a modern processor/GPU.
Phi 3 can run on pretty low specs (requires 4gb RAM) and has relatively good output
it’s a lot slower that chatgpt but on my integrated graphics i7 laptop it ran decent, def enough to be useable. Also there’s different models to play around with, some are faster but worse and some are smarter but slower