I'm glad somebody out there is brave enough to push back against the "personal ChatGPT usage is terrible for the environment" message https://andymasley.substack.com/p/a-cheat-sheet-for-conversations-about
"If you want to prompt ChatGPT 40 times, you can just stop your shower 1 second early."
"If I choose not to take a flight to Europe, I save 3,500,000 ChatGPT searches. this is like stopping more than 7 people from searching ChatGPT for their entire lives."
@simon Using it promotes AI as a viable business model as a whole as well as (in the case of non-local models) provides usage data and training content (and money if you pay) for the companies to do further bad AI business with.
So of course, while the direct impact of one query on the environment doesn't really matter, the economical impact of you using it does indeed have an influence on the environment!
The goal has to be to make AI and its connected surveillance capitalism a non-viable business model, and refraining to use it and shunning everyone who does is a way to do that!
@simon
Give me an approach which:
and maybe then we can talk about actually deploying this in the real world. Until then they should stop their venture-capital induced hype bubble fueled by ex-crypto-bros und get back the the drawing board. I'll take human work over an AI every day until that happens...
@the_moep we have some of those today:
- Does not waste massive amounts of energy (and water) for training: DeepSeek v3, still one of the strongest models - trained for just $6m, way less than the biggest USA models
- Does not require expensive rare materials to run: can't help with that if it rules out laptops
- Doesn't spit out wrong crap: yeah that's still unsolved! Newer models are "better", but honestly an AI tool that never makes a mistake feels like it will always be science fiction me
- Can actually run locally in a private environment: yes! We have that now. The models I can run on my laptop got really good starting from about six months ago
- Is free (as in libre) to be used by everyone: we are there too. The Qwen models are under an Apache 2 open source license. Plenty of other good models are "open weights" which is almost good enough to allow "free to be used by everyone"
@simon
Pretty much all of that does not hold true in my opinion:
DeepSeek v3, still one of the strongest models - trained for just $6m, way less than the biggest USA models
That's money, not power or resources. Any monetary cost claimed by a Chinese company can't be compared to actual free countries as a way larger part of that cost is offset by reduced environmental and worker protection (and partial slavery), so totalitarian in general. Also they allegedly based their work on OpenAI so you might have to add their costs on top too...
Does not require expensive rare materials to run: can't help with that if it rules out laptops
It's about their requirement of specialized hardware to train (while the models might run on "normal" CPU nowadays, they cannot be trained on a cheap phone or laptop. A normal program can be created there no problem.
Can actually run locally in a private environment: yes! We have that now. The models I can run on my laptop got really good starting from about six months ago
They really aren't. They are slow and can't handle actual reasoning or even remembering things like the "big" models can. (And of course they are anything but intelligent.)
But it's all just a word prediction system anyways I guess it now just predicts more words at a basically cost linear to how much words you input and want to have predicted, so with the current approach a local machine will always be behind what one can do in a huge data-center hence why I want a different approach that aren't just llm and inference-based.
Is free (as in libre) to be used by everyone: we are there too. The Qwen models are under an Apache 2 open source license. Plenty of other good models are "open weights" which is almost good enough to allow "free to be used by everyone"
All they release is the finished model (and in the case of Qwen their weights) which of course is nice, but it does not allow reproducing or even forking their work. As long as they do not release the code which made the model and the training data under a free license it imo. cannot be considered free.
Them licensing it under any kind of free software license might actually not be valid as it's not based on work that was available under a free license. I would even go as far and say that most models are released illegally as they are derivatives of copyrighted works. Them sticking a free software license on it does not magically safe them from the copyright the material they used is under.
A good comparison in the classical software development work is the CraftBukkit project which used Mojang's Copyrighted Minecraft code and got taken down not by Mojang/Microsoft but by a contributor because their approach violated the GPLv3 license, most "open" models run into the same issue
@the_moep shunning people who use LLMs because they are being environmentally irresponsible feels dishonest to me
There are plenty of credible arguments against irresponsible usage of LLMs, I don't like seeing people waste their time on the ones that are least credible
@simon Well while I agree that one should use the correct arguments for criticizing something I also don't think one should completely discard the personal environmental-impact argument when it comes to usage of AI.
I see the issue with that argument more in that the real cost of AI-queries are hidden from the user (and independent researchers), it's not the energy spent on the single query but the amount of resources spent to create the model and to further the development of the current approach. (Which includes discarding existing hardware or potentially even models)
Of course that is way harder to calculate but I believe that you would get closer to other things that are now generally accepted as being bad for the environment like short-range flights.
@tasket You are of course right, sorry. It seems that I too fell victim to all the "AI" hype talk, but I feel the actual meaning of that term is now (unfortunately) generally accepted as meaning "generative AI, especially LLMs".
I do not have such a big issue with actual useful "AI" doing stuff like voice recognition or protein folding.
Simon Willison
Max Lee
"Wish the sun to stand still.." 🌞