“There’s no way to get there without a breakthrough,” OpenAI CEO Sam Altman said, arguing that AI will soon need even more energy.
Optimizing power consumption? Why?!
In fairness the computing world has seen unfathomable efficiency gains that are being pushed further with the sudden adoption of arm. We are doing our damnedest to make computers faster and more efficient, and we’re doing a really good job of it, but energy production hasn’t seen nearly those gains in the same amount of time. With the sudden widespread adoption of AI, a very power hungry tool (because it’s basically emulating a brain in a computer), it has caused a sudden spike in energy needed for computers that are already getting more efficient as fast as we can. Meanwhile energy production isn’t keeping up at the same rate of innovation.
The problem there is the paradox of efficiency, making something more efficient ends up using more of it not less as the increase in use stimulated by the greater efficiency outweighs the reduced input used.
It’s not so much the hardware as it is the software and utilisation, and by software I don’t necessarily mean any specific algorithm, because I know they give much thought to optimisation strategies when it comes to implementation and design of machine learning architectures. What I mean by software is the full stack considered as a whole, and by utilisation I mean the way services advertise and make use of ill-suited architectures.
The full stack consists of general purpose computing devices with an unreasonable number of layers of abstraction between the hardware and the languages used in implementations of machine learning. A lot of this stuff is written in Python! While algorithmic complexity is naturally a major factor, how it is compiled and executed matters a lot, too.
Once AI implementations stabilise, the theoretically most energy efficient way to run it would be on custom hardware made to only run that code, and that code would be written in the lowest possible level of abstraction. The closer we get to the metal (or the closer the metal gets to our program), the more efficient we can make it go. I don’t think we take bespoke hardware seriously enough; we’re stuck in this mindset of everything being general-purpose.
As for utilisation: LLMs are not fit or even capable of dealing with logical problems or anything involving reasoning based on knowledge; they can’t even reliably regurgitate knowledge. Yet, as far as I can tell, this constitutes a significant portion of its current use.
If the usage of LLMs was reserved for solving linguistic problems, then we wouldn’t be wasting so much energy generating text and expecting it to contain wisdom. A language model should serve as a surface layer – an interface – on top of bespoke tools, including other domain-specific types of models. I know we’re seeing this idea being iterated on, but I don’t see this being pushed nearly enough.[1]
When it comes to image generation models, I think it’s wrong to focus on generating derivative art/remixes of existing works instead of on tools to help artists express themselves. All these image generation sites we have now consume so much power just so that artistically wanting people can generate 20 versions (give or take an order of magnitude) of the same generic thing. I would like to see AI technology made specifically for integration into professional workflows and tools, enabling creative people to enhance and iterate on their work through specific instructions.[2] The AI we have now are made for people who can’t tell (or don’t care about) the difference between remixing and creating and just want to tell the computer to make something nice so they can use it to sell their products.
The end result in all these cases is that fewer people can live off of being creative and/or knowledgeable while energy consumption spikes as computers generate shitty substitutes. After all, capitalism is all about efficient allocation of resources. Just so happens that quality (of life; art; anything) is inefficient and exploiting the planet is cheap.
For example, why does OpenAI gate external tool integration behind a payment plan while offering simple text generation for free? That just encourages people to rely on text generation for all kinds of tasks it’s not suitable for. Other examples include companies offering AI “assistants” or even AI “teachers”(!), all of which are incapable of even remembering the topic being discussed 2 minutes into a conversation. ↩︎
I get incredibly frustrated when I try to use image generation tools because I go into it with a vision, but since the models are incapable of creating anything new based on actual concepts I only ever end up with something incredibly artistically compromised and derivative. I can generate hundreds of images based on various contortions of the same prompt, reference image, masking, etc and still not get what I want. THAT is inefficient use of resources, and it’s all because the tools are just not made to help me do art. ↩︎
It’s emulating a ridiculously simplified brain. Real brains have orders of magnitude more neurons, but beyond that they already have completely asynchronous evaluation of those neurons, as well as much more complicated connecting structure, as well as multiple methods of communicating with other neurons, some of which are incredibly subtle and hard to detect.
To really take AI to the next level I think you’d need a completely bespoke processor that can replicate those attributes in hardware, but it would be a very expensive gamble because you’d have no idea if it would work until you built it.
Some of the smartest people on the planet are working to make this profitable. It’s fucking hard.
You are dense and haven’t taking even a look at simple shit like hugging face. Power consumption is about the biggest topic you find with anyone in the know.
The human brain uses about 20W. Maybe AI needs to be more efficient instead?
Perfect let’s use human brains as CPUs then. Not the whole brain just the unused bits.
It’s what matrix would’ve been if the studios didn’t think people would too dumb to get it, so we ended with the nonsense about batteries.
I would love it (if there exists a FOSS variant of that) imagine being able to run a LLM, or even LAM in your head,
wait…
🤔
That would require a revolutionary discovery in material science and hardware.
And yet we have brains. This brute force approach to machine learning is quite effective but has problems scaling. So, new energy sources or new thinking?
If only we could convert empty hype into energy.
So AI can’t exist without stealing people’s content and it can’t exist without using too much energy. Why does it exist then?
Because the shareholders need more growth. They might create Ultron along the way, but think of the profits, man!
There’s no way these chatbots are capable of evolving into Ultron. That’s like saying a toaster is capable of nuclear fusion.
It’s the further research being done on top of the breakthrough tech enabling the chat bots applications people are worried about. It’s basically big tech’s mission now to build Ultron, and they aren’t slowing down.
I think we’ve got a bit before we have to worry about another major jump in AI and way longer for an Ultron. The ones we have now are effectively parsers for google or other existing data. I personally still don’t see how we feel like we can get away with calling that AI.
Any AI that actually creates something ‘new’ that I’ve seen still requires a tremendous amount of oversight, tweaking and guidance to produce useful results. To me, they still feel like very fancy search engines.
So AI can’t exist without stealing people’s content
Using the word “steal” in a way that implies misconduct here is “You wouldn’t download a car” level reasoning. It’s not stealing to use the work of some other artist to inform your own work. If you copy it precisely then it’s plagiarism or infringement, but if you take the style of another artist and learn to use it yourself, that’s…exactly how art has advanced over the course of human history. “Great artists steal,” said Picasso famously.
Training your model on pirated copies, that’s shady. But training your model on purchased or freely available content that’s out there for anyone else to learn from? That’s…just how learning works.
Obviously there are differences, in that generative AI is not actually doing structured “thinking” about the creation of a work. That is, of course, the job of the human writing and tweaking the prompts. But training an AI to be able to write like someone else or paint like someone else isn’t theft unless the AI is, without HEAVY manipulation, spitting out copies that infringe on the intellectual property of the original author/artist/musician.
Generative AI, in its current form, is nothing more than a tool. And you can use any tool nefariously, but that doesn’t mean the tool is inherently nefarious. You can use Microsoft Word to copy Eat, Pray, Love but Elizabeth Gilbert shouldn’t sue Microsoft, she should sue you.
Edit: fixed a typo
The models get more efficient and smaller very fast if you look just a year back. I bet we’ll run some small LLMs locally on our phones (I don’t really believe in the other form factors yet) sooner as we believe. I’d say prior 2030.
I can already locally host a pretty decent ai chatbot on my old M1 Macbook (llama v2 7B) which writes at the same speed I can read, its probably already possible with the top of the line phones.
Lol, “old M1 laptop” 3 to 4 years is not old, damn!
(I have running macbookpro5,3 (mid 2009) on Arch, lol)
But nice to hear that M1 (an thus theoretically even the iPad, if you are not talking about M1 pro / M1 max) can already run llamma v2 7B.
Have you tried the mistralAI already, should be a bit more powerful and a bit more efficient iirc. And it is Apache 2.0 licensed.
Because it’s a miracle technology. Both of those things are also engineering problems - ones that have been massively mitigated already. You can run models almost as good as gpt3.5 on a phone, and individuals are pushing the limits on how efficiently we can train every week
It’s not just making a chatbot or a new tool for art - it’s also protein folding, coming up with unexpected materials, and being another pair of eyes that will assist a person do anything.
They literally promise the fountain of youth, autonomous robots, better materials, better batteries, better everything. It’s a path for our species to break our limits, and become more.
The downside is we don’t know how to handle it. We’re making a mess of it, but it’s not like we could stop… The AI alignment problem is dwarfed by the corporation alignment problem
How about an efficiency breakthrough instead? Our brains just need a meal and can recognize a face without looking at billions of others first.
I mean, we can only do that because our system was trained for hundreds of thousands, millions of years into being able to recognise others of same species
Almost all of our training was done without requiring burning fossil fuels. So maybe ole Sammy can put the brakes on his shit until it’s as fuel efficient as a human brain.
While that is true, a lot of death and suffering was required for us to reach this point as a species. Machines don’t need the wars and natural selection required to achieve the same feats, and don’t have our same limitations.
Erm.
I recall a study about kids under a specific age that cannot get scared of looking at pictures of demons and other horror stuff because they don’t know yet what your everyday default person looks like.
So I’d argue that even people need to get accustomed to a thing before they could recognise or have an opinion about anything.
“recognize a face”
Who’s? Can the human brain just know what someone looks like without prior experience?
Your ability to do anything is based on decades of “data sets” that you’re being constantly fed, it’s no different than an AI they just get it all at once and we have to learn by individual experience.
