Hi all, I’m working on a Solarpunk world building project and I want to know your thoughts on one of the main features of my world. To preface all this and provide some context, my world is an alternate-history with a divergence point sometime in the 2020s. The divergence was caused by a vocal and technically-skilled group of Green-Anarchists that labelled themselves as “Dawn”.
Dawn did a whole host of things to ween people off of Capitalism and into my Solarpunk world, I’ve gone into immense detail on this but I doubt it’s relevancy to my question so I’ll omit all those details, but there was a tipping-point in which Capitalism crumbled and gave into Dawn’s Anarcho-Solar world.
To make sure the world stayed Solarpunk and to give people stress-free lifestyles, they developed 1 AI and 1 AGI. The AGI manages all Dawn technology, such as Dawn power generation, carbon-capture, a global hyper-loop etc and the AI makes sure no one tampers with the AGI (For those unaware, AGI is Artificial General Intelligence, so for example Skynet is an AGI since it can think and do many things, but ChatGPT is an AI because it can only do text).
Most people in my world wont ever have to think about the AI and AGI, it is taught in my education system to make people aware in case of catastrophe but it mostly manages itself and is monitored by the longest-serving Dawn members.
I simply want to know if machines like this can exist in Solarpunk with it remaining Solarpunk, and if people like the idea or not. If you want to know more about my world building then feel free to ask! Thanks for your time in advance :)
P.S. I should mention that AI and AGI are mirrored across 8 different instances and for the most part work independently of each other, meeting only when strictly necessary. This is to give even more defense against tampering and error.
Edit 1: Changed title from Overlord to Background, Overlord implies oppression which the system doesn’t do.
We know how to generate MWs out of solar power and some models run on raspberry pis. Our immediately available tech has this. I run a very decent LLM (Mistral-7B) on my machine and it must draw 500W max, that’s extremely doable.
500w is what my PC w/ a 3080 runs at with stable diffusion, with everything including monitors and etc it usually caps at about 700w - while the analog equivalent IIRC was about 80w max under load (I’ll have to double check that though). Powering the peripherals like monitors and speakers draws more than sifting through an LLM with that tech! And that’s a more overarching issue when you scale it up per-home, worse if per individual.
I don’t necessarily think it’s infeasible, it’s more about applications and purposes, like why use a 500w dedicated power supply if the same purpose could be scaled down leaving 420w for other tasks. Similarly, we use cloud clusters for stable diffusion, mid journey, dall-e and such all aren’t running on one computer alone but a bunch of computers connecting to a server to process the loads in queue. So while the PC’s aren’t working on “the same task” they are all working on the same goal. For a community server like a stable diffusion discord this work is spread across volunteer PC’s that are (mostly) already being used for day to day tasks anyway. But even then, say just 5 people volunteer their PC’s that’s now 2,500w all just to speed up a queue. Whereas a specified instrument cut down to 80w ran by 5 people is only 400w.
Basically what I’m getting at is right now almost regardless of what technology we have available, we are brute forcing AI. Using CPU and GPU tasks to accomplish something that can be done with an analog computer in a quarter of the time with 98% accuracy. For former is just not sustainable by comparison.
On one hand, making breakthroughs towards making current technology better and feasible is a good thing. On the other hand, why not use all available options and utilize the benefits of each. Consumers being able to use AI locally is awesome, but consumers using an energy efficient option would be even better.
I came to this conclusion with the solarpunk ethos in mind, and again it’s not that I think it’s infeasible or shouldn’t be done or anything like that, just that we have a bit of tunnel vision with digital computing and it would be interesting to see what could be accomplished with a wider scope. Especially since it is somewhat just a breadboard with copper wires ;) lol
Agreed, there are certainly more optimal ways to do. The applications of LLMs are being made right now, and multimodal robotics systems are coming out of the lab this year. My point is that we should not believe that they pose any kind of energy waste problem if they do a useful function.