AI has a ridiculous resource cost in electricity, and so far the news on productivity gains has been mixed or uncertain. Meanwhile, OpenAI has been partially successful at trading vaporware (stock price increases) for hardware (graphics cards), which is reckless to an absurd degree and an omen of a potential bubble.
There's no question that AI is a qualitative leap in technology. But aside from specific cases like decoding ancient tablets, what exactly is produced by better google searches and faster concept iteration that we didn't have before? And whatever that is, is it worth the cost?
For example: AI is famously just as good as junior software developers, or better, at writing code. But if it costs billions of dollars to accomplish that, is it worth it? What about replacing artists and journalists that are already paid like crap anyway?
AI also has the potential to replace huge swathes of white collar, like personal assistants, secretaries, various types of managers, HR, statisticians, etc. But a lot of those jobs are female welfare in the first place, so if efficiency was the point, they would've been gone already. Does anyone believe the entire girlboss economy is going to be thrown out on its pantsuit?
Also, does anyone know how liability will be handled with AI agents? I'm pretty sure that's one of the significant reasons we don't have autonomous cars yet. Contra the name "Full Self-Driving," Tesla requires you to have hands on the steering wheel because they can offload all responsibility for your death if you take a nap in the back seat and the Tesla headbutts a tree (which has happened). Under a fully autonomous system, do they pay out the lawsuits? What kind of sanctions will come from regulatory bodies?
Somewhat related, an AI that gets things right 99% of the time would seem to be unacceptable for certain applications, like driving or medicine. For example, getting 1% of medical facts wrong, while outperforming a real doctor on average, would inevitably result in catastrophic consequences for a random member of the public and make a human check & balance necessary, which nullifies a lot of the potential efficiency advantage.
I think there is a strong possibility that we'll look at today's landscape as the good old days when virtually unlimited queries were possible for a ridiculously low cost. When AI actually has to demonstrate ROI, there may be a future where each generation costs $5, or worse.
Am I wrong or is it only the training that uses a ton of energy? In any case I think the real future of AI is in entertainment. Procedural generation, character chat, shit like that.
Yup, the training is the real energy suck, and anything used at scale will have to be trained by big tech and suffer as a result. I think you're right. It's a fun toy but I've yet to find any serious use case for anyone with an above average IQ. Helping jeets cheat their way through freshman level programming classes seems to be the extent of its capability. For entertainment the porn models will develop at light speed since big tech left that field wide open which will have mixed results to say the least.
Lol, Ive heard of people "code vibing", but ive never seen any good projects ever result from it. Its better used as a aide in that situation, to write some simple boilerplate code but I cant even get it to do that. But it does pretty well as a teacher and I dont feel to bad for bugging it a thousand times in a session.
But, for simple things, lol why would you pay someone a living wage for example, data entry when a ai can do it much faster and with less errors as well.
It does pretty good at predicting human nature as well so its used heavily in advertising and data collection.
Its not perfect, but its getting better, and faster too. Its not a technology were ready for yet in my opinion.
The real neat trick that the elite want will be putting them into robotic bodies so they can do anything a human could. Of course this would require some tremendous leap in "known" technology. For all we know the military may have things way more capable behind closed doors. In terminator 2, the government is building all these autonomous kill machines, thats exactly the kinda future some of these weirdos envision. Robots "shouldnt" rebel when they dont get enough to eat.
The rich have to much invested in this and they dont care if they burn the world down in the process or wipe out humanity.
Just the risk of becoming a god I guess.
Another good example is the game developer who won a nobel peace prize for his work with ai and drug development.
https://www.theguardian.com/science/2024/oct/09/demis-hassabis-from-video-game-designer-to-nobel-prize-winner-google-deepmind-ai-
In my experience it really depends on the size and quality of your code base. If you have a huge, well written code base it does a lot better than starting from scratch - it references existing code so generally tries to produce a similar style.
I tried "vibe coding" (hate that fucking faggot term) a project from scratch and it mostly produced stuff, that while it technically worked, it was something a jeet would shit out.
I do hope AI replaces a lot of outsourced jobs though.
I use it rarely, but when I do I mostly use it to do shit I don't want to (and that doesn't really matter).
On the otherhand you have NPCs that have started off-loading all their thinking to it and crazy bitches forming emotional bonds with it.
Future? Were living some of that now.
That’s the present of AI!
Inference uses about the same amount of electricity, but it only takes a few seconds (per request, obviously it adds up when you're serving thousamds of queries per second) instead of weeks or months.
I suspect the cost of hardware is the limiting factor for end users since they would be fine tuning at most, and probably not even that. In terms of companies taking a blowtorch to their workforce the cost would be a nonissue since they would expect it to pay for itself.
You can do a surprising amount of AI fine tuning/training at home, especially if you're using it for image generation, on a GPU that costs under $1000 and maxes out at 350 watts.
Interesting. I looked into it on the LLM side of things a while back and they were recommending top of the line graphics cards as a bare minimum, and while my 4070 Ti Super is decent it's not top of the line. I haven't done much image generation. I tried my hand at it and wasn't impressed with the results. I don't know if it was the model or user error but I never went back to it. I've mostly been experimenting with models that get around leftist censorship.
Even if they do a lot of these companies are eying up onsite nuclear power as a way of decreasing running costs.
Hahaha nuclear power is finally accepted and utilized to power endless ai wifus
"Nuclear powered AI waifus" is not a sequence of words I thought I'd ever hear in my life.
And oddjob quickfix solutions. Not sure how to convey a proper example right now, but just think of situations where you've had to spend up to 5 minutes manually copying and pasting lines of text or swathes of files. Renaming multiples files and directories, stuff like that, where it'd also be a tedious hassle to try and write or find a script to automate the process for you.
Or put another way, autohotkey without the extra steps.
And yes, you're 100% correct about training being the main energy drain. Even for smaller consumer models, since most kinds of training (aside from LORA models) require so much computing time/power that it would still take eons to do on a beefy consumer grade computer.