Raw LLM Responses

I'm not convinced. The 'AI' as it stands is a decent enough tool for someone skilled to assist them, but using it to replace workers will likely be less fruitful than expected. Recent studies and videos suggest that only 5% of the AI systems implemented to replace workers (Phone, drive through, scheduling, note taking, etc) is effective and the rest of the humans spend a lot more time checking to make sure it isn't making mistakes, or making stuff up whole-cloth. It's a lot like sending all your tech phone support to india, and then people can't understand them, time differences, cultural differences, jargon differences, etc just bites them in the ass and they have to spend just as much time/effort to clean up what failed in the first place. Then we are talking technology, do you have any idea of how much processing power is needed for these neural networks? For a halfway decent 70B i require a tonne of ram and a tonne of GPU processing to get decent speeds, and by 'decent' I'm talking 2-4 tokens a second. Sure i suppose you can offload those to datacenters, unless of course you lose connection for a few moments and your robot cooking your food suddenly makes a fire or you crash your car into another car, or you all the manless ride-sharing cars suddenly stop because the internet is congested. This means you should really be running the systems locally, but that's on par with probably a $20k computer everywhere you want an AI... Unless you just enjoy living life risky. Now let's look the data centers running GPT, Chat GPT4 is something like 450B models, (with B assume that's Billion so you have 450Gigs for a single model). Sure you can quanitize it, assuming 4bit is good enough rather than 8 or 16bit, that's still 225Gb Ram/Vram. And running on raw ram vs Vram isn't worth it. Likely the later models they are going to push are in the Trillions of nodes... Bump that computer to $50k... Add to the riskiness, power. Have a system that cooks fries and hamburgers, great it's probably plugged in. Wireless? Damn are you really going to run a 2k watt machine on a battery? I mean you can technically, but expect it to be a VERY large battery and computer, or very short periods of time using it and lots of time recharging it. (Not to mention the iffy nature of batteries at present with spontaneous combustibility, i suppose you could use generators, say goodbye to 'green energy'). Personally i think Neurual Networks is a terrible idea. It's throwing crap at the wall and some stuff sticks. Again as a tool it has it's uses, cleaning up audio, upscaling images, etc etc. But it's not intelligent, it doesn't understand nuance that well, and you can't debug it, you literally don't know what it's doing. Blindly installing and using these systems is bad and stupid. Also if you remove people from the work force, people won't get money and the companies can't get money anymore because people don't have the money. This is not rocket science. They'd probably rush headlong into it, and then within a year die suddenly because all the sales dry up. I believe Ford was asked why he was paying his assembly line workers so much, he answered 'I want my workers to be my customers too'. Lastly, people need something to do. The IQ based on the bell curve says at least 30-50% of people need simple/grunt jobs. Building things, doing roads, planting, harvesting, etc. Doing something like UBI isn't a good solution, better to just have jobs for them to do. TLDR: So in short... Maybe they will replace SOME workers but you need a lot of trust, a lot of hardware, and the models need to get a LOT better. You can't take 10-20% failure on everything and then clean it up afterwards hoping to see what is real/fake/mistakes. In 50 years, who knows. Right now? Don't trust it blindly.