>>44758 This post increased my awareness of the problem.

We've been well aware of the problem with LLMs on here for some time, it just hasn't been discussed lately since not much has changed on that front. I'm not sure if this topic really warrants its own thread at this time, but I'll leave it up for now and let Chobitstu be the final judge on the matter. >>44759 Don't respond to obvious trolls. Just make a report and Chobitsu, Kiwi, or I will deal with it.

>>44760 :D

>>44756 LLMs may be primitive compared to sci-fi, but so far, it's the best we have to give our waifus personalities that can react to real life events. I guess this a good reminder that robots are still partially sci-fi. >In truth, we are about as close to a true AGI as we were in the 1960s - we are in a modern messiah delusion. That's just wrong. Imagine showing ChatGPT 3 in 2012, 10 years before it publicly released. Before ~2019, AI was still mostly a sci-fi term. I think we've been too accustomed to AI being in our daily lives. It's like how a smartphone would be miraculous to someone in 1980. >So how do we break free from this paradigm? I truly want AGI so I'm honestly mad at how delusional and dishonest we are about these systems, even for highly structured information like code these systems are still at about "competent normie with dysfunctional hippocampus" level. No offense, but my response is basically "do you have anything better?" I personally think AGI is still far away. It's like interplanetary space travel or practical flying cars. >At least we are expanding our compute infra, but our entire species is still insanely compute poor compared to a puddle of shitty water. Do we really need more compute? Seems like we have enough compute already, if not too much. We have data centers everywhere, whether we like it or not. There's computers in home appliances. I read an anecdote about how a guy got logged off his lightbulbs. If anything, we need less compute, and more optimization.

>>44762 LLMs are fancy toys, maybe we moved from 0.001-0.002% of the way to AGI, but my point still stands. Do I have anything better? No. But we are actively sabotaging the effort with retarded transformer scaling that just nets you a slightly more precise amnesic retard. It's fun and kind of interesting but at the same time incredibly frivolous and a massive waste of resources. A single eukaryotic cell beats us profoundly in compute, it's honestly pathetic. I just want the delusion to go down so we can use all these sparkly new datacenters for something actually productive that isn't mass surveillance or slop slot machines.

>>44766 Well, AGI is a nebulous term. Does it mean intelligent in the real world? If so, then you should look into VLA models (>>44521) >But we are actively sabotaging the effort with retarded transformer scaling that just nets you a slightly more precise amnesic retard. It's fun and kind of interesting but at the same time incredibly frivolous and a massive waste of resources. Then what's the better option? You can bemoan the inefficiency and overambitious zeal of the AI companies, but what should we do instead? How will you get my Galatea to talk? >A single eukaryotic cell beats us profoundly in compute, it's honestly pathetic. What? >I just want the delusion to go down so we can use all these sparkly new datacenters for something actually productive that isn't mass surveillance or slop slot machines. That's why they were built in the first place. The only other reason I could see them being used for is scientific things, like SETI(at)home and Folding(at)home. But ((((they)))) wouldn't allow it, that actually helps people

>>44767 If I took a single cell from your body and measured it perfectly I wouldn't be able to find a single compute cluster on earth that could dynamically simulate it. Now multiply that with hundreds of billions and you have a brain, trillions and you have a mammal. We can estimate behavior through statistical measurement and modeling but the fact we're working with digital computers will continue to fuck us up thermodynamically. We can build massive measurement/modeling systems that can autoregressively simulate communication, but it's as brittle as any digital modeling system and works more like a stochastic state machine - if you go to any unexpected state it quickly falls apart, I think most LLM curious people are well aware of these out-of-distribution states, they're very easy to induce. The car wash meme is a good example. The reason cells aren't brittle is because they aren't trying to simulate a cellular system based on measurement and behavioral statistics, they are a cellular system. Analog computing gets us a bit closer, we escape the digital penalties but then we run into conformational penalties - the system can't efficiently physically restructure itself as the residual evolves so the state will distort at some point or be locked into a specific space. Maybe an analog FPGA.. LLMs are interesting technology and they're fun toys, but they are a digital approximation of a discrete analog system, that approximates a continuous morphing analog system, they can only scratch the surface of complexity and dynamics, that's why it's called slop.

>>44768 >If I took a single cell from your body and measured it perfectly I wouldn't be able to find a single compute cluster on earth that could dynamically simulate it. I think you would be able to simulate one cell. Maybe not perfectly, but well enough. But that's a slightly different topic. It is compute related, but that's physics simulation, not personality/intelligence. I don't need Galatea to fold proteins. Most of the movement is probably brownian motion anyway. The rest of what you're saying is absolutely true, but as I alluded to before, there's stuff that is still beyond us currently. I personally wish we had more space infrastructure. The best thing to do is make the best of what we have, inspired by the sci-fi we look to and dream about.

>>44770 If you read Michael Levins research you will see that cells are actually intelligent agents. That's why I picked a cell, that tiny thing can be an agent because of the depth of internal dynamics it can express using a pW of power. We are stuck with sequential calculators and these hideously power inefficient unstable retards. I can't figure out if I should go into EE/Analog ASIC design, SW, nanotech, math, physics or bio. Where is our leverage? It sure as fuck isn't 100GW of Jensen's space heaters.

>>44771 I can't give good career advice, sorry. I would look into which one leads into better careers.

>>44772 I don't care about money, friend. I just want our species to have more power so we can force the world to suck less.

>>44773 >I don't care about money, friend. Me too on principle, but you need money to survive, and you need some if you want to do your own projects. Obviously, don't get greedy or immoral. >I just want our species to have more power so we can force the world to suck less. 90% of the reason the world sucks is my own species. The majority of problems are social, not technological.

>>44774 ADDENDUM This entire board was created because of conflicts with 50% of our entire species

>>44774 The species reflects the amoral system that generated it, intelligence can potentially stop the zero sum game for the majority which would make us even less evil. Predators still exist because evolution fucking loves them, that seems to be fundamentally built into the world, a longer term kind of problem. Not much else we can do.

>>44762 POTD >Do we really need more compute? Seems like we have enough compute already, if not too much. >If anything, we need less compute, and more optimization. Well-spoken, GreerTech! I hardly need remind any regular here the massive datacenter approach is (((by-design))), ofc. >"Plenty for me, but none for thee." as is commonplace with these wretches. >tl;dr Only pursue opensauce projects that are focused entirely on edge computing, such as ggerganov's work * is my advice, Anons. Cheers. --- * Developer of the famed llama.cpp https://github.com/ggerganov

>>44756 Hello, Anon. It's an interesting question and one we've delved into extensively here on /robowaifu/ , as @Greentext anon pointed out. >AGI Don't hold your breath! :) >>44760 >but I'll leave it up for now and let Chobitstu be the final judge on the matter. Heh, I'm inclined to let it stand if neither of you two have objections. Later we can merge into a larger-scoped thread if we need the thread-budget headroom? In the meantime, as long as the thread remains active with discussion I don't see the harm in it.