I feel like current neural network based AI is like half of the solution of the general AI problem.We have the fuzzy (and unreliable, hallucinating) part of the problem solved, but we're missing the strict and verifiable logic part.And no, I'm not arguing for symbolic AI - I know it was mostly a failure, not really going anywhere.What I feel would fit that part is... formal verification.I strongly believe that if we had by now - and by now I mean for at least few decades - a practical, intuitive and commonly used - by commonly used I mean it would not only be used strictly for programming or math, but often also to describe scientific and engineering models - a good formally-verified programming language and all the data available due that use - just like we do with code - due to open source, then we would already have general AI by now.Maybe not super-intelligence still, but a tool that could reliably perform general and verifiably correct work.Current AI is as is, because it lacks the tools to verify its own answers - human may make mistakes but he/she will be verified by reality, AI is shielded from it, it has no senses, so it has to have some other means to judge its output.And honestly we seriously lack in the department of tools that can automatically verify correctness of solutions to all but the most basic of problems - not even due to the lack of theory fundamentals, but due to the lack of effort and investment in the area.
>>16872164It actually has to go a little fuzzier. You need a quantum computer in the loop somehow in order to enable data transfer between the physical world and the Duat, thereby emulating one of the functions of the soul.
none of the problem is solved, drift and hallucination will get worse because they don't know what they are doing, but you don't need quantum anything to figure it outeverything necessary to build a clear backwards reproducible true thought machine is already here, all the ideas and components are here, it's just big corp value asset in infra, that's why they sprawl with hardware, always follow the money, or in this case, meaningful physical asset can easily translated for shareholderseverything is already figured out in academia, there are labs developing ....thought clarifiers and dags, they just don't know how it can all mesh together, from what i see, they won't all magically gather like some kind of gravitation well, they will all go their separate little add onsi don't know about agi, i don't even call it ai, just llm or neural net, but trust me when i say the literature for a non-hallucinating no-drift bicameral mind can be made with current research, you got to believe me, also they have a high wall, if you email them with your idea it goes directly to the trash, you cannot contact them at all, they are building the tower of babel, and it only goes one way
>>16872164>We have the fuzzy (and unreliable, hallucinating) part of the problem solved, but we're missing the strict and verifiable logic part. So, garbage in, garbage out.Thanks Hawking.
>>16872164>human may make mistakes but he/she will be verified by realitySanta Clause isn't real. Grow up.Nobody bothers to verify niche useless research. The scientific method is not operating as advertised.Do you really think academia hasn't been a giant disinformation/redirection effort since the 40s? Nukes changed the game. And anything else dangerous/powerful.Do you think we weren't trying to do exactly what AI is trying to do right now since the 40s?https://en.wikipedia.org/wiki/Emanuel_GoldbergOCR in 1914. Strange similarity to emanual goldstein?https://en.wikipedia.org/wiki/Expert_system#Early_developmentFits my timeline.(And closer to what OP might want)https://en.wikipedia.org/wiki/Neural_network_(machine_learning)#HistoryAllowed (de-fanged) published papers begin in the 60s and 70s which makes me think US and Soviets got there before.My guess is the seemingly recent AI breakthroughs are actually the second (at least) go around in the US.And the data for the recent shit is dirtier than the first go around.Eisenhower warned us in 61https://www.youtube.com/watch?v=OyBNmecVtdU&t=521
>>16872164General intelligence is such an unnecessary idea. If you use specialized current AI and simply tokenize motor inputs/actions to commands you can build up a system for widely capable robots without needing general AI. In fact it's safer, because it's better to have a machine that works in 99.99% of situations and then spazzes out in 0.01% rather than a general system which works all the time but is plotting to kill you.
>>16872247Assuming the strategy was to only allow benign research to be published in academia, what would it look like if somebody published something that was thought to be benign but was later found out to lead to something significant?1) Ignore to avoid Streisand effect?2) Have "experts" swoop in to discredit?3) Use a forest of bullshit to hide the trees?4) Stealth retraction?You might be able to strike gold (or become a muzzled slave to your state) with this guiding heuristic.
>>16872169
>>16872258not possible, academics still read journals, they might be surprised but really good ideas won't escape their notice
>>16872164What are you on about dummy? The problem is memory.
>>16872368kind of but not really
Smash
>>16872354What if the idea is useless by itself and the utility only emerges at scale?Or requires another component that is classified in order to be really useful?Do you really think there is a full proper vetting of EVERYTHING?Most researchers don't want to waste time on bullshit. And there is a publish or perish treadmill to limit low probability pursuits of needles in haystacks.And anyone that does strike gold gets muzzled so you wouldn't hear about it.Go find the paper that tells how to make nukes. If you can't find it then nukes must not be real.
