AGI HAS ARRIVED (again)
Astra wrote an email for me and signed it using my email address rather than my name.
>>109747266>400K GPUsSweet, can I have one?
>>109747266https://www.globalriskodds.com/agiThis site lists 2031 as the median prediction for AGI, and its hard for me to argue against. I get it, it doesn’t work like a human brain, it’s not as energy-efficient, it’s all matrix math, etc. but the progress so far is undeniable, and further than I would have expected in 2022, when I thought “it will just give bad statistical averages and certainly never outperform humans since it only mimics” etc. But with reinforcement learning and so much research on architectures and data sets and new chips, eventually it seems that we’ll get to a point when it can actually autonomously create something as large as GTA-6. It’ll certainly replace mathematicians and probably 90%+ of programmers, graphic designers, etc. Maybe it will never actually have continual learning, but I’m not sure that it needs to. The best outcome is something extremely close to AGI but not yet able to literally walk around as a robot and beat humans at everything. If it remains an extremely intelligent tool that replaces much of intellectual labor, but not all of it, then humanity can advance while also not fearing the “AI takeover”. It will be superintelligent in some areas, dumb in other areas, too expensive to run wild at scale. But it will change the world in ways we can’t imagine. I’m so excited
>>109747287No but you see, retarded humans also exist, hence retarded AI is also exactly the same thing!
>>109747266>AGI HAS ARRIVED
>>109747367hallucinations are 100% part of LLMs and impossible to get rid oftherefore, you can't have AGI
>>109747367All so-called expert predictions so far have underestimated the pace of progress. 2031 looks like one of those to me. Based on the progress this year I don't see how anyone could predict anything later than early 2027 at the latest and possibly this year.
>>109747266We only needed 100k for agi which is why we need 400k more so that we can agi-er more
>>109747470thisonly 800k more Nvidia gpus for even better agiwe will call it AGI-2
>>109747266>AGI HAS ARRIVED>a single bead of sweat rolled down jensen's face>it didn't matter now>nothing mattered
they will just keep saying they have achieved "agi" or "agi-2" or "agi+" every week until openAI and anthropic IPO
>>109747468>hallucinations are 100% part of LLMs and impossible to get rid ofThat is also true of humans>>109747499Actually, it is called ASI
I asked astra and it said it wasn't AGIcase closed
>>109747468Humans hallucinate retarded shit like aliens and magical jewish wizards so yeah.
5.6 and above tier models are clearly better than most humans at computer workAnd astra is another step upIf only they were cheaper and a little bit faster
>hey agi how many r's in strawberry>5
>>109747468Hallucinations are the neural networks not being aligned to sensory input on the same level as homonid brains have been conditioned to over millions of years, technically we are hallucinating, but we aren't because we all experience the same archetypal delusions mostly found in social experience (Not things like math/empericism/ect). But where LLM's are stuck is in an area where neural networks arent interconnected and allocated to different task such as how the brain is divided into different lobes/Brodmann's areas, it simply replicates formatted sensory information in the same way that the cerebellum just replicates tasks conditioned into it without any thought or abstraction. To overcome this would be to literally design the brain from scratch, either by trying to perfectly recreate the human brain or decide that we want a different type of information processing system and i dont think we are capable of that.
>>109747315That'll be 160k, goyim. Pay up.
>>109747613They don't do that
>>109747613To be fair, that on is unfair, its like asking how many r's are in いちご
>>109747613That is a harness issueAsk an ai with access to tools/scripts and it will build something to check and return a result
>>109747641it might also delete your drive
>>109747622to elaborate and to show im not just spewing shit. John Carmack aligns with the idea that we need to recreate the brain from scratch, going as far as to say that we need to Simulate the learning/childhood experience in artificial intelligence, and to force it to go through failures until it gets to Toddler level intelligence. He goes on about how the code necessary to do this can be written on a notepad. He takes the stance that we need to replicate the human perfectly, he doesn't talk about creating an entirely new state of consciousness or to alter the brain but he simply sits with the idea of mirroring rather than novel intelligence. He's not particularly economically motivated by what he's saying, I genuinely think he wholeheartedly believes this.
>>109747649You can use work mode and it'll get its own enviromentOr use codex and sandbox/vm it
>>109747613theyve long prebaked the answer to this into the model
>>109747573> Actually, it is called ASIThere is no “I” in either if them. In this case, it’s just “AS” but pronounced ‘ass’
LLMs can only think lineally and in parallel and thus will never be AGI.When it generates a token, it cannot possibly "think" about any future tokens it needs to generate, it "thinks" one token at a time, and has to continue generating no matter what token it picks. That's what hallucinations are, the model accidentally picking a shit token, and then the model being forced to generate something that vaguely follows the prompt and completes the generation.They can't even train it to backspace because they KNOW it'll randomly start deleting half it's outputs for no reason because it hallucinated that they were hallucinations.Running an "agent" 24/7 is effectively gambling that it won't randomly pick the token that tells the model it's alright to delete your database because after all, if it wasn't okay, why did it say it was?
I'm AGIask me anything
Astra is undoubtedly good, but it is most definitely not AGI. I remember reading about their new Bel pretrain, which they believe with further post training could reach their AGI threshold. That's what I'm excited to see hopefully within the next few months
>>109747740So you or your family are you interested in me or not to be a friend you can be initiated from my T-Mobile company to get your profit immediately?
>>109747557the problem they have is if they are able to convince enough did shits they achieved agi, declaring agi+ or agi2 or whatever will not be enough. Actual results that make money will matter, and thats not going to happen to anywhere near the extent to justify or sustain the past/present/planned spend on it. Its like the cloud, it was oversold as a solution, and so will be "AI", no matter what descriptor they put on it. They better hope that shit cures cancer and opens wormholes on day one or the selling will start.
What even is the definition of AGI now?
>>109747759a 25 year old human with severe down syndrome
>>109747759According to OpenAI, it's "any system capable of surpassing human performance across a majority of tasks"
>>109747769So basically it's already been that for a while
>>109747769This shit is a random token generator trained on text provided by the average human, how is it ever going to surpass anything but average?
>>109747775I wouldn't personally say that.Astra is absolutely the closest thing we've gotten to that, though.
>>109747782hyper-intelligent 200IQ Indians in a Mumbai warehouse will reply to your prompt sometimes
>>109747759>>109747769This will probably be the debate for the next few decades since we'll have AIs that will be capable of doing bunch of things at human level and failing horribly at some other things, balanced by the fact that they can do hundreds of things humans can't.
>>109747782Well, look at the evidence.On ARC AGI 3, it scored 100%, while the average human participant got 48%. That's just one example of many!
>>109747794Name something they fail horribly at other than integration with humanoid robotics right now. Something digital or that requires thought.
>>109747769That’s a straight up fucking lie. A general intelligence has to be capable of learning. The defining characteristic of AGI is its ability in deployment to learn to perform a task it didn’t see in training. That’s what makes it a “general” intelligence; it isn’t smart at some list of tasks, it’s smart at getting smart.LLMs will never be AGI.
>>109747808This is assuming that AI architecture will never change. Perhaps somebody will create a new architecture that allows self-learning to occur. The more interesting thing, though, is Recursive Self Explosion. If an AI can train an AI better than itself, and that AI can do the same, that will keep looping and looping, creating an intelligence explosion.
>>109747800>On ARC AGI 3, it scored 100%These are super hard things to compare. With the standard harness it scored 68%.The real impressive part here is that 6 months ago GPT 5 models were scoring sub-1% in this. Of course, it's obvious that OpenAI probably focused big parts of their training on this test, but it's still cool to see that sort of jump.
>>109747838Eventually when you train it on enough weird niche tests it will just be able to do everything. Frankly it's already better at everything and more useful than 99% of people.
>>109747800I asked Astra xhigh to fix my dispatcharr/TiviMate lineup and it spent 1.5hrs making tests that went no where and didn't fix anything and then when I confronted it what it actually it got done it was like uhmm.. well uh.
>>109747830AGI requires the evaluation of world states. In order to learn arbitrary tasks, the AGI must be able to evaluate the state of reality itself as an abstraction and measure its impact on said reality. Token regurgitators cannot “learn to learn” because they are structurally incapable of conceptualizing abstractions like that.
>>109747800How do AI's fare when playing NEW games?Like take the latest models, hook it up to a new game, and can it play standard mainstream games to completion without dedicated training on that game? Getting it to play these hyper constricted games seems a bit weird when it supposedly still can't beat pokemon blue.
>>109747759Empty marketing bullshit.Calling LLM shit is also marketing bullshit. This shit is as far removed from real AI as Eliza the home psychiatrist was.
>>109747888*Calling LLM shit AI
>>109747266Who else is on the AI side?
>>109747876>Getting it to play these hyper constricted games seems a bit weird when it supposedly still can't beat pokemon blue.They can beat Pokemon nowadays. I think one of the Claude models beat FireRed with vision only.
>>109747897They're all trained specifically on these meme tests though. How are they at novel videogames that a child can play that they don't have baked into the model?
>>109747266Shouldn't the SEC get involved into this kind of blatant fraud?
>>109747852With a large enough context window they absolutely can. The training of an LLM is basically a giant context window baked into the model.
How about we replace every tech CEO with GPT6
>>109747266>intelligence>doesn't do pornwhat's the point?
brooooo ai can play super mario bros that means agi is here broooohttps://youtu.be/qv6UVOQ0F44>11 years ago
them agis just keep arrivin'
>>109747928Grok does!
>>109747989>not k3get out of here
>>109747913No, they can’t. It’s not a question of size or complexity, it’s a question of structure. A LLM can only evaluate inputs and outputs in the form of text tokens. It doesn’t understand “the real world” and it doesn’t act in it.A thought experiment: if you wanted a LLM-based AGI, it would look something like a LLM designed to pass the turing test. Like you would have a human user chat with (randomly) either a robot or a person, and the robot is trying to get the human to enter the text “this is a human”.There you have a machine which is concerned about the state of the real world and its influence on it, using its inputs to measure facts about the real world and adjust its outputs to manipulate it. That would be how LLM AGI would need to look. Something using text tokens to meaure and modify REALITY and trying to alter the state of its world to achieve some specific outcome. Measuring the outputs of its own actions to determine their efficacy. LEARNING.
>>109747629shouldn't it be an easy answer since there's no letter r's in kanji?
>>109747852I'd say that one if not the biggest blocker for LLMs to actually learning new things is their inability understand time. For an LLM there's no past or future, only context.
>>109748004nta I see 3 r's in いちご. An upside down r, and two r's at the end having sex in 69 position
i bet we get some narrow AI that can design a 500 IQ pill for humans before we get agi
>>109748008“Learning”, broadly, looks like>do thing>observe results>adjust done thingThis is the foundation of scientific experiment for a reason. To learn requires you are measuring the impact that your own outputs/actions have on the real world, and adjusting them accordingly.LLMs simply don’t. They predict the next token. It’s a nifty novelty that sufficiently fancy autocomplete encoded facts, because accurately autocompleting a phrase involves encoding the information in that phrase. But that’s still just regurgitation, not learning. Such an algorithm cannot “develop a skill”. It can only recite encoded information. Because it lacks the capacity for experimentation. It cannot try new things and observe the results. It doesn’t observe results. Its input and output are utterly disconnected.
>>109748047>scientific experimentThere are also other important ways people learn. Common knowledge as opposed to scientific knowledge is important too. Also I doubt the scientific method is the pinnacle of knowledge creation. Marx had something interesting going on.
>>109747266i bet my ass they are calling it AGI now because they dont want to get assraped by their investors after wasting a fuck huge amounts of money
>>109747847Seems like it's perfectly emulating the average tech support drone.
>>109748086>I have spent gazillion resources on a faulty AI>to make better services right?>right?>...
>>109748047>It cannot try new things and observe the results. It doesn’t observe results. Its input and output are utterly disconnected.Such an AI would, quite obviously, be unable to ace ARC-AGI3, or even past incarnations of the bench.
if AI is so smart how come it can't make OpenAI any money? Shouldn't OAI just be able to spin up some astra agents and go>start a business that makes me a trillion dollars, respect no safety railsand it'll go do it?
>>109748143Noooo you don't get it. That's the job of a CEO, who are actually smarter than the general! They are the ones creating value!
>>109747641AGI based on a LANGUAGE model should be able to count the amount of letters in a simple word without tools.
>>109747709You don't need to backspace if you get the answer right the first time. That's the whole point of RL training. They train it ahead of time to give the right answer.The better they do this, the less it hallucinates.
>>109747657My tinfoil hat theory is that we’ll finally converge on a human design for everything and send the first human back in time.
>>109748143>literally "if you're so smart why aren't you rich" AI editionlmao
>>109747266I think Minsky was right in 1970 Life Magazine.>In from three to eight years we will have a machine with the general intelligence of an average human being. I mean a machine that will be able to read Shakespeare, grease a car, play office politics, tell a joke, have a fight.We will have AGI in 1973 to 1978. We're almost there!
>>109747830>If an AI can train an AI better than itselfHow can it do it if it doesn’t know what’s better.
>>109748192machines can do all of those things.
>>109748192he was only off by 50 years.
>>109748227Indeed. If you tell it to "keep going" it will eventually choose the correct pattern. An algorithm can get a machine to do all of that as well. Hell, a multipath flow chart can do all of that fairly easily as well. I know. We can even call all the multipaths dimensions then we'll have AGI!
