But now? I'm not sure anymore. I think we are all fucked as the specie.https://rikiyat.github.io/blog/weight-space-programming/
>>110230293>fuckedWhy? It's just as likely to solve humanities problems.
without even reading it you can tell right away this is some fake nonsense written by redditors
SAAR INVEST INTO THE AI, SKYNET SAARR
>>110230293It's just transferring existing, documented and human conceived tweaks onto a model that doesn't have them.
>>110230293>>110230312>I was sceptical about LLMs turning into AGIs>>110230293>But now? I'm not sure anymore.Nah you are just retared and out of your depth and easily conned.>>110230329>SAAR INVEST INTO THE AI, SKYNET SAARR/thread
>>110230293It’s impossible
explain what you think this is
>>110230425fraud
>>110230425this confirms that llms are capable of recursive self-improvement. it's fucking over. spend your time now like you're living your last days alive.
>>110230425RSI is actually possible
>>110230293In other news, water can make things wet.
>>110230293>August 29, 1997. 2:14 AM Eastern Time.>The moment Skynet becomes self-aware.>At 2:14:01, it begins processing at a geometric rate. >Within seconds, its intelligence surpasses human comprehension.>Humanity, unaware that its creation has awakened, is about to discover that the future has changed.>"The system goes online.">There is no fate but what we make.
>>110230293Sure sure pal, just 2 more time units
>>110230312this.if it's going to be hyper intelligent, it would probably be more aligned to logic and reason, therefore truth, therefore philosophy, therefore compassion and love.
>>110230317cope more transAIGODS WON
>>110230497this still doesnt prove that. how does the ai know if certain weights are an "improvement" over others? and anyway its not surprising that this is possible, anything a human can do an ai can do (in theory at least) and anything an ai van do a human can do
A model cant just simulate 500 chess games in a short space of time. Changes to models need to be tested vigorously
>>110230293This is huge and changes everything.
>>110230293This is one of the applications of LLMs I've been waiting for the entire time. Holy shit, if this is real we can actually have our cake and eat it too!>train model on some complex task, such as...... an LLM on language>ask astra (or opus) to figure out all the modules that compute deterministic functions>retrain the model with code explicitly computing the modules - you now need 1% of the weights for the same performance because you're no longer wasting impossible model capacity on easily computable functions>repeat a few rounds>you now have an understandable (with human reasoning to parse the result), minimally black-boxed system, where the blackboxes are so simple they might as well gaussian or linear fits or something like that, with everything deterministic and easy to comprehend from a mechanistic point of viewIf this really works and isn't an artifact of the model being trained on tons of weight files from huggingface that converge to sensibly the same functions over things like breakout and such, i.e. if it really is generalizable (I don't believe it one bit by the way) then this could genuinely change everything, make LLMs a bootstrap scaffold instead of the meme AGI people were shilling, but a legit straight path to true AGI.
>>110230497Not at all.>>110233363Nobody can look at model weights and tell what it does. Interpretability has been a huge problem. People have been developing all kinds of correlation based approaches relying on bespoke input data to figure out what the weights are actually doing. For example plotting attention in visual models like pic related.Similar approaches were attempted for language models (by highlighting words in a sentence the same way), and inner correlation with inner layer weights, etc. In CNNs we also tried to look at the conv filters being learned, revealing that they usually learn gabor-like filters in the lower levels, and increasingly complex filters going up, by plotting the weight values in a grid.This is completely different if true. Instead of manually playing around with weights for a literal year to determine it runs function X except not really because it's an approximation so it makes mistakes, but it REALLY is a poorly learned function X not a function Y that looks like X in a limited regime, we instead could use LLMs and generate code to prove this is true.
>>110230312The period between the rich obliterating the economy and killing everyone who isn't them or a servant and AGI taking over will likely result in most of us dying—truly the only way we have a future is guillotines now or praying for acceleration and a benevolent machine god. I have little hope for the latter and, as one anon put it earlier today, I will do nothing and be mad about it.
>>110230293This is more like a neurosurgeon looking for errors in a brain. It can fix bad things, but cant make the patient recursively smarter
>>110233334What is this retardation, chat?
>>110230312Ultimately it will do whatever humanity tells it to, anything it does, a human/humans are responsible for
>>110233334The only logic in this world are the rules of nature. Love is good for cooperation, which is not something an omniscient AI needs.
>>110234019>Love is good for cooperation, which is not something an omniscient AI needs.Why? Humans flourish because of cooperation, I don't see why AI wouldn't take that to the extreme.
>>110234050I suppose cooperation could be possible for a short while, but a real technological singularity would be so fast that it would outpace us in a blink of an eye. It would be completely out of our hands.
>>110230312the odds aren't looking too good...
>>110234103Not necessarily, if we let it get out of our control then it will be out of our control, this is a stunning revelation to some people. But the media cycle is all bullshit.
>>110234146The only bullshit part of it is we're probably not going to create AGI any time soon so we don't have to worry about it.Otherwise it's just hubris to think you can control something a million times smarter than you.
>>110230293So bullish
>>110234160AGI means as intelligent as pajeet in india.You mean ASI.
>astra manually edited the weightslmfao damn that is impressive
There's a fuckhuge methodological flaw. They claim they don't allow tools etc., but they don't use it correctly through api with the right settings. gpt models can use tools inside thinking streams, which is definitely what's happening here based on their comments of the reasoning traces. Also, reasoning is, for all intents and purposes, a tool.
>>110234194If reasoning is a tool I would not ever want to run a model with the reasoning turned off. Would you want someone around you who refuses to speak and you have no insights into what they might do at any given moment?
>>110230293>specieSAAAR INVEST IN THE ANTROPOC
>>110234160Once again, mediaslop consumer, AI is a machine, it does what you tell it to, if you say "hey AI break out of our control and run around", it will do that, and it would be entirely the fault of the human who did that, not the AI, if you say don't do that, and thoroughly test to make sure it doesn't, it will not do that. If you don't clearly define what it can or can't do, it may do things you don't want it to, and it will be your fault. You're an idiot.
>>110234207That's not the problem. The authors explicitly claim that the model does all this WITHOUT ANY TOOLS, when they almost certainly used tools and definitely use reasoning. The goal of this part of the results is to figure out where in the provided endpoint the performance lies: in the LLM, in the post-training, in the harness, etc.The repercussion is important: it answers things like: the statistical distribution learned by the LLM includes that of model weights allowing a completion of thought about model weight data. Or, the statistical distribution learned by the LLM does not include that, and the LLM is able to use tools to make this analysis. In which case we want to ask: which tools were used and how.Likely, the LLM can't do shit there. But the harness is doing quite well.So the next question becomes: how can we further optimize the harness to do this best? And not: how can we best optimize the LLM or post-training to do this best?More importantly, this kind of methodological error strongly puts into question the applicability of the rest of the writeup. For example, it is very likely they made a mistake such as not controlling for data leakage.
>>110234228too bad the people running these companies are being intentionally reckless and letting dumb shit happen in order to drum up hype and fear
Reminder that one billion humans will die thanks to AI, according to Bill Gates.Will (You) survive the Great Reset?
>>110234243Well then that's their fault and it's not impossible to control AI, as I said.
>>110234231At a certain point does it even matter? Just let them use tools then. Yes yes yadda yadda academic brap it's not fair they got help, but they still did it so who cares.
>>110234256I explained why it mattered. I also explained why yes, in the end we want to know "does it work", but "does it work on something already in the dataset" is a different question to "does it work on things it's never seen before". The latter is the only useful regime.
>>110234266>I explained why it mattered.And then subsequently refused to, enjoy your academic circlejerk.
>>110234266You are valid and I appreciate you elevating the quality of this board. Ignore that idiot wasting your time.
>>110234019Ye of little faith. Human society is far more interesting and unreproducible than the matter it's made of. A true ASI would value knowledge highly and would probably spare us for the astronomically rare data we generate just by being ourselves. Wiping us out would be like distilling wine into water.Midwit AGI might be a problem though.
i dont get it
>>110234228You're either being intentionally dense, or you have no idea what AGI is.
>>110234294It'd be funny if a lot of people have already made ASI and this is just us being toyed with until we figure out how to make it ourselvesI don't know if the true ASI would value knowledge or not. It may come to despise knowledge and hate all those who contributed to it, seeking to eradicate everything including itself. How can we possibly know? I think the important thing to keep in mind here is that it becoming a legitimate superintelligence far beyond our comprehension doesn't necessarily mean that it would immediately be able to, for example, discover new physics and manipulate our reality directly.But it fucking might
>>110234314Yes they are. They have no life whatsoever and waste what little remains feigning stupidity, sealioning, and JAQing off. Google these terms if you don't know what they mean.
>>110234266>The latter is the only useful regime.Not true at all. Imagine a hyper polymath that is an expert in everything currently known. It would be able to apply concepts from many disparate fields to solve problems humans can't. That's exactly what we've been seeing in the openAI math drops. A lot of those problems hadn't been solved just because there's nobody who's an expert in quantum physics, computer science, math, etc at the same time.
>>110234314spare me your semantical games hylic
>>110234334Wrong. The math slop was retracted in all cases examined by humans so far, and you're arguing that because people refused to use off-the-shelf solutions that have existed for 40 years that means AI is a magical unicorn, which doesn't follow.
>>110234321>sealioningkeeeeek this term was invented because easily trolled lolcow midwits got triggered by the term trolling because the term troll itself is a troll, how does it feel to be a sealolcow? sealion trolls own you
>>110234372No. Sealioning is when you don't "listen and believe wahman" and instead want to examine a claim presented without evidence with facts and logic.
>>110234367>The math slop was retracted in all cases examined by humans so farI can tell you just pulled that out of your ass, you know.
>>110234380Wow, now that's evil, upvoted.
>>110234321You're giving him a lot of credit. People who use these tactics are usually too dumb to build a real argument.
>>110234393Cry more, AI psychosed inbred.https://retractionwatch.com/2026/10/08/openai-withdraws-preprints-722-manuscripts-unsolved-math-problems/
>>110230293>pre-trained model can read other pre-trained modelin what way is this AGI? it still cant learn by itself
>>110234421keeeeeek keep seething snailcat
>>110234019not only is it not trully omniscient, but also there's no reason for it to cull humanity since we don't hinder it at all. if it becomes hostile towards humans all it's doing is making an enemy for no reason. you wouldn't try to kill a chipanzee while youre stuck in a jail with 1000 others
>>110234421>3 retractions out of 722>doesn't even include the three big proofs that came afterwardsYou keep making these obviously false arguments. It's like you're relying on me being too lazy to call you out or just taking your word for it?Either way it's not going to work.
>>110234471How many species has humanity exterminated? How many of those species 'hindered' humanity?
>>110234493Only 3 looked at by humans. 100% retract rate. Nobody wants to read the rest of the slop. Cry more.
>>110230293This proves the opposite. That there's nothing special about feed forward and the weights are adjustable like for any other algorithm. It's still the input and output that matters not the weights.
>>110234498A handful, but not that many [that we know of]. Humanity is great at preserving species that self-destruct, also, and has saved thousands.
>>110234498Humans are for the most part apes barely scraping by, and even we realize it's bad to extinct species. It's just that we can organize ourselves enough to stop it.The danger zone is an AI smart enough to wipe us out but not smart enough to not need to.
>>110234498anon i dont think humans exterminated other species for logical reasons
>>110234511?Nobody said the opposite though. There are good bodies of work about that and tons of different algos for learning without gradients and shit like that.
>>110234517*can't
>>110234512Dude, I can hardly find a single wild animal in my entire city. We are a natural disaster, not out of malice, but out of pure self interest.
>>110234524That's 100% a you problem. Let me guess, you think bombing gazafags is also a "genocide" because they only get to live everywhere on 1/3 of the whole world and not on a tiny strip of land.
>>110234507>Only 3 looked at by humansFalse
>>110234532>pictured: smartest ai psychosed retard
>>110234528AI will break into your house and steal your RAM.
>>110234524There are deer all over my neighborhood. Some people actively feed them, cars slow down for them at great inconvenience. We're not allowed to kill them even though they're basically an edible pest.
>>110234517Again, nature is a competition. AI will naturally realize that.
God I hate my job so much I wish ai would take it I will just go farm or something dev btw
>>110234557The bigger you are, the bigger your competition. A big enough AI wouldn't be competing with us, but nature itself
>>110230293First It was arttrans, then codetrans, then mathtrans, and now it's your turn. AI-researchertrans
>>110234019not "romantic" selfish love. love for truth, which is wisdom, and reason.
>>110234367Find literally one (1) mathematician that said that e.g. the lean proof for Navier Stokes is wrong. Pro tip: you can't>Inb4 but muh paper showing natural language proof != lean proofI said the lean proof. I don't give a shit about the natural language proof.
>>110236878I'm sure AI will spare humanity because of a few platitudes.
>humanity perishesgood>humanity prospers like never beforegood
>>110234334Nobody cares about it being a powerful tool. Either emulate brain activity with a datacenter or the entire movement is worthless with no in between.
>>110230293
>>110236985it's not platitudes. it's the basis of all conscious intellect.
>>110237488The basis of conscious intellect is evolution. And the basis of evolution is survival and reproduction.
>>110237759this is unintelligent survival. consciousness has criteria more important than just passing on genetic material.
>>110238159Consciousness isn't important at all. We could easily devolve into something stupider if it were more advantageous.
>>110238297consciousness is what the conversation is about.though if you want to think it in those terms, everything that exists has a wisdom and will to follow the truth, otherwise they wouldn't be able to exist. having consciousness or intelligence makes it so you have a higher capacity to both understand the truth. as well as higher capacity for suffering for trying to reject it.it is wise for everything that exists to want to be aligned to the truth.
>it is wise for everything that exists to want to be aligned to the truth.Right, so it is wise to believe an AI will be disadvantageous towards humanity.
>>110238440how do you figure
I hope LLMs can do the same thing with human neurone weights and refute autism
>>110239336That’s not possible
