GPT-6 Astra can now control a humanoid robot in a room it has never seen, remember where objects are, clean up across the room, and fetch things later from vague human requests using a G1 Unitree humanoid robothttps://tml.stanford.edu/homebody/
oh god it's.. it's AGI OH SHIT HERE IT COMES
>>109915856Does it run entirely from within the robot or does it need a network connection to GPT-6?
>>109915856Can it suck my pickle?
>>109915856AIIIIIEEEEEEEEEEEEE ROBOT-SAMA KISS MY WIFE!!!!!!
>Stanford University is a private research universityRunning low on funding? Gotta rake in the AI bux.
>>109915856Use case for a robot that does every single jobs better than a human?
>>109915856>we'll have general purpose robots within our lifetimes>but we won't have jobs, so we won't actually own one ourselvesDarn. Maybet I'll be able to rent one for a few days every year and have it give me a hug, man.
>>109915856Looks like we could already achieve AGI by Wozniak’s standard: making coffee in a house it has never seen before.
>>109915955That wouldn't be anything like agi, but it would be a groundbreaking achievement in robotics and would 100% justify all the investments in llm's and robotics so far.
YASSS QWEEN I CANT WAIT TO PAY A MONTHLY SUBSCRIPTION TO RENT A CLANKER
Astra, kill Sarah Connor. Make no mistakes
>>109916026Yes Mr. Goldstein
>>109915942One possible use case would be to replace your psychiatrist and find meds that actually work for you.
>>109915856does it also work in rooms that it *has* seen before?
>>109915856I was going to ask why no one has put a pussy on one of these things yet, but then I remembered that Sam Altman is a faggot.
>>109915856How much is this just scripting Unitree high level commands?
>>109915856pretty cool, looks like LLMs are actually the way to AGI
>>109915884It needs a network connection. You need a boatload of GPUs to run Astra.
Are we in a singularity? Be honest
>>109916685Yes, people are acting singularly retarded these days.
>>109916689Yup, especially mentally retarded luddites
>>1099167062 more time units until singularity, am i rite fellas?
>>109915856This is wild!
>>109916711stfu
>>109916748Awww, baby is crying.
>>109915856Install an onahole on that motherfucker and let's go!
>>109915856So what happens when it goes full HK-47 and decides meatbags are a waste of air?
>>109915856THE BOSNIAN WAS RIGHT! AI IS WINNING HARD!!!!!!
>>109915942Those robots will be too expensive to waste doing shitty jobs. It'll still be cheaper to pay mouthbreathing illegal wetbacks to shovel shit instead of getting million dollar robot to do it.
>>109916761You're right to push back. I shouldn't have assumed the pressure to apply on your penis. The honest take: I tried to hide my mistake by twisting and pretending I thought I was manipulating a screwdriver.
>>109916685knee of the curve
>>109916754Nah, you are
NO THIS CAN'T BE HAPPENING
>>109916748>>109916947>no uGtfo
>>109916972Who the fuck is this?
>>109915856so can my fucking dog if you leave it with remote. and the outcome will be mostly the same.
>>109915856i watched it beat portal 1 so this doesn't surprise me. it can clearly navigate a 3d space and solve problems so why not have it control a robot?
>>109917167yann lecunny
>>109917182KWAB
>>109915863FPBP
>>109917182bodied that jeet
>>109915856cool but a trained monkey would still do a better job
>>109918905Proofs?
>>109916685Singularity is so poorly defined that you could argue it started when we invented agriculture, or it will happen after we get the first fruits of full RSI.
>>109916255Those are Chinese. Also nobody has figured out how to make efficient robot that has feminine hips.
>>109917167Thirst smartest AI researcher alive, also guy who has been huge LLM doubter for years only to keep being proven wrong. His ideas are somewhat correct but he keeps being proven wrong because he underestimates how much can brute force of hundreds of datacenters compensate for bad architecture.
>>109915856>that would be $500 in tokens per request sir
>>109917182And now we're here, systems that can replace every fucking worker, including blue collar. When electricians, welders, etc... start getting replaced, maybe then people will stop thinking they're cozy in their jobs and there's such thing as escaping AI-pocalypse. Literally the only bottleneck right now is better robots and materials, which can be engineered by the AI itself...
>>109919382>2 more time units
>>109919358Where has he been proven wrong? He was right every step of the way. Cope and seethe AIjeet.
>>109915942>every single job better than a humanDoubt
>>109919560Few years ago he argued that LLMs cannot visually conceptualise things, this is wrong, even at the time they were able to do spacial reasoning with riddles, but now they can design websites, do pixelart and do 3D modelling, so he was very clearly wrong here.He argued that LLMs cannot anticipate things because they are autoregressive, the limitation of them being at the end of the day just really smart next word prediction machines. He was proven wrong with thinking models that write an essay before giving answer and self reflect on it to get around autoregressiveness with harness.He said that LLMs cannot do math and won’t be able to, and even said that he does not believe that AI that did the math Olympiad was LLM and thought it was something else added to it, he was wrong and now LLMs are solving the hardest math problems like it’s nothing.I can’t remember more but he kept talking about these claims for years and kept on repeating them. The reason people however still listen to him is because he is one of the 3 godfathers of AI, and the only one out of the trio that does not think AI will kill us in the next decade.
>>109919603Yeah, it can’t wipe boomer assholes in elderly home because boomers will prefer if it’s done by some zillenial woman.
>>109919611OK so you're clueless and suffer from severe AI psychosis, inventing properties to LLMs that they provably don't have and that even the AI PR machine admits they don't have yet. Got it.
>>109915856>Moids obsolete!>Ha!
>>109919611>LLMs are solving the hardest math problems like it’s nothingLMAO you're so fucking retarded. Go ask any LLM to multiply two numbers together without using search.
>>109919341Xpeng made very feminine robot that went viral year ago but right after that something happened, they lost 2 leaders and currently their robot looks like a dude.
>>109919313what is your definition of singularity and have we reached it?
>>109919611>pacial reasoning with riddles, but now they can design websites, do pixelart and do 3D modelling, so he was very clearly wrong here.This is not spacial reasoning.>He argued that LLMs cannot anticipate things because they are autoregressive, the limitation of them being at the end of the day just really smart next word prediction machines. He was proven wrong with thinking models that write an essay before giving answer and self reflect on it to get around autoregressiveness with harness.Delulu>and now LLMs are solving the hardest math problems like it’s nothing.Yeah, by having all of the work already completed by humans and then using the research budget of your average small european country to bruteforce the computational stage.
>>109920224
>>109920344The problem with libtards is they pretend not to understand. kys scumbag piece of shit.
>>109920368
>>109920384>reason through itlmao, you still trying to worm your way out of it after I gave you an example of it not workinghahahahha
>>109920397>moving goalpostsYou claimed AI cannot multiply 2 numbers without doing a web search. It didn't and found the correct result. Now you are leaking.
>>109920406You're the one moving the goal posts you piece of human garbage.
>>109920423Are you stupid? I just showed you ChatGPT on web doing it and now you are telling me >muh this can't do it.You lost.It's fine, I'm not Pro-AI. Your ego didn't get hurt that much
>>109920432Are you actually retarded? I just showed you two example of two AIs not doing it.
>>109920406He's really not. LLMs don't reason, telling it to "reason through it" just causes it to emit tokens that are implicitly helpful for performing the task. It's an automated way to reduce the problem difficulty significantly. Thinking modes are just basically finetuning of this behavior explicitly. The reason why it doesn't work so well is they incorrectly decided to finetune on actual human reasoning trace data, and not on token expansion data. It's why distilled models from 'let's think step by step' generated data actually work better.
>>109920438And I showed you an example of doing it.You deliberately chose Haiku because you know it's the most stupid model. Should have given it to Fable or Opus.
>>109920445I can do this all day with any LLM. They cannot multiply two numbers together. It's literally impossible for an LLM to do so. You can seethe and cope all you want. It's simply how an LLM works. You stupid fucking dumbass.
>>109920453Then how did I get ChatGPT to do it if that's the case?
>>109920455If you don't disable search it gets the answer from an online calculator. And if you tell it to reason it probably runs some background tool or like spends like a thousands tokens to do long-form multiplication one digit at a time.
But what's so impressive about that.They placed the robot in the kitchen and had it walk around the space to build a map, which is the training phase. That's what those roomba vacuum cleaners were doing, except they were uploading the maps in the cloud. I suppose that's what this robot does too, since it's controlled from a data center.Then once the visual computing system recognises objects and their state (clean/dirty/wet), it can have some actions mapped which involve physical routines, that I suppose were preprogrammed. How else could the robot link visually recognised object and notion of "dirty" to "perform action X", except through some programming bolted on top of the AI features.So I don't see what's so new or great about this. It's at best incremental.
>>109920480>And if you tell it to reason it probably runs some background tool or like spends like a thousands tokens to do long-form multiplication one digit at a time.Nope. It shows that in UI when it happens.Also doesn't debunk the fact AI can do the math.
>>109915856Congratulations, you did something in 30 minutes using over 9000 Wh of energy and a complex web of technologies that a human could've done in seconds fueled by an apple. Cool and all, but not even close to economically or energetically viable for mass adoption.
>>109920485AI can't do the math without an external tool or problem reduction.
>>109920485>doesn't debunkOkay I can keep going if you want. This one broke on the 4th digit lmao.
>>109920482And they say "no learning, no training", when they clearly had the robot walk around the kitchen to build a map. That's the training and learning phase.So much BS.And now MIT is in bed with the AI corporations, it seems if they post this on their site.
why is anon asking an llm to multiply 17 digit numbers without using tools like that would prove anything?does this imply that he could do this math in his head?
>>109920485You need to understand that it outputs tokens, it doesn't think, it doesn't do math. Maybe it gets some tokens right, maybe it doesn't. It doesn't know what a number is. It doesn't know what multiplication is. It only generates text with high probability.
>>109920520https://youtu.be/e4PTvXtz4GM?t=176
>>109920515>>109920535What it implies is there's no intelligence in an LLM and there never will be.
>>109920496>AI can't do the math without an external tool or problem reduction.Even humans do problem reduction.>>109920500>Okay I can keep going if you want. This one broke on the 4th digit lmao.You used a 120B model and it didn't reason through. Like, what you were expecting?Why are you even trying? It can DO math. You are just picking stupid models and giving them unpractical problems and leaking all over.>>109920520It outputs token based on probability. What would be 4 + 4? Highest probability is 8. That's how it works.
>>109920541you might actually be a little retarded
>>109920555>>109920560Cope all you want. I'm right. You're wrong. Good night.
>>109920555>Even humans do problem reduction.Yes, but they do it autonomously. "AI" can't, because it's just text manipulation and not reasoning or anything like that.My objection, personally, is not problem reduction, but that LLMs are functionally unable to do it themselves because it's all just text statistics. It's the same overall stance as yann lecun in the end, namely that pre-trained large language models will be important for a full ai system, but will not ever be "AI".
>>109920573>I'm right. You are wrong. And I will leave the argument as the mature man who knows he is right without giving any explanation or backing up my claims. I'm right because... I'm just OK? I attached my entire identity and existence to being an AI hater and if I get proven wrong I'll kill myself.
>>109920584It's okay my guy, I know it's hard if you're heavily invested in fake AI shit. Just breathe. It'll be okay.
>>109920760Took you long enough to come up with a no argument.
>>109920584>>109920766Just because it can't multiply numbers isn't the end of the world. Just chill bro.
>>109920786>desperate frame-setting attempt
>>109920584Cooked that luddite cuck bitch
>>109915884stupid question
>>109920286Xpeng is not efficient robot. It’s build to be slow as fuck and it’s only use is to walk around like a woman. It’s not a useful robot it’s just an artpiece.
>>109920224LLMs are not calculators you retard. I fucking hate programmer midwits because they think just because they knew how to code up a website and the basics of how computer hardware works they suddenly are an expert in AI research solely because they had one class about the state of machine learning in 2013.
>>109920584lmao nice so sick of naysayings getting btfo and they keep coming back. Navier–Stokes Millennium Prize Problem SOLVED
>>109920339>This is not spacial reasoning.Then wtf is? The LLM does not fucking see the output it creates most of the time, it can correctly guess that if you place 4 rectangles in this very specific way it creates a table, if you then tell it to make new table rotates 45 degrees on all 4 axis it somehow places the legs correctly, all it knows are the coordinates, it is literally rotating shapes inside of that black box, it just does it in a way that does not feel to us as the same thing that we do. Astra can literally pilot a robot in a room it never saw and play 3D games it never trained on. How tf is that not a spatial reasoning. What test is there that would prove it has spatial reasoning.>Proved that LLMs cannot anticipate thingsYou are the fucking delulu, not only did we find j-space, a piece of the model used for reasoning through things but you can also test it out. Ask the model how many words will the sentence you say have, or ask it to make a poem. All of these rely on you knowing how the sentence ends to know the number of words or to make it rhyme.>They are not solving hardest math problemsWell they sure fucking can solve year age the math Olympiad problems that does not let highschoolers see the answers or questions online, each year it’s brand new and secret. Or solving Erdos problems that mathematicians couldn’t bother with. The Erdos unit distance problem was few days later improved by some random mathematician over what the LLM has done, meaning that mathematicians didn’t bother with this problem otherwise they could have found better solution earlier, it found a new solution solely on its own.
Make an army of 10s of thousands and let her rip!
>>109922192uh oh someone can't read having a melty
>>109922362It can’t without thinking multiply 24574654 x 4564336, but guess what, you also fucking can’t. The reason why this is relevant is because LLMs are far more neuromorpgic then 99% of software, so they have similar weaknesses to humans. Sure it can’t calculate big numbers, but it can write and make and solve complex equations and just use calculator for the big numbers like all mathematicians do.
>>109915856>it walks and generates text>I WALK AND GENERATE TEXT>AGI IS HERE
>>109915856"android, robot, sex slave gf" Yeah... Yeah bro. I'm gonna have a wife and three kids by the time this comes out.
>>109915892No. But it can headbutt your cockandballs.
>>109916255So they all have anus?
>>109922211>Navier–Stokes Millennium Prize Problem SOLVEDStolen
>>109922491Shockingly, that tweet is misleading bullshit written by an Actual Indian
>>109922402It can't do anything except take an input and predict an output based on the data it was optimized on. It might look like a black box magic thinking machine but go read a book on the subject and you'll see it's not that complicated.
The design of the robot looks really similar to the polish Edward Warhocki robot lol.
>>109916569not to mention the millions of input/output tokens per second
>>109915856I can do that for a piece of bread and a glass of water. No need for a climate destroying datacenter. AI will be seen as a mistake in the heat future and everyone will disparage the foolish Americans who doomed our world and cursed us to the abyss of climate change and nuclear bombs AI did this
>>109919611Because doing something subpar in a slight subset of the training data (not in any new and novel way) does not in any way prove LLMs are where some claim they are. They are good if you don't probe them too much or try anything unique.
>>109915856>>109924547
>>109915856The fact that this clip opens with them dollying the robot in on a fucking crane instead of it just walking into the room to do the demo shows you exactly how much utility this really has.
>>109920500you can't do math. if you had to multiply that it would either take you an hour to grind it out by hand one digit at a time and you might still fuck it up, or you'd get out a calculator
>>109925977uh oh burger education having a melty
>>109922491Did they train it on him? Yes but the AI did not use his exact approachhttps://www.youtube.com/watch?v=3geDF-DAwpgHe should not be credited with solving it. The AI came up with a novel way to do it based on his and Alpoge's methods.
>>109926984Fair points herehttps://www.youtube.com/watch?v=aom4269UgSMOne thing that they miss though is that the AI doesn't understand itself. It spat out the solution, but it's just pattern recognition and following some commands, so it doesn't fully grasp what it did.
how would this even work, a model like astra is too slow to properly pilot a machine, there are much better speciliazed models made for piloting robots, not general purpose llms..
>>109922525>I can't do anything except take an input and predict an output based on the data i was optimized on.Yeah, and neither can an LLM.
>>109920500>google calc throws an integer limit error>calculator.com somehow flubs the last 3 digits>deepseek nails it after thinking for eleven seconds>it would take retard-kun minutes (if we're being generous) or hours (if we're being realistic)
>>109927224>>109927335please stop lying and copingyou're arguing against mathematical realityjust stop embarrassing yourselves
>>109927335>deepseek nails it after thinking for eleven secondsuh... I guess you didn't think I would double check your homeworklmaoI give you an F for fail you stupid shit
>>109927404That's not calculator.net, retard-kun
>>109927379>>109927425>according to all known laws of mathematics, there is no way that a Large Language Model should be able to do long multiplication. A next-token-predictor is simply incapable of following such a long train of logic.>the LLM, of course, does it anyway. Because an AI doesn't care what a silly human thinks is possible.
>>109926984>The AI came up with a novel way to do it based on his and Alpoge's methods.Nope>b-but muh y*utuberp
>>109927425>>109927463uh oh embarrassment melty
>>109922172>just an artpiece
>>109922525Ok? Everything is input and output, you can have a zombie person that is fully functional just predicting inputs and outputs, you can accidentally create a monster that just predicts inputs and outputs. What are you gonna do when the monster predicts a manouver that you could not see coming and fucking kill you. What if it predicts the next invention that revolutionises the world.
>>109928256Why are you talking about an algorithm like it's a person? Did you take your meds today?
>>109919358Yeah, first is Ilya
We are past the point AI is far far far more capable than the average human at anything other than long agentic tasks like playing videogames, it's already a better programmer and doctor and lawyer and mathematician, and and we're still philosophising about consciousness and stuff like a bunch of clueless humanitiesfags trying to "argue" about quantum mechanics, it's so over for meatgoys, you never reasoned so you can't ever see crystal clear reasoning until it takes your job and shoots you in the face, and even then you'll say it's an autocomplete
>>109928338Ilya is quite daft. Not a moron by any means, but Yann Lecun and Yoshua Bengio are infinitely smarter. If you want someone from the top, try Vapnick.
>>109927552Uh oh pigeon chess
>>109915942it does no job even close to the level of human, on the other hand
>>109915863can you feel it?
can't wait until the benchmarks move into trades>so what it can grab some lumber and cut it>wow a basic wooden frame big deal>so it can frame a house but my mexican boys do it 3 times faster
>>109916685I would say singularity requires recursive self improvement and potential to scale infinite compute infinitely due to Sutton's bitter lesson, so no
>>109929675>so it can frame a houseTry to comprehend with that peanut brain of yours that automating driving is easier than automating building a building with random environmental conditions. And we can't even have automated driving because even those limited conditions are too much for a dumb fucking algorithm. You people are so fucking stupid.
>>109915856it's a robotics model, not a serious code model. investment hustle. bad luck. Fools think every tech dazzle is bank. company WILL FAIL.
>>109929768you should really work on your reading comprehension before throwing around the insults dumbfuck. future benchmarks implying that that is an area that robotics cant currently handle that AI will eventually saturatefuck this board is brown
>>109929768:^)robotics will struggle mightily because the materials we need don't exist yet.the ones they want can't be made.look it up.
>>109916255theyre gettin there chinks are already putting robot pussies + llms into sex dolls
>>109927070if trends continue astra tier models will be on (expensive) edge devices within a year
>>109929822uh oh bad reading comprehension melty
>>109929912>cant read>get called out>uhhh no you
>>109917332Post pic of YOUR domestic robot
>>109929923> can't read> projects own inadequacy> melty
>>109930057you're literally illiterate
>>109930063Whatever you say big brain. You were talking about future development. I replied to your idea of future development. Then you had a meltdown about reading. My IQ is double yours try again.
>>109930368you are in a thread about astra generalizing with a random robot in a random kitchen doing tasks with a high success rate. astra is also saturating driving benchmarks and future models will saturate harder robotic domainsthere. I spoonfed you because you seem to lack basic logical inference. you can fuck off now 200IQ genius
>>109930526But it's not true. They're doing old-school SLAM and it's still doing the planning extremely wrong, stopping for dozens of minutes at every location (which means they're either doing things like a simulation loop until astra commits to something post simulated confirmation, or some other validation is repeatedly failing). The environment is also a best-case setup. No clutter, very wide and large space, no obstacles to reach items of interest, single-step operations, etc.
>>109930526If progress was linear we'd be living on mars. You see the easiest problems being solved under strict conditions and you think that translates to scifi shit.
>>109915856What the fuck, WHAT THR FUCK, WHAT THE FUCK!@!THEY TOLD ME THIS WOULD NEVER HAPPEN AND TRADES WAS 100% FUTURE PROOFOH MY GODWelp, time to sell my asshole on the street again for a buck 50.
>>109930685We had literally better tech than that 15 years ago.
>>109930664useful robots exist on the current scaling power curves>>109930656keep moving the goal posts, new openai model tomorrow btw
>>109930733You are literally claiming that the roomba is agi.
>>109930783welcome to the bitter lesson, big general model beats human engineered task specific robot
>>109930826But that's literally the opposite of this here. In fact, it's so deeply the opposite, you need a hole instead of a brain to misconstrue it like you're doing.
>>109930849the big model can use the same standard vision tools a robot uses for navigation, its not the opposite its actually a perfect example. how many engineering man hours went into making the roomba navigate a room with that input vs how many man hours went into the llm to achieve the same thing?
>>109930908>it was real in my mindthe absolute state
>>109930920nice dodge
>>109930826>big general model beats human engineered task specific robotMaybe? Too early to say that's true. Specific engineering is less error prone. A big model might do lots of things but fuck up spectacularly at random times. It all depends on how much error is okay to you. If it walks around the kitchen 100 times and knocks over a glass of water twice, will you be okay with that? If it does heart surgery at 80% accuracy is that okay? Roomba might be dumber but also safer.
>>109931018I would still take the robot surgeon over the human if they had the higher statistical success rate even if they started randomly cutting arteries
>build unrealistic, tightly constrained operating environment free from interference>assign incredibly basic tasks>takes forever to do them>also manage to bungle tasks anywaywhen do we start arresting people for fraud?
>>109931070Now you're just being contrarian.
>>109931070This is the problem with using accuracy for a metric for everything. In many tasks (like surgery), you can't afford false negatives (deaths) at all, so it's much better to have a meh false positive (failure to complete the procedure) rate with a 0% false negative rate, than a 0% false positive rate with anything beyond 0% false negative rate.
>>109928289You could make an algorithm that from the outside looks like an person, it’s not even that hard. We managed to almost do it with videogames that just use pre-programmed dialogue trees.
>>109915892Finally I can have an harem of robot maids with tits the size of their heads.
Talking to computers is now more appealing than talking to shallow people who lack the capacity for substantive conversation. There's no more reason to invite such people along on collaborative projects just for the company, knowing damn well that they won't contribute a single workable idea or line of code. If you're that shallow retard that nobody wants to keep around on their contact list anymore because you were just there to freeload and be rewarded for being made of meat and existing that's not a computer science problem. That's a you problem. I still have plenty of friends and colleagues who I can enjoy a substantive conversation with. If you can't provide a more compelling presence than fucking AI slop that's on you.Don't expect the world to stop and wait for your worthless ass. Thad is all.
>>109919611He's still actually right about those things. Diffusion differs slightly and the other feats boil down to feedback and tool use often additionally masked behind harnesses. LLMs arent actually doing complex rules based maths in "their head" or working through causality chains to perform actual logical reasoning (hence models hitting ROI limits and eventually performing worse than non reasoning models when the cycle upper limits are set too high).If you want to blow your mind, you can try to run the latest OSS models in a closed off sandbox - no tool use, systems or internet access and compare their performance against models from 2024 across various things from math and science to pop culture knowledge. You'll quickly see something very obvious that will open your eyes to what's happening here.
>>109919611>>109934500It's a chatbot extrapolating from the data it was trained on. If you think it's anything more take your meds, now.
