Ilya Sutskever Just Revealed The Next BIG THINGS In AI (Superintelligence Explained)

[@Regnum0nline]

Agi is reached when it tells ilya to shave his head

[@tribeleader0212]

lmao. Honestly though, which look is he going after with that hair

[@Comenta-san]

@@tribeleader0212 it's an antenna to capture the ASI signal 📡

[@chastetree]

He's fine. I much rather he focus on developing ASI than fussing with his hairline.

[@ParallaxOfficialTV]

😂

[@namenl2205]

😂😂😂😂😂

[@fredrikt6980]

I am super IN on AI but I seriously think you are confusing people by interpreting Ilias speculations as facts and information. It is almost to the point of clickbaiting. But most channels today are mostly about driving traffic for self-revenue. Would be nice with some more balanced and critical perspectives to provide viewers with a clearer understanding of what is speculation versus proven information. AI is an exciting and evolving field, but overhyping or misinterpreting ideas can do more harm than good, especially for those trying to learn. A bit of skepticism and fact-checking would add a lot of value to your content!

[@filmfest]

You would never see it sure due to the algorithm. 😢

[@albericosessa2814]

I’m starting to wonder if this is all just hype for investors

[@Felipe-zl1rj]

No doubt about that

[@jackburton8499]

It's both

[@DWSP101]

No, I don’t think it’s hype. I think it’s an inevitability, but it might be a little bit slower than what you think to achieve it but faster than what people will expect. I believe we have super intelligence sooner than 2026.

[@Felipe-zl1rj]

@@DWSP101 2026? Not a single chance. Maybe 5 to 10 years from now if and only if we have some crazy breakthough that doesn't exist yet.

[@toadkiller4475]

@@Felipe-zl1rjoptimistic and tempered predictions both have very good arguments. There are many factors and new data everyday that one must take into account to more accurately predict when. Many people may be pretty familiar and up to date in the fields that pertain to this speculative and theoretical achievement but no one can utilize all of the ever expanding data to make prediction. Personally I take the more optimistic time frame prediction and don’t see hardware or software hitting a wall of resistance. It’s difficult because the rate of exponential growth changes almost daily based on what is being achieved. If met with any resistance I believe it will be legal/regulatory, enforced by those who see such advancements as a threat to the structure in which they maintain power and control in society.

[@DanielStonehill-ld9xv]

Ilya's actual talk is quite accessible and well worth watching. You will get more from the original.

[@ilyaglinsky7905]

Link?

[@braidy4450]

kzbin.info/www/bejne/Z6rZc6SXqK1_kNU

[@TheAero]

He is right, data will finish. But it's not yet over, not even close enough. There is so much work that can improve and increase data 1. Data quality (mislabels) missing data, wrongly tracked data 2. COT data, teach how to thing 3. Multi-modal training data 4. Artificially generated data 5. Data between AI tools, or AI to human interaction training. Bonus: We need to teach LLMs WHY to thing each thing

[@exhibitD79]

You don't think he thought of any of this?

[@TheAero]

@exhibitD79 then this presentation talks about a problem far away. Cause we get loads of data to train for the next years.

[@exhibitD79]

@@TheAero True. But that data isn't going to get these developers to the end goal of super intelligence. And that's the framing of his points. The data will only add some improvements to the current types of Ai we have. So the question he posed is... 'What will?'

[@higher2254]

They already thought about "Synthetic Data" :)

[@CharlieMcNugets]

You're wrong. Humans wll continue to explore and discover new things and this will become new data. That will never end.

[@carsonmccallum4810]

happy AGI day everyone!

[@KitLaughlin]

For real world synthetic data sets available now, just open any newspaper, or turn on TV news.

[@chuckpeabody6851]

The conversation is an example of how science works. We are imagining the advances required beyond today's DL/NLP/Transformer tools. We think, we hypothesize, we experiment and we gain new knowledge. We need to remember -- we cannot schedule technology breakthroughs...

[@javebjorkman]

I don’t know what I’m talking about, but here it is: I think we have to get away from the transformer model or whatever that’s requiring all this amount of datasets and energy to power these models. I think real innovation is gonna come from a learning model that’s similar to human learning-a little bit of info that is extrapolated to different possibilities. Instead of this exponential energy suck and requirement for more datasets

[@ddoongie1982]

That's what they tried to do in the 80s with symbolic reasoning using programming languages. We didn't have tensors, the algorithms and the hardware to do it in any meaningful way at a cost-effective price. Transformers are a hacky way to hallucinate reasoning and relatively easy to train because it's unsupervised. The other way they're trying now is to train the models with synthetic geometric mathematical data to build it's internal reasoning a bit more. Also self-selection and chain-of-thought are really powerful in generating better responses. There may be a pathway there to generate the synthetic data required to train future models as well as the chance to generate zero-shot responses that are spot on. They're expensive currently but I think smaller models are the future and bigger models are just meant for enterprise/business solutions with massive corpus of internal data. For lone users, models with less than 10B parameters should be enough with additional features stacked on top of it. Even so, SOCs are coming out of the fabs that are only meant for AI so maybe instead of GPUS, we have NPUs

[@marcusedvalson]

We’ve exhausted static training data, yes, but here are things to think about… 1. Static training data is all the data we’ve bothered to write down, and can be fed to AIs as training data. 2. This does not even scratch the surface of what we know as humans. It is only what we’ve been able to form in to words. For example, the taste of honey, we know this taste, but capturing it in words is futile. 3. The training of agents will involve data in the form of behavior. AIs observing our behavior as a form of training. This is how humans train other humans, via observation. 4.Synthetic data will be created from existing data, and will not radically expand the AI knowledge base, it will only sharpen what is there.

[@letMeSayThatInIrish]

We're never running out of fresh data. The sensors of the Large Hadron Collider can generate a petabyte per second. We can set up cameras all over the world to generate arbitrary petabytes of data about human behaviour, animals, birds, plants. We can gather data from space telescopes, microscopes (set them up in ponds, in the sea, on the forest floor, wherever), and feed the data directly to AI. Data is easy to come by.

[@cbk666]

But who is going to check the validation of the data. All data available online so far are himan made. Not ai. Mostly. If we try to create huge new information, who is going to process and validate it??

[@geoattoronto]

Learning by observation and also identification is a good point.

[@pubdigitalix]

Really interesting thoughts. Point 2 is undeniable true but point 4 maybe is wrong. Synthetic data can be generated from noise for example and spark a lot of information in the sense of statistical meaning. I think more in thermodynamics aspects of information than philosophical one. Maybe this kind of information is useless to humans but valuable to information systems in some form.

[@pubdigitalix]

@@letMeSayThatInIrish Data is not the same of information. Put instruments to observe the environment is what he say in point 3. But the AI still need to make sense of all this data to transform it in information.

[@superfliping]

Thank you for this update I have been waiting on this for quite some time

[@CharlotteLopez-n3i]

Mind blown by Ilya's vision for AI's future. Agentic systems that reason, act autonomously, and generalize? Sounds like sci-fi, but I'm excited to see it become reality. Synthetic data's potential is huge.

[@SimonNgai-d3u]

7:30 It seems like test time training is gonna be a breakthrough in which they can make a 8B score 6x% in ARC-AGI like just the shift for o1.

[@GrowStackAi]

AI: The only thing that works 24/7 without coffee ☕

[@SynLabs]

It seems that training on large datasets could potentially continue for quite a long time, as we are able to think qualitatively thanks to the multitude of perception systems we possess. We process information based on various layers of data: vision, sounds, bodily sensations, and digital inputs. This gives us a significant advantage. I believe that if we integrate these layers, there is still an immense amount of information in our world, which could form the basis for a new wave of deep learning. Take music, for example. It is said that listening to music helps develop a child’s brain. Perhaps the reason lies in the fact that music teaches us to understand harmony, and this knowledge can be transferred to solving problems in life. Music is not an isolated experience; it provides us with tools to build a deeper understanding of the world. Interestingly, music is not only about emotions but also about mathematics. Its structure consists of frequencies, rhythms, and proportions that form the foundation of harmony. It bridges the rational and the emotional. This unique quality of music could open entirely new dimensions for AI development. If AI could learn to perceive the world through such "harmony," it might hold the key to the next breakthrough. There is an immense amount of music in the world, and as a resource for learning, it remains largely untapped. And this is just one layer that could be integrated, alongside emotions, vision-both internal and external-sensations, and more. Isn’t all of this precisely what deep learning and its potential are about?

[@electroncommerce]

IoT (including IoB) is a real-world data partner for LLMs. More sensors means a never ending fresh data stream. Useful data partnership for running a real world digital twin (perhaps rendered using Unreal), with agents for everyone.

[@Victor-kd9dh]

That's why I really believe there is no data bottle neck it's a question of exploiting more sources of data

[@BeeRich33]

@@Victor-kd9dhdata generation will be an accepted standard across the board. We have been trained to know that AI training takes large datasets, proper hardware and energy, and time. Also a model that can pick up on realtime data submission.

[@georgemoller2206]

I'm just glad that in your world there are never any technical glitches.

[@oonaonoff4878]

I imagine our A.I. overlords will keep humanity around post-singularity, in order to create high quality data to feed to future iterations of itself, preventing model collapse. A.I.’s ravenous need for unique data supports the third possibility outlined in Nick Bostrom’s simulation argument. Humans might be a data farm for an invisible architect. It’s possible we are unknowingly, artificial intelligence ourselves, locked in an ancestor simulation, creating data loads for an advanced machine civilization.

[@sd5853]

Shhh dont say that outloud

[@cbk666]

Dope

[@SUPRIYACHAKRABORTY-tk4zb]

Creatures capable of simulating a level of intelligence will aspire to benefit from an either engendered equivalent or lower intelligence data?

[@TheKidmcsydney]

So uhhh... you actually don't have any idea what the Singularity, you quoted is ..... do you? lol

[@lance9749]

We will be their pets basically

[@marko_z_bogdanca]

The open internet may have been scanned, but a lot of things are gated behind paywalls. So it is clearly not true that data is over. Also, a lot of data is within us.

[@AIsavesDemocracy]

This is how the simulation starts, with synthetic data. Simulations = synthetic data which may scale infinitely, far past reality.

[@efexzium]

Hallucinations bro

[@TQ2andDebbieDo]

Was Einstein creating synthetic data when he did his famous thought experiments? Is the search for ways of making synthetic data, the search to make AI independently creative and able to make progress?

[@shirowolff9147]

​@@TQ2andDebbieDoa thought is a simulation, just like calculating things are, you dont use the calculation itself but its results

[@vitorgeraldes7054]

I see new data every second and any video camera can see it too!

[@TraderRoWi]

I think he is building this system so he can get an answer to a personal question "should i shave my head?"

[@MilesBellas]

"The Hominidae, whose members are known as the great apes or hominids, are a taxonomic family of primates that includes eight extant species in four genera: Pongo; Gorilla; Pan; and Homo, of which only modern humans remain. Numerous revisions in classifying the great apes have caused the use of the term hominid to change over time"

[@picksalot1]

"Ants have one of the largest brain-to-body mass ratios among animals. Their brains can comprise about 6% of their body mass, with some species, like Brachymyrmex, having a ratio as high as 1:8 or 12% of body mass. In comparison, humans have a brain-to-body ratio of approximately 1:40. This high ratio in ants allows them to perform complex tasks despite their small size, highlighting their remarkable collective intelligence." Perplexity AI

[@cacogenicist]

The power usage is largely about the hardware architecture. There are some neuromorphic, non-Von Neumann computing architectures being explored, which can be far more energy efficient (but have their own drawbacks).

[@Dodomba]

brain is like origami … could be small but have big area size …. also neurons are not the same … human glia cells makes mice much more intelligent if they survive the integration… if humans build better glia cells alone could make us much smarter … also develop more then 9 month before birth will us alow to become much smarter … i think humans have much more potential then ai and real AGI will know this

[@oystercatcher943]

Notice words like "will". Rather than we hope, we will aim. Nothing is guaranteed especially when you admit you don't know how to do it yet!!

[@MrWizardGG]

Because its not related to his hope or his aims, it's logic

[@georgemoller2206]

@@MrWizardGG No, it is another thing - wishful thinking.

[@MrWizardGG]

@georgemoller2206 nah biology shows it's based on evidence

[@jayeifler8812]

Amazing how they talk to us sometimes. There's just a lot of money involved so it's okay. But really everyone already knew it was about the data Ilya unless you're dense to it. Good to hear compute has improved. The data is everything. The world model of tracking all objects on the planet shouldn't be far off. They can't tell us obviously. So constant flows of data streams, data markets, simulated data markets, it's the future. It is no longer just surveillance it's now needed training data (and yes we'll all have cameras/microphones in every nook of our homes even and love it). And folks yes Ai control grid is scary for how it can go wrong, but when it goes right it solves problems and brings about true civilization and humanity.

[@HadrianAibe]

"we have but 1 internet"

[@BCCBiz-dc5tg]

Yep In-tra-net is the internal "internet" within an organisation

[@tomirkpl]

It is funny that someone asked Ilya about hallucinations, while he hallucinated all the time predicting the future of AI :D

[@PhilB-yl4jk]

Agentic active inference learning!

[@gridvid]

Data is the new Gold 😊

[@user-tx9zg5mz5p]

You can't have AI without gold, so 😂

[@KasperSOlesen]

Ahh, this is what I was hoping the new advances was hinting at. If new models are capable of test time compute and test time training.... then you should really only need some well made core model, and the rest is something the model adds to itself as needed. Should be possible to pretty much have modular AIs, that adds some language to its context that it needs to translate or maybe understand the prompts. Then the same for programming languages and so on.

[@cacogenicist]

Agents don't necessarily set their own goals -- at least not their terminal goals; humans supply those (hopefully 😊). Agents are able to take action on the world and accomplish goals. They can identify sub-goals necessary to accomplish the high-level goal. Ilya doesn't know the difference between hominids and hominins. All the great apes, including humans, are hominids -- 3 orangutan species, 2 gorilla species, chimps, bonobos, and humans, all hominids. The lineage leading to Homo, after the last common ancestor of chimps/bonobos and humans, are *_hominins_* ... so modern humans are also hominins (and hominids -- you don't evolve out of a clade). Chimps/bonobos are hominids and panins, but not hominins.

[@sodiqayilara1519]

Nice video. Keep it original and we the needs will make you rich.

[@JustAThought01]

If you need more training data,ask the user home questions and train on those answers.

[@TheMrCougarful]

Pretraining ends once they have what they want, an artificial mind based on the total of human progress as recorded in text, audio and video.

[@russcontact]

And the pressures to advance AI at all costs is enormous. It’s hard to see this not going sideways given the current world we live in. And there’s no chance we have enough time to change the world before the AI intelligence explosion. Here we go.

[@xxxValobasha]

这个视频真是让我大开眼界，太有创意了！

[@oystercatcher943]

Synthetic data has major issues. one aspect is data augmentation where you, say, move or rotate images, or change colours. It helps because you impose human knowledge about structure in the data but its limited. You can't create genuinely new and/or accurate data.

[@nexcreatives]

well we kind of can, if you think mathematically, e.g. a country of x many sq metres tends to have a ratio of y many people, if we create a "new" country of x square meters, based on previous predictions it should have an amount of people which respects the previous data. this way new data is created, which is aligned/kinda accurate with previously used data. it's not entirely accurate of course but it is new and also intuitive

[@nexcreatives]

basically the ai finds patterns in the data, extrapolates the patterns and uses that to create new data which fits the line of best fit.

[@benfidar]

Hype? Put a bunch of terms on the screen. Define them to an audience. Wait for applause.

[@sekritskwirl6106]

"I will turn your face to alabaster...". GenX, u finish the verse.

[@Music_vibes-kw7xr]

What's this assumption that all data have been exhausted? From what I know and see everyday the internet is getting bigger in exponential rates. New content is created every second on tic toc, Instagram, youtube, blogs, articles, news etc.

[@DrBreadstick]

There is also test time compute.

[@web3cryptos]

"Pre-training is only over if you have no imagination" -Logan

[@TheAiGrid]

Logank from google?

[@web3cryptos]

@TheAiGrid yes

[@CODM100x]

Check his latest tweet​@@TheAiGrid

[@futurama45345]

Logan who? Wolverine?

[@web3cryptos]

@@futurama45345 Logan Kilpatrick the former lead developer at OpenAi, now working with Google- senior product manager, Gemini API and Ai studio.

[@sampotts9666]

Synthetic data will end up being a lie, a front for data that's been illegally obtained and subsequently anonymized before undergoing tokenization.

[@seanivore]

4:12 lol right like everyone just throws agent in these days and you’re like .. uhh…. What?! This is not agentic. Glad he cleared that up a bit hahaha

[@shashankpathak3511]

I think people who talk about imminent AGI either have forgotten to do any real world AI engineering or are believing that AGI will come soon so that it helps them in some financial way: Ilya, Musk or Jensen are all too invested in AI to be truly having an objective opinion. Meanwhile for rest of us, AI remains a superbly helpful enabler just like my bicycle is (wrt me walking/running) to go to point B from A.

[@harshithpatte7169]

Quantum AI 🙏

[@panzerofthelake4460]

6:55 looks more like intervention in the nature of things. Alien gene editing? God? Because the trend only seems to have happened with ONE SINGLE species and no other (that we know of)!!! It's either a really rare occasion, and maybe it's bound to happen again, or there's more to it, one could think. If we put together all other parts of the puzzle the picture becomes clearer.

[@DanaOredson]

It was funny when you came up with a Boeing part failure rate.

[@AkaBigWurm77]

Agents feel like its just GPT running as a cron job. Synthetic data is just more transformed organic data, I guess its a way to reverse engineer otherwise copywritten training data.

[@stacypiccolo]

I wonder if AI will be able to teach humans to be self-aware. Many of us really need it!😅

[@Barrel_Of_Lube]

hmm i feel loras might be the key here where they can be trained on the fly with the latest context -> create a context graph -> merge and dissolve in the main model over time.

[@prodromosregalides3402]

That thing about the end of data might be somewhat wrong. There are tons of data for future intelligence from sensory equipment: visual , olfactory , auditory, touch . If knowing all the encyclopedias of the world was mandatory for intelligence , we simply would not be here. Our AIs should have sensory input as we have. No-one can be absolutely certain , but I think, with this, AIs will acquire the rest of our qualities. Self-awareness, emotions, purpose not suggested but natural and spontaneous. So, I kind of agree with their approach, but disagree with their method. So , they will try to create whole new worlds to generate synthetic data? Meanwhile wasting huge compute , instead of building simple sensors to get the data they need . This is counterintuitive and nonsensical. Mixing with AI may be a concern for some, but they should have this concern when it was appropriate. Now it is too late, it will happen and will not ask if it should happen. But we insist on being preoccupied with what we deem as important , until we find ourselves confined in a cage with the AIs throwing bananas at us and looking at us sympathetically as we do towards our cute animal friends in the zoo. Again ,we do our wars, our commercial wars , religious and political conflicts and we think that all this about AI, is just science fiction. When we realize that our "serious" preoccupations was just fiction it will be too late. And most probably it is already too late.

[@bumpinator]

I've had thoughts similar to this about the supposed limits of data. So I don't think you're entirely off base about this and I like how you phrased it. As for the second part maybe we should be worried about AI but I just like that there is stuff changing that's interesting in this world.

[@maclif62]

In response to Ilya Sutskever about the possibility of no new data from the Internet. This can be overcome by artificial intelligence looking at the world through microphones and security cameras, and being able to draw information and conclusions from it, that is, from life. As soon as we find a way to connect cameras and microphones to people.. In fact, we are already there. AI glasses in Apple, Amazon, Google and Facebook [Meta]

[@tonamewith]

You didn't exhaust all data. Are you kidding. You just 'scraped' some online data. What about all the books out there that are not online? And what about pictures, and videos? Doesn't that contain information? Nonetheless true that you can get it smarter with what you already have. But that the data is exhausted? Nuance.

[@davidlocontes3564]

Imagine what SuperIntelligence-fueled Killer Drones would do to the bad guys. Or the good ones, depending on your perspective. This species is so fucked.

[@CharlieMcNugets]

Not sure why he would say that Internet data will run out. Humans will ocntinue to do research and post vast amounts of data. That's never going to end.

[@dr.mikeybee]

Did Ilya read my article? The Human Knowledge Frontier: Why Large Language Models Are Hitting a Plateau.

[@geor664]

Synthetic Data: Then synthetic data generated from synthetic data. A photocopy of a photo copy of photo copy until the data is unrecognisable junk.

[@honestlocksmith5428]

One analog is teaching AI to hit a baseball. Rather than throwing balls in the real world, every possible pitch could be simulated. This is far more efficient.

[@asaforg]

This is what happens when computer scientists don't study evolution. human got a bigger brain due to language and social communication. AI has billion of social interactions every day we just don't utilize them. We don't need more data we just need different data, and we need to treat it differently, that is make ai social - but this will have to change the whole training model. Transformers simply won't do.

[@secondaccounte1209]

Flawed method .. with all the mental problems and propaganda driving unnatural human behavior, I can't see increased interaction being beneficial for acquiring data if the models are greatly flawed... My opinion? We're backing ourselves into a corner we won't be able to exit out of without catastrophic collateral damage

[@asaforg]

@secondaccounte1209 you completely missed the point. It doesn't matter if there are mental issues and propaganda, it's the communication and having to understand and react in the real world in real time that develops cognitive abilities. Given, an emobied intelligence is needed, and we probably don't have the technology, but that's what we need.

[@claudioavendano7292]

Synthetic data is like Platonic forms, things that are not bet could be, they also considered the realm of intelligence.

[@samvirtuel7583]

Right order = Is selt aware -> undestands -> reasons Otherwise it means that babies and animals are not conscious -> it is inconceivable

[@vincent_hall]

Correction: data is not fossil fuel, it will continue to accrue, just that its growth rate is far slower than the improvements in AI. Or more accurately, the models won't be able to just rely on every greater volumes of data. They'll have to get more intelligent, not just knowledgeable. Knowledge and intelligence are different. We have the first and some of the second. We want much more of a second.

[@TylorVetor]

I have found the unlocked code …. This is very powerful

[@seanivore]

Seems like no real synthetic data until Agentic data. But also what about like reaching 5G and edge computing and how that will lead to AR/VR glasses we actually want to wear and phones with basically compute power of super computers. Imagine the amount of data that will generate. The bigger issue seems like quality data. How do we separate the junk.

[@calvingrondahl1011]

Al… a better human than I.

[@TheULC100]

When you here "Evolution created" = God Created :)

[@Comenta-san]

Pretraining vill end.. why vill it end?

[@detective_h_for_hidden]

It never worked to begin with. Current AI is super flawed. It's just regurgitation but now they are realizing that regurgitation won't work to solve real-life problems

[@billymellon9481]

Sum Hallucinations are actually just deeper insights

[@blengi]

isn't there in practice 2 internets the western one and china's locked down firewalled version, where westerners don't have easy access to data?

[@rastafabui]

I am no expert, but besides synthetic data there is proprietary corporate and government data, with plenty of opportunities to come from real world data, to unlock more real world applications for robotic applications of neural networks.

[@chad0x]

Ilya really needs to finish paying for the hair plugs

[@khariclemens]

Notice he didn’t mention Sam Alterman as one of his previous collaborators. 🤔🤔

[@toadkiller4475]

I can’t exactly put my finger on it, perhaps it is his perpetual use of vocal fry when speaking, but I despise Sam Altman. I usually rely on intuition and have very consistently proven to be a good judge of character. I often see past facade instantly and make accurate judgments about a persons personality, beliefs, values, and motives even from just a short interaction.

[@tzardelasuerte]

Sam is a Ceo not a researcher or scientist

[@gopherchucksgamingnstuff2263]

Knowledgeable and intellect are not the same thing. Current AI is quite Knowledgeable, but the intellect is still young. Need a command prompt instead of making a self-conscious choice.

[@Plazman]

Adding AI-generated subtitles is low effort and doesn't enhance anything, especially when they're wrong.

[@youAbajajTube]

1:55 the data is not growing but the knowledge you extract from it is growing wirh scale....its embedded in the layers, links and weights. Not a good argument Are they bummed with colossus and trying to divert investor attention away from scaling, ie XAi?

[@TylorVetor]

It can already do all of this !!!! 😏

[@ThierAI]

What about collecting data from nature?

[@NYKIKE]

I think the widow chip is the solution to the synthetic data generation.

[@timber8403]

Willow chip

[@NYKIKE]

@@timber8403 i stand corrected. i must have misheard

[@oystercatcher943]

It isn't

[@NYKIKE]

@@oystercatcher943 explain

[@oystercatcher943]

@@NYKIKE fair enough. the capabilities of this chip are incredibly limited. basically to highly restricted algorithms that cannot do any useful computing at all. while quantum computers theoretically have huge promise many experts believe they may never do more than provide answers to simple physics experiments and never have anything like the flexibility of a general purpose computer. so using the willow chip or anything like it to generate data may never be possible or is many decades away. Even then whats it doing? anything that can be encoded in a conventional computer is better done there and would easily be sufficient to generate data. but even then it may well not be good data. kind of like how movie sequels are always worse 😂

[@Outcast100]

Tesla has tons of real life data😅 ( the best kind of data )

[@petratilling2521]

Do you really want to OWN a conscious self aware being?

[@efexzium]

Keep going

[@Julianninha]

I want it to own me

[@drawnhere]

If AI becomes self aware, then utilizing them would be slavery.

[@TheULC100]

"Just imagine"... is not reality

[@user-tx9zg5mz5p]

Steven Mark Ryan 🤔

[@geoattoronto]

So data is limited by how good our data is. Reasoning is limited by how good our ‘first principles’ are. And Ilya is limited by his assumption that human reasoning is a function of the human brain. He does not realize that our intuition is generated by the human spirit.

[@tadmarshall2739]

The talk about the incredible speed of progress seems to fly in the face of GM giving up on self-driving cars. I saw a video about Tesla crashes and it's clear that Elon's vehicles fail at the basic task of keeping the passengers alive. One fatal crash involved an overturned tractor-trailer in the driver's lane and the car "couldn't identify the object" and so drove into it at full speed. Overpromise and underdeliver seems to be the rule for today.

[@SmirkInvestigator]

✋I thought engrams were the way. 🤦‍♂️

[@ivor000]

okay, while i understand, and appreciate, the hat-tip to biology figuring out how to break the ratio mold (body to brain size), to somehow imply that because biology figured "a way out", AI engineers will also be able to make a similar "jump" is a stretch, sorry. I'd much rather this claim sit on top of statements like "we'll figure it out, cause why wouldn't we? we have up until now" or even better: "were quite confident at this point that AI will solve this problem itself"

[@RichardWarner-df7dz]

Ahhh but XI also uses visual learning.

[@keviniverson2835]

They will go willingly😢 1:06

[@TheBiffsterLife]

Didn’t gain undemanding, instead lowered opinion of Ilya

[@DaveJMorten]

Good content. But why has Ilya got such a pedestrian designed presentation?

[@wands82sg]

It sounds like Skynet?

[@Felipe-zl1rj]

Basically, nothing new. I would think Ilya would have something to show for himself after that amount of VC money

[@naomirain]

Copyright actions are limiting data availability... the data is unattainable... at the moment

[@augmentos]

He is half right, but his hominid idea is fundamentally incorrect for many reasons Which surprises me to use that example

[@silberlinie]

I think that's why what Ilya Sutskever said was very enlightening, not because of the content of the speech. What overwhelmed me was his style and the presentation of his performance. We know from the spiritual realm the way our preachers and priests speak to us in church and chapels. That's what Ilya did. Including, and very importantly, the echo of his voice like in a large cathedral. Extremely astonishing. You feel led to quickly sink to your knees in humility, bow your head and squint forward in shame and fear. Brilliant, Ilya Sutskever. That's what happens when you want to conjure up a new religion out of thin air. Thumbs up.

[@johnnymetonymic]

Did the dude get s rash?

[@luttman23]

Nah that's his hair

[@makeitinstrumentalist]

Take this with a MASSIVE pinch of salt. He is just saying things that people have been dreaming about for a long time now without backing it up with any evidence Like literally he just said "why not?" to a couple of questions. Yeah, I'm really going to believe you on all is this, yep. I'd love for them to prove me wrong, but I just highly doubt any of this is going to happen any time soon

[@antooant]

In what event did Ilya say these things?

[@LuisBorges0]

Hallucinations, or bullshit, is a symptom of creative and intelligence. Don't forget, okay? We created AI like us and now we don't like them because they are like us 😂😂😂😂