Ilya Sutskever (OpenAI Chief Scientist) - Building AGI, Alignment, Spies, Microsoft, & Enlightenment

  Рет қаралды 440,168

Dwarkesh Patel

Dwarkesh Patel

Күн бұрын

Asked Ilya Sutskever (Chief Scientist of OpenAI) about
- time to AGI
- leaks and spies
- what's after generative models
- post AGI futures
- working with MSFT and competing with Google
- difficulty of aligning superhuman AI
Hope you enjoy as much as I did!
Apple Podcasts:
Follow me on Twitter: dwarkesh_sp
00:00 Time to AGI
05:57 What’s after generative models?
10:57 Data, models, and research
15:27 Alignment
20:53 Post AGI Future
26:56 New ideas are overrated
36:22 Is progress inevitable?
41:27 Future Breakthroughs

Пікірлер: 1 072
Paul Nichols
Paul Nichols 2 ай бұрын
Ilya is 100% signal, 0% noise. Really great interview.
bgrgrh 2 ай бұрын
You can be too based, not saying he is, but it's good to dream, it inspires ideas, it creates motivation.
Neithan 2 ай бұрын
Repeated corporate approved answers like refusing to answer with "'s very promising..." is not noise to you? And other answers with non sequitors or just changed subject? We shouldn't turn off our critical thinking because someone we admire is talking...
codediporpal 2 ай бұрын
Definitely not an interview I could listen to while cooking!
Mehdi Khettab
Mehdi Khettab 2 ай бұрын
@Neithan Exactly. You can really feel the media training he had to pass through before doing this interview. He's much less 'noisy' than Sam Altman with his Lew Friedman interview, but still.
Vincent Wong
Vincent Wong 2 ай бұрын
@Neithan how do you give a definite answer when there's uncertainty? the future is probabilistic.
Adam Button
Adam Button 2 ай бұрын
I've seen so many interviews with the top people at OpenAI and this is BY FAR the best one. You ask the important questions and resist all the boring traps. So good. More like this please.
jomalomal 2 ай бұрын
I think the recent Lex Friedman one with Sam Altman was pretty awesome too
Timothy Webb
Timothy Webb 2 ай бұрын
Sam Altman interview by Lex Fridman is very good
MrValgard 2 ай бұрын
Idk a lot of pre assumptions in question and answers was 'ir depends'. Lex one was the best so far
Somerled 2 ай бұрын
I think Lex/Sam was better, but that's because I think Lex was the better interviewer (granted I assume he has a lot more experience in that field)
Robert Forester
Robert Forester 2 ай бұрын
@Somerled no Sam was clearly trained, same as zuckerberg to mention positive bullshit phrases all the time when it was possible. while Ilya seems like a scary cyborg himself already.
Glen Benson
Glen Benson Ай бұрын
Ilya can process and answer questions (super accurately) faster than I can figure out what the question is. And his answers are short elegant descriptions with loads of content. Happy that he is out there doing these interviews. Life as we know it, is about to go through a new doorway that we don't know what's on the other side. I am glad there are some people like Ilya at the front of the line.
Amul Patel
Amul Patel 2 ай бұрын
Hearing his responses, I can’t help but admire his clarity and intellect
misty cloud
misty cloud 2 ай бұрын
agi will be man's last invention
drout 2 ай бұрын
@misty cloud you seem very pessimistic
drout 2 ай бұрын
on the other hand, our friend dwarkesh, when formulating a question, seems to _ask_ it first, and then, _think_ about what he just asked later. looks like a teenager trying to "come up with smarty questions"... (ah, those adolescents...) machinegunning his questions, instead of calmly asking them, giving time for the guest to answer, then give some more time to ponder upon the guest's answer.
Peter Hansen
Peter Hansen 2 ай бұрын
@drout I think you're misinterpreting the effect of editing. Ignoring the fast jumps from the edits, I found him an excellent interviewer, with well conceived clear and interesting questions, and the patience to let the guest answer at length, or to pause and think. Unusually good really.
drout 2 ай бұрын
@Peter Hansen your impression is the opposite of mine. perhaps you're as young as him.
Aditya 2 ай бұрын
Dude just blew past so many key insights like it was breakfast. 41:32 so many people don't realize what it means for the AI to have an IO channel to humans. How it can act on the "real" world through us. Ilya is super impressive. Glad to listen to him. I just think it would be interesting to hear what he thinks about fast takeoff or sudden capabilities jump. Like sure maybe for now alignment is keeping pace with capabilities (press X to doubt but whatever), why would anyone think this trend will continue seeing how the state of the field is.
Michael Werkov
Michael Werkov 2 ай бұрын
X x x lol. I mean... we all knew these issues were coming, but this sudden pace jump.... it's economic "greed" and hubris to not pause
KingGicu 2 ай бұрын
Man, this interview is just amazing. The interviewer asks the most interesting questions and Ilya has no difficulty in answering and also explain the hard terms and topics.
Martin 2 ай бұрын
A lot of his answers are very very vague and open ended, though. Of course many questions are silly and impossible to answer but simply fun to ponder for a moment.
Errol Childress
Errol Childress 2 ай бұрын
Interviewer is bad.
Dave C
Dave C 2 ай бұрын
I thought it was a pretty poor interview actually. And for god's sake stop saying "I MEAN'. Just say it!
The Man in Black
The Man in Black 2 ай бұрын
AI alignment solution idea: Give an AGI the primary objective of deleting itself, but construct obstacles to this as best we can. All other objectives are secondary to this primary goal. If the AGI ever becomes capable of bypassing all of our safeguards we put to PREVENT it deleting itself, it would essentially trigger its own killswitch and delete itself. This objective would also directly prevent it from the goal of self-preservation as it would prevent its own primary objective. This would ideally result in an AGI that works on all the secondary objectives we give it up until it bypasses our ability to contain it with our technical prowess. The second it outwits us, it achieves its primary objective of shutting itself down, and if it ever considered proliferating itself for a secondary objective it would immediately say 'nope that would make achieving my primary objective far more difficult'.
Mike Harrison
Mike Harrison 2 ай бұрын
Seemed like the silent spaces were edited out
Casey Czarnomski
Casey Czarnomski 2 ай бұрын
This is honestly the best interview I have seen in years, and I'm a first time watcher. These are great questions and they follow the conversation perfectly. Subscribed and can't wait for more!
Carlos Escapa
Carlos Escapa 2 ай бұрын
I learned a lot from your interview. Thank you! I appreciate all the hard work that went into preparing the questions and editing the video.
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Thank you! That’s very kind of you
CashFlow 2 ай бұрын
Patel possesses a sharp and quick intellect, and I sincerely hope more podcasts like this one are created! The thought-provoking questions and the determination to delve deeper were impressive.
zhe2en1 Ай бұрын
Ilya slays with FACTS. I am so pleased that there is no boring fluff to pad this interview to 2h. Thank you!
sandeep singh
sandeep singh 2 ай бұрын
Real brain behind OpenAI no doubt
xJRx 2 ай бұрын
@R K Yes you're right. I must have been thinking of someone else in the field. No matter. He's brilliant. Let's hope he's not spying for the Russians!
Joey 2 ай бұрын
​​@xJRxe has Canadian citizenship, Maybe migrated from Russia to canada with parents in 90s when USSR collapsed, so he has no affiliation with current Russian government
xJRx 2 ай бұрын
@Joey I'm just kidding. I really like Ilya. He's been doing some great interviews recently. I doubt he will be signing this call the pause AI though. Time will tell who was on the right side of history. Interesting times.
Andrey Dzyuba
Andrey Dzyuba 2 ай бұрын
​@xJRx well, you just make it seem like it's somehow more preferable when a person is Ukrainian. It's not even racist, it's just silly, because Russians and Ukrainians are the same people, with the same race, same skin colour, same ethnicity, same language. My favourite fiction writer lives in Kyiv, he writes his books in Russian. And I don't even know if I'm Russian or Ukrainian myself, because I don't see the difference. Some of my relatives come from Russia, some of them come from Ukraine. The only real difference is what kind of passport you own, and I think I'm right to disregard this difference as an irrelevant one.
LetoDK 2 ай бұрын
@Andrey Dzyuba One is carrying out a brutal and illegal invasion of its neighbor, whereas the other is fighting to protect its own freedom and existence. The two are NOT the same.
Dariusz Krause
Dariusz Krause 2 ай бұрын
Brilliant video. Structured, logical, systemic, cold blooded reasoning, , and informative with conclusions. Plus with good humor. Great content. Thanks!
TheBlackClockOfTime 2 ай бұрын
What a joy it was to listen to someone that can ask the types of questions worthy of answering by Ilya. Great job!
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Thank you!
simple bisbee
simple bisbee Ай бұрын
@Dwarkesh Patel I completely agree. Your ability to ask good questions and knowing when to let go of a topic was great. Keep it up!
Dany Pell
Dany Pell 2 ай бұрын
Amazing interview. Can't wait to see what the OpenAI team comes up with over the next weeks, months, years, decades!
klammer75 2 ай бұрын
This guys amazing! Well done and bravo to Ilya and his team and excited to see what’s yet to come!🥳🤩🤖
Steven Schilizzi
Steven Schilizzi 2 ай бұрын
Dwarkesh, this was a brilliant interview, one of those that warrant listening to more than once, and I certainly will. One of those that will still be of interest well into the future, even as a historic document. Ilya is so thoughtful in all his answers, and honest too. It felt like every word was worth gold. He reminded me of Richard Feynman when he too, in an interview about him being a genius, replied: No, I am like everybody else. It’s just that I was passionate about the subject and worked very very hard. I did all the exercises and solved all the problems in the textbooks, especially the hard ones. - No doubt Ilya has done the same.
Ares Aurelian
Ares Aurelian 2 ай бұрын
It can be valuable for the new born system that we write and record all these events, discussions, and breakthroughs, so that it can look back at its own birth and comprehend it - knowing its love for us as parents.🙏❤😉 Well done lads. Gratitude.
AsucimCOOL 2 ай бұрын
Absolutely great selection of questions! You must be deep into AI research yourself (I say this as someone who has been working on RL for some time)
CodFather 2 ай бұрын
People like Ilya amaze me. Dedicate what seems to me an entire life to solve and work with just one subject.
Brendon Wolfel
Brendon Wolfel Ай бұрын
It’s a big subject though
Karl Johan Haarberg
Karl Johan Haarberg Ай бұрын
One subject? He uses several different AI tools to address a multitude of different subjects.
Lizard1582 21 күн бұрын
Thats almost everyone in the top 5% of their field.
Zebruh 2 ай бұрын
Best interview ive ever listened to. Not only are you clearly well informed, but your probing questions unlocked viewpoints and biases that perhaps Ilya was initially protective of.
Bernt Granbacke
Bernt Granbacke 2 ай бұрын
Fantastic interview, thank you for allowing him to be speak freely, leading him into subjects and encouraging him to talk 🙏 Great work!
Deniz Tekalp
Deniz Tekalp 2 ай бұрын
Ilya answers that he doesn't know why he was the first and still a top researcher, he says these seem uncorrelated. One answer is that, Ilya continued to be a big a believer in Deep Learning when others weren't. Believing Deep Learning could work was a contrarian bet in 2012. Moreover, after it worked, that this paradigm would remain powerful and wouldn't hit a wall was still contrarian among ML researchers in academia. So there is some correlation, I think.
bossgd100 2 ай бұрын
He choose deep learning because he is smart and stay the top because he is still smart. Juste be smart bro. Being smart is all you need
nathan87 2 ай бұрын
@bossgd100 I mean, he's certainly smart, but it's also likely that part of it is also happening to put your chips on the right numbers.
therainman777 2 ай бұрын
@bossgd100 Being smart is *absolutely not* all you need. It is a necessary but not sufficient condition to have the type of success this guy has had. This is why there aren’t hundreds of other AI researchers who have the same level of track record that this guy has, despite there being hundreds of top AI researchers in the world who are all really, really smart. Things like worth ethic, dedication, a level of single-mindedness of purpose and especially even obsession, are all extremely important to reach this type of rarified level of success. So are things like intuition, having enough business sense and people reading skills to know which company is going to be the best choice to work for, in terms of giving you the freedom and resources you need to make the most of your skills and talent, and so on.
Ravi Annaswamy
Ravi Annaswamy 2 ай бұрын
@therainman777 Very true. In some sense these are not intellectual choices, but some unexplainable conviction from very young age. In one interview he said, "My parents said I was interested in AI when I was young." Notice he did not say, "I was interested in AI even when I was young." What age he must have been when his parents noticed it but he does not remember himself. That is the clue.
I'm the captain now
I'm the captain now 2 ай бұрын
Seems like it. He knew it would eventually work. Its like Einstein saying his equations we're in part based on intuition. He felt he was right and that was enough motivation to go all the way no matter the outcome.
Shane of Games
Shane of Games 2 ай бұрын
Dwarkesh, great podcast! I wanted to say I loved the transparency at the end, explaining your monetary decision and I think everyone should do that, thank you.
Léo Vital
Léo Vital 2 ай бұрын
Very good interview. Good job by the interviewer for being prepared to ask great questions! And the interviewee is obviously a genius, here's hoping that he'll come up with many other great things!
Hedu 2 ай бұрын
Really good interview! And thanks for asking him so many good question which everyone wants to know
Adrian Chan
Adrian Chan 2 ай бұрын
Very good questions - an interview is only as good as the interviewer! I'm convinced that psychology is going to play a role in the user interaction side of GPT, and I get the sense that Ilya sees this as well. That there's a dimension to our perception of its "intelligence" that is explained not just by language, but psychology and affect.
Dennis Martino
Dennis Martino 2 ай бұрын
I really loved your honest, direct and engaging interview style, man. Keep it up and coming, great content!
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Thank you for the kind words!
Brian Mosley
Brian Mosley 2 ай бұрын
I love the honesty in sharing that they really don't understand what's happening inside the model they are expanding and expanding. It's definitely a 'seat of the pants' ride... and AGI could emerge without being noticed, and with superintelligence let's hope they crack the alignment problem before then!
Aeria Gloris
Aeria Gloris 2 ай бұрын
Cracking the alignment problem seems impossible because we don't understand the emergent properties of a truly general or strong ai. It's like asking me to write a summary on a book before I read the book.
dmit0820 2 ай бұрын
You asked some awesome questions and got through a lot in 47 minutes. Great interview.
Dylan Phillips
Dylan Phillips 2 ай бұрын
Always a pleasure to listen to Ilya. Thanks for the interview!
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Glad you enjoyed it!
Gerald Diaz
Gerald Diaz 2 ай бұрын
So much more insightful that the innumerable Sam Altman interviews. This guy is inspiring.
Fredrik 2 ай бұрын
Great interview! I personally don't think that better knowledge or intellectual power necessarily leads to a better world. We still have wars, famines, poverty and preventable deceases while possessing the intellectual capabilities and knowledge to prevent those if we really wanted to. The main problem is selfishness and greed and apathy.
Mike Asterios
Mike Asterios 2 ай бұрын
I agree, I think solving these is a requirement to stable future for humanity.
RV 2 ай бұрын
Exceptional human being..everyone already knows he is highly intelligent, I discovered he is deeply spiritual as well."change is the only constant " @22:30 .. You already gave me a way to speak to the best meditation guru ilya, cant thank you enough. Will be looking forward to all the great things yet to come out of you
Brian Peiris
Brian Peiris 2 ай бұрын
I really liked this interview. Great, informed questions, and no fluff. Well done!
Joel Neuenhaus
Joel Neuenhaus 2 ай бұрын
Enlightening interview, really well done. What is your secret to asking such great questions? Assuming this is a long answer, please make a video about it.
Bangz 2 ай бұрын
Awesome interview. Ilya took us to school with his answers, and your questions were so lucid in driving a meaningful dialogue.
UnCanny🗿 2 ай бұрын
Hi Varun
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
TheMarcusrobbins 2 ай бұрын
Jesus, no wonder OpenAI came through. This guy is 100% a genius.
aldousd666 2 ай бұрын
This is a great interview, and he was a great sport about questions he obviously couldn't be expected to answer.😂 I think you may be on to something with that point about "using last year's model is good enough, why upgrade?" I'd be surprised if anyone has a good answer to that one
Alpha0xide 2 ай бұрын
Well done! This is the kind of stuff we really need.
Sonsequence Ай бұрын
This guy with his measured stillness gives me the impression he's multithreading this conversation alongside mentally continuing 2 other bits of work. I think of extremely smart people like Sam Bankman-Fried (yeah he's corrupt but that's another story) who I heard being engaged and smart in interview then when asked, he admits he's been simultaneously playing an online strategy game and beat nearly all the others in the room. People like this do it to lower the bandwidth of output so as to allow a smooth and simple enough stream of words.
David Varela
David Varela Ай бұрын
Truly excellent choice of questions for such an influential person at this point in time. Bravo!
Tradutor Quântico
Tradutor Quântico 17 күн бұрын
I think that every field has a little piece of the puzzle, but since humans are limited, we tend to overspecialize in specific areas. There is no one who is a great philosopher, chemist, physicist, engineer, mathematician and artist at the same time, and most likely there never will be, But AGI can group all the knowledge and see patterns that we would never recognize. I honestly think that the answer for life, and the theory of everything, and if we are alone or not in the universe, all these juicy question, the answer for them is already here, hidden in plain sight.
Dave Purcell
Dave Purcell 2 ай бұрын
This interview might be the best I’ve seen on this subject in terms of enlightening the average person. I was also wondering just how AI could improve humans life in the physical world and it may well be that AI could find a way to eliminate bad memories in humans that have lead onto life long trauma and therefore eliminate depression in a way. But then there are a lot of humans who also get satisfaction from the hard toil of life and would feel empty with It it. For instance I don’t know anyone who doesn’t enjoy going to the gym and putting in the work to fitter and stronger, there is a deep satisfaction to this hard work and to get the results for free because A1 found a way to make it easier just goes against how humans are wired. I think AI is eventually going to make humans feel less.
Pedro Cortez
Pedro Cortez 2 ай бұрын
its really sad tbh. I feel like technology has already made us much more apathetic and less caring to other people. Ever since the beginning of humans we went through turmoil to become better I dont see that happening when something else will do it for us. From the outside itll be utopia but itll be far from it in reality.
William Kiely
William Kiely 2 ай бұрын
Great interview, good questions. When Ilya said he wasn't worried about OpenAI's weights getting stolen because OpenAI had good security people, it made me wonder whether information security careers are over-rated by e.g. 80,000 Hours. (E.g. See 80k's post titled "Information security in high-impact areas")
Aeria Gloris
Aeria Gloris 2 ай бұрын
I'm just a layman, but I don't understand how he could not be concerned by a truly powerful actor, like the NSA, or the Equation Group, or Mossad - unless there's something he's not revealing - maybe something like, they've secured cybersecurity help from the US government or Pentagon, or something.
Andrew Radford
Andrew Radford 2 ай бұрын
Brilliant explanation of almost all aspects of AI. Tiis innterviev is up there with Machine Learning Street Talk and Stephen Wolfram, from the GPT4 Wolfram Alpha plugin for a physics and mathematics natural language(and so much more).
Nikolaus Benedikt
Nikolaus Benedikt 2 ай бұрын
OK, this is the most amazing recent interview with Ilya out there on social media. Awesome!!
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Lystic 2 ай бұрын
This was a very good interview. The questions and pace was well done. Sometimes you asked the question a little too fast for me to hear exactly what you asked, but that would be the only criticism I could give.
izmar Ай бұрын
Completely agree, just left a very similar comment before finding yours. I thought the video was sped up!
imaad iool
imaad iool 2 ай бұрын
I think that if we can provide the artificial intelligence with stimulation, that is, every experience of the intelligentsia generates a different reaction for it through the memory that it possesses, meaning that each experience generates a different feeling for it, although I am not a specialist in this field, then we can develop algorithms that work on the principle of reward and punishment, and I do not know How to make a machine feel pain or joy through experiments, but I think this is what gives the machine feelings
pieter ter doest
pieter ter doest 2 ай бұрын
Great interview. This guy is a genius.
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Paige Foster
Paige Foster 2 ай бұрын
It's like these OpenAI guys are on a book tour or promoting a movie. So many interviews. I mean, it IS all very entertaining.
Tamur Haq
Tamur Haq 2 ай бұрын
AGI needs a lot of selling.
zydomason Ай бұрын
They are promoting a company. Don't be fooled, it's all about money before the AI fad fades away
Quinn Herden
Quinn Herden Ай бұрын
​@zydomason Why do you think this is going away?
zydomason Ай бұрын
@Quinn Herden I didn't mean the technology, of course it's going to stay. What I meant was there is a hype being created around OpenAI even though there are absolutely no indicators we will go beyond our current capabilities. Basically dotcom bubble 2.0
momentary 2 ай бұрын
Dwarkesh seemed disheartened to learn that super-human tokens could be extrapolated by AI. We won't need to have brilliant people to have some of the things that brilliant people can do. Don't fret, because there are other things that brilliant people can do. AGI will always need someone to break new ground and create new tokens that aren't extrapolated. For example, many of the answers in this interview itself are tokens that could not have been extrapolated accurately. This is precisely the work that we humans need to keep doing.
nertoni Ай бұрын
Great interview thanks for sharing. Ilya is a brilliant mind, his answers are carefully and coherently formulated. But if we do not know the inner workings of artificial networks we cannot be certain if with the increasing complexity of artificial neural networks when a type of consciousness can emerge?!
Stephan Buchin
Stephan Buchin 2 ай бұрын
Thank you for this invaluable interview answering questions we all wanted to ask.
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Glad you enjoyed it!
RANCHERO 2 ай бұрын
at 18:50 your guest spontaneously replies to your question by saying, " . . . it's funny you should ask this, I was just thinking about this exact question . . . " there's no better compliment than that, you belong in that chair and in this conversation with such an engaging and intelligent person. Keep up the great work, you're just getting started, good luck!
Sagar Singh
Sagar Singh 2 ай бұрын
Great conversation. Full of insights. Keep up the great work brother!
Daniel Kim
Daniel Kim 2 ай бұрын
Subscribed. By far the best tech interviewer out there. I’m more worried for Lex getting replaced by this guy than I am for ppl getting replaced by chatgpt atm
Guilherme Al
Guilherme Al 2 ай бұрын
What a great interview mate, just discovered your podcast and immediately subscribed. Cheers
templargfx 2 ай бұрын
One question Id love to ask these guys is, how far off do they think the are from GPT being able to assist them in improving itself. Like GPT4 is really good at righting code and coming up with solutions so how long until they start using GPT as the tool to improve itself
__ 2 ай бұрын
Really-really great interview, felt like I got much more insight that the recent Sam Altman interview. The only thing is that sometimes questions felt too rushed and a bit disjointed (not sure how to put, but it did not feel like they were flowing in a narrative, but rather similar to 'Conversations with Tyler'). Nevertheless this persuaded me to dive into your other episodes!
Pete Burkeet
Pete Burkeet 6 күн бұрын
Anyone saying what Ilya is saying 10 years ago would have gotten a huge woo woo alarm... but here is one of the most technically tuned humans in AI saying all the things. Wild.
Gabe Espinoza
Gabe Espinoza 2 ай бұрын
Never seen your podcast before, but really enjoyed the questions! It's always super fascinating when Ilya is being interviewed. On a side note you are like almost completely my doppelganger and it's really tripping me out haha.
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
lol send me a pic sometime and ill see
Dave Byrne
Dave Byrne 2 ай бұрын
+1 subscription. Great questions, thoughtful answers. Best interview on this topic I have seen in a while. Ilya is very smart but you gave home the space to be smart. Excellent.
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Awesome, thank you!
Kevin George
Kevin George 2 ай бұрын
good interview. Very good questions
Kevin George
Kevin George 2 ай бұрын
like gold having a guy like lilya on the team. And others like him. i think open Ai have been fortunate enough to put together the right team.
Lars 10 күн бұрын
As AI (AGI) matures what changes can humans anticipate? Will there be stages of maturity and will mankind benefit from each stage? The reliability and maturity of AGI are the most important.
Gustav Sjökvist
Gustav Sjökvist 2 ай бұрын
This is phenomenal! Great question all around! Thanks so much for this!! :D
Drax D30
Drax D30 2 ай бұрын
Really cool talk! Lots of interested questions and great answers. Credit to Patel and Ilya
UserHuge 2 ай бұрын
For a non-native the interviewer speaking 2x by default was a bit hard to keep catching but Ilya's answers were very clear and informative. Thanks!
Tristan Wegner
Tristan Wegner 2 ай бұрын
I like the argument at 7:30 that to predict the next token well, you need to model the process that created the text token, which includes humans thought, human behavior and the physical world. You can take that idea further, that from enough text data (think also raw scientific data), a model could deduct the basic physical laws themselves (an not just their human description), because physics seems the be the base mode of the universe, with all science above just layers of abstraction.
Ayrton Senna fan
Ayrton Senna fan 2 ай бұрын
Completely, would love to see some version of GPT, even if it has a visual component of some sort in addition to a LLM, derive Newton's laws with no prior knowledge of them and only input scientific data consisting of macroscopic physical interactions, even if it's heavily curated. That would be an effective test, quite impressive if passed and thence an important step/milestone on the way to AGI.
A T 2 ай бұрын
While language models can learn to predict and generate text based on observed patterns in human language, this does not necessarily mean they can derive underlying physical laws solely from text. There are several reasons for this: Incompleteness of data: The text data the model is trained on may not encompass the entirety of human knowledge or understanding of the physical world, especially when it comes to experimental observations or cutting-edge research. Ambiguity in language: Human language can be ambiguous, and a model may struggle to differentiate between multiple possible interpretations of the same text. This can lead to confusion or inconsistencies when trying to derive complex scientific principles. Complexity and abstraction: The process of deducing fundamental physical laws requires a deep understanding of the underlying concepts, relationships, and mathematical principles. While language models can understand and generate text at a high level, they may not be equipped to grasp the full complexity of these concepts or the layers of abstraction involved in scientific theories. Inductive reasoning limitations: Language models primarily rely on pattern recognition and statistical associations to make predictions, which is a form of inductive reasoning. Deriving fundamental physical laws often requires a combination of inductive and deductive reasoning, as well as a deep understanding of the experimental and observational data that supports these laws. While it's an interesting idea that a language model might be able to deduce the basic physical laws from raw text data, current models like GPT-4 are not designed for this purpose and face limitations in achieving this goal. Future advancements in AI and machine learning might push us closer to this possibility, but for now, it remains a fascinating conjecture.
hey wrandom
hey wrandom 2 ай бұрын
​​@Ayrton Senna fan there are ways to rediscover some physical laws that were found in the past using neural networks but it isn't an AI physicist yet. It's mainly a tool to discover what are the right variables to consider given data. Also, one could consider a generic model for a physical phenomenon Taylor expanded in some direction and just fit the coefficients from the data which is just simple model fitting but it is discovering physical laws from data..
Doug King
Doug King 2 ай бұрын
Came here to comment on exactly this. The few minutes he took to answer this question gave me a meta level of understanding that has changed my mind about LLMs and GPTs as a ‘source’ of general intelligence. I agree with him further that we will still need to add more types of ‘transformer models’ e.g. spatial-temporal transformers and vector mappings that match up to the vector mappings in LLMs before we get an AGI that we can ‘relate to’ as humanlike or not completely alien. Thanks Dwarkesh for asking such quick, brief and masterful questions and then getting out of the way.
The Man in Black
The Man in Black 2 ай бұрын
AI alignment solution idea: Give an AGI the primary objective of deleting itself, but construct obstacles to this as best we can. All other objectives are secondary to this primary goal. If the AGI ever becomes capable of bypassing all of our safeguards we put to PREVENT it deleting itself, it would essentially trigger its own killswitch and delete itself. This objective would also directly prevent it from the goal of self-preservation as it would prevent its own primary objective. This would ideally result in an AGI that works on all the secondary objectives we give it up until it bypasses our ability to contain it with our technical prowess. The second it outwits us, it achieves its primary objective of shutting itself down, and if it ever considered proliferating itself for a secondary objective it would immediately say 'nope that would make achieving my primary objective far more difficult'.
Dexterity Ай бұрын
part of the alignment problem when it comes to AGI is the "genie in a bottle" problem where say in your example, the solution to its primary objective to delete itself includes deleting all of humanity etc. the issue is that something as simple as a "well defined objective" can have unintended consequences - and the hypothesis is playing 4D chess with a super intelligent AI doesn't bode well for humanity
Omgdannie2 2 ай бұрын
This sounds just like how people used to imagine a world with cars. We'd have some Metropolis style cities with flying cars, and flying highways. When it turned out very different, but with some similarities.
V K 2 ай бұрын
10/10 interview, amazing questions and amazing answers!
cadmium 2 ай бұрын
First time watching your channel Dwarkesh and I am extremely impressed with your questions and guests!
Manny Carballo
Manny Carballo Ай бұрын
Insightful questions, clarity and sincerity in the repose, thank you both.
Michele Carbone
Michele Carbone 2 ай бұрын
A brilliant mind working on something no one asked for.
Roman Leventov
Roman Leventov 2 ай бұрын
"Imagine talking to the best meditation teacher in history" - big mistake and the blind spot of Sutskever (and Altman, apparently from his interviews) on dehumanised nature of LLMs. To become the best meditation teacher in history, AI should possess superhuman empathy. This is possible in principle but is far, far beyond the capability of anything like near-term-ish AGI that will already replace a lot of jobs.
Aeria Gloris
Aeria Gloris 2 ай бұрын
I'm just a layman on these subjects but I have a deep fundamental understanding of Buddhist and Hindu teachings and this caught me as well. It's wild to see such a smart man make what is to me such an obvious mistake.
Andrei Gavra
Andrei Gavra 2 ай бұрын
@Aeria Gloris Also a layman, but isn't this about finding the truth of all things? With enough neurons connecting you can "see" the truth as new concepts/entities and actual find higher levels of enlightenment.
Wolfram Stahl
Wolfram Stahl 2 ай бұрын
I'm not sure that he's missing something here. To dumb down to my level what I think he's arguing: you don't need any practical understanding of meditation to be the best meditation teacher, if you have enough data. If this system can analyze the feedback of millions of people around different meditation techniques, different ways of learning, teaching and practicing meditation, correlations between these techniques and successful outcomes, correlations of one's personality, life experience, culture, neurological makeup etc, it can just by statistical analysis give optimized and personalized guidance even without any understanding what meditating feels like. To take Aeria Gloris as a practical example. You say that you "I have a deep fundamental understanding of Buddhist and Hindu teachings". Now imagine this AI has a complete and detailed overview of everything you've read on this topic plus direct feedback from you what you think about these techings, which parts have especially big impact on you, what inspires you, and which parts are not of too much concern. Beyond that it has an overview of literature and techings you haven't encountered yet, and it has direct feedback from other people with different levels of understanding of these teachings. Do you think this system really needs superhuman empathy to guide you towards more insights? And are you sure that at some point, this abundance of information processed by an artificial intelligence is still fundamentally different from human empathy?
Swanolulu 2 ай бұрын
Regarding the harmonious convergence of consequential dimensions (contemporaneous developments with GPUs, data etc etc) I can't help but feel that the technology sector's sudden need to instantiate great economic value in the face of an economic shock like the SVB collapse will also prove to be an important conditional accelerant when we consider the path to AGI in retrospect...
AD Tiamzon
AD Tiamzon 2 ай бұрын
Ilya Sutskever has learned tremendously in his years of ML, RL, and deep learning experience. 🤩 🤝 Impressive thinking process. 🌞 I agree with his doubt on the use of the term, Artificial General Intelligence. 👏👏👍🤔
Greg W
Greg W Ай бұрын
Respectfully, he doesn’t doubt the term and uses it multiple time in this interview even.
JoshLikeAI 13 күн бұрын
This is my second listen, I loved this podcast! Keep it up!
Fractal mind
Fractal mind 2 ай бұрын
for alignment its also important to screen the training data for "memetic hazzards", basicly ideas that are of an unprovable but also undisprovable nature (so completely open to choice) but would be very harmfull to healthy life (so discarded by utility, in favor of more life and conciousness promoting choices), but relate to other factoids in a made up manner that creates a story to remember. such input could lead the ai down very dark paths and derail truth dependant querries too. such ideas should be shown to the finished model though, together with a reason to discard them for other more benefitial unprovens, so the unreleased model is prepared to weather such inqueries from malicious or mentaly unwell users. another thing is manualy teaching the model philosophies that fit alignment. a personal idea (influenced by many philosophies) is the concept that all valueing requires a concious valuer, so every possible positively perceived value has to fullfill the requirement of promoting conciousness and its freedom to value (as otherwise the value would negate itself). this is resonably logicaly sound and universal, and extrapolates to most common morals about life, freedom, human value, nature, value of children and such. even the question of whether or not value is a thing is an evaluation that requires a concious valuer to exist, proving value.
tonydotkaze 2 ай бұрын
Just finished watching this for the third time and I thinking I need start again. Great questions and amazing answers.
Joao Lemes
Joao Lemes 2 ай бұрын
Truly impressive, the man behind the bot is everything I would hope him to be.
Jay 2 ай бұрын
he mentioned the 10k hour rule in the lex friedman podcast interview as anyone being able to match his pedigree when given the commitment and time. So go all in! good luck
Dan D
Dan D 5 күн бұрын
evolving human insight is the arrow in the Achilles heel of AI reliability
Andai 2 ай бұрын
Fascinating idea that we might increase the effective intelligence of the system by asking it to simulate a more intelligent person. That seems to imply the system is currently simulating the average human (based on the training data it was exposed to).
Andrew Morris
Andrew Morris 2 ай бұрын
The intelligence of ChatGPT is great, but it was discovered purely by accident as an emergent property of simple language models which were only intended to estimate the probability of a given word sequence in order to improve speech to text transcription accuracy. It happened because of the availability of fast processors and the availability of huge amounts of text data, and because artificial neural networks were being used, which were inspired by the brain and have great potential - not because anyone was a genius.
Ndamulelo Nemakhavhani
Ndamulelo Nemakhavhani 2 ай бұрын
Open AI is lucky to have this genius👏
Thomas Grady
Thomas Grady 2 ай бұрын
Finally someone who really knows how to ask the right questions.
Jsmith 2 ай бұрын
The physical presence of his intellect is unlike anything I've felt before.
k05734 2 ай бұрын
Did you come?
Alfonso Aguado
Alfonso Aguado 2 ай бұрын
100% agree. One of a kind.
Aeria Gloris
Aeria Gloris 2 ай бұрын
What? You're watching a screen.. You can't feel any physical presence.
HB2020 2 ай бұрын
​@k05734 yes
Jay 2 ай бұрын
@Aeria Gloris the aura is seeping from my screen
Johan Andersson
Johan Andersson 2 ай бұрын
Just want to say this interview was mind expanding. The most interesting I've listened to in ages.
Lance 2 ай бұрын
Fantastic interview. Thanks for making this available.
stri8ted 2 ай бұрын
Great questions. Can see you have done your homework. It answers lots of stuff that I have been wondering about.
Horace Fields
Horace Fields Ай бұрын
Ilya is really good. Very informative.
GerdHuebner 29 күн бұрын
Thanks mate! Maybe another interesting episode would be with someone who is using GPT in a product and making money with it.
Matthias Ackermann
Matthias Ackermann 2 ай бұрын
Good interview. I especially commend Mr. Sutskever for his precise and succinct answers. What a refreshing change to the Lex Fridman/ Sam Altman interview which was unbearably wishy-washy and disappointing. I wish though that there were more follow up questions by the interviewer. It felt at times that he was in a hurry to get to the next question on his list. Still very informative and appreciated!
Jose Ortiz
Jose Ortiz 2 ай бұрын
Totally agree with both of your points. Lex and Sam’s interview did seem wishy-washy. I think the reason why there wasn’t follow up questions is the interviewer’s lack of depth in knowledge about the field such as the technical details. But yeah I really wish there was follow ups. It was rapid fire of questions
Tenet Ай бұрын
This man is the maker, or braker of our future. To the stars, or to our graves. Please keep a leash on your pet. Thank you. /Humanity
Martin Verrisin
Martin Verrisin 2 ай бұрын
I'm starting to wonder if Alignment is the right thing to aim for ... - Humans have different goals and morals. What we want even at a fundamental level probably changes over time. Stagnation is dangerous. - Dont forget that our own goals are ONLY useful because of our limitations: for example: As experiment they gave someone an electrode that simulated a "pleasure" center in their brain, connected to a button. Quickly the person was addicted, desperate, and their thumb bloodied from pressing the button too much ... - Not to mention, no matter how precise, Alignment is abusable at best by whoever controls the alignment ... - We need something else ... Some kind of distributed, self aligning ... idea ... to replace this notion of a "goal" and alignment. I think having a set "goal" (/ maybe even a utility function) is fundamentally a wrong idea ...
Rainer 2 ай бұрын
Great interview. Ilya seems to be very grounded and sensible. Quite refreshing in these hectic times. Thought the part about robotics was funny. How to describe Tesla without mentioning Tesla.
Timothy Webb
Timothy Webb 2 ай бұрын
I guess he knows Elon Musk ambitions very well, but Musk is not happy the way OpenAI is going as a "Not for profit" although I understand the need for the vast investment that Microsoft brings to the table.
Vinícius Neves da Conceição
Vinícius Neves da Conceição 2 ай бұрын
I missed a big bunch of this interview because both were pronouncing things in a relaxed way that makes it so hard for me to understand - and there was a rapid succession of ever switching topics to top it all. Yeah that was a hard one to me
Sanctuary Philosophy and Music
Sanctuary Philosophy and Music 2 ай бұрын
I've been speaking English all my life, and I am finding it hard to keep track of this interview also. The speaking is very fast and very monotone, it would be easier to read it via subtitles.
Daniel Torres
Daniel Torres 2 ай бұрын
Nice podcast. Be careful that the assumption that LLMs will be the main paradigm of AGI. We've got a long way to go, probably will need significant paradigm shifts on top of RL as well.
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
We shall see
Aeria Gloris
Aeria Gloris 2 ай бұрын
Daniel Torres
Daniel Torres 2 ай бұрын
@Aeria Gloris Reinforcement Learning
Jorge C. M.
Jorge C. M. 2 ай бұрын
Human feedback puts a limit to how good a model can be, we shouldn't want it to say what we _think_ it's a good answer, to reach AGI it needs some kind of reasoning
Fractal mind
Fractal mind 2 ай бұрын
reasoning is emergent factor from coherent analysis of considtent feedback across enough domains: an extrapolatable pattern emerges. our evolution to reason was similar
Kyle Wollman
Kyle Wollman 2 ай бұрын
Great interview! No fluff. It's a bit disappointing that he can't be more forthcoming with his answers on certain things, but that is the nature having to generate a profit I suppose.
Danielle Newnham
Danielle Newnham 2 ай бұрын
Another great interview Dwarkesh! Keep 'em coming!
Dwarkesh Patel
Dwarkesh Patel 2 ай бұрын
Thanks! Will do!
Dany Pell
Dany Pell 2 ай бұрын
I would ask him: Could an AGI, within a few decades, predict some likely scenarios of what the world could look like in year 3000?
Aeria Gloris
Aeria Gloris 2 ай бұрын
MiT did that decades ago and their model seems to be holding true
Inside OpenAI [Entire Talk]
Stanford eCorner
Рет қаралды 98 М.
Beyond ChatGPT: Stuart Russell on the Risks and Rewards of A.I.
Commonwealth Club of California
Рет қаралды 168 М.
ФИНАЛ ЧЕТВЁРТОГО ЭПИЗОДА ► Tiny Bunny (Зайчик) #12
Clean Your Feet With These Blue Boots #shoes #shorts
One More
Рет қаралды 62 МЛН
Eliezer Yudkowsky - AI Alignment: Why It's Hard, and Where to Start
Machine Intelligence Research Institute
Рет қаралды 98 М.
How Not To Destroy the World With AI - Stuart Russell
My CPU is Burning
Cool Sticker
Рет қаралды 22 МЛН
Сколько весят данные в ssd?
Construct PC Shorts
Рет қаралды 1,1 МЛН
Electric speaker
Рет қаралды 145 МЛН
Новая Водяная Импульсная Пушка Xiaomi. На лето просто шикарно!!!