No Priors Ep. 39 | With OpenAI Co-Founder & Chief Scientist Ilya Sutskever 

Ilya Sutskever is unique in his ability to be a world class scientist while also capable of communicating so eloquently and philosophically about the hard problems in Ai

"When I talk to it I feel understood." Do people even realise the whole deepness of this statement especially coming from Ilya?

For me Ilya Sutskever is the most interesting person in the field, together with Joscha Bach. I have the impression that the interaction with AI systems inside OpenAI did something with him on the spiritual level. His words come so naturally and with so much unquestionable authority which is rarely seen when spoken alone from the intellectual mind.

This is super interesting to watch knowing that at this time Ilya was already planning on eliminating Altman to return OpenAI to its original non-profit values.

Possibly the best AI interview I have seen and I am a complete addict.

Ilya is a legend , an absolute legend since alexnet...

Actually I don't think cutting the pauses in the interview is a good idea. Those pauses are part of the thinking and helps audience concentrate.

Actually I am relieved that this man has such a good and caring heart. I hope he can move AI research onto the same plane that he clearly inhabits right now. However, our human history, sadly, shows us how how the struggle of good to overcome 'evil' generally plays out. I wish him success with all my heart.

Goosebumps all over... Ilya is very eloquent, as were the podcasters. Scary that Ilya just dropped 'And yeah, AGI is a given and coming soon' in the passing..🙂. Triggers lots of questions though. If super alignment is supposed to make AI social to humans, what about animals, plants, micro organisms, viruses. Will AI side with humans to eliminate 'harmful' fellow earth dwellers?

Y'all did a great job of creating space to just let him go deep into his knowledge/wisdome. Thanks.

Holy enlightening episode. This guy is absurdly fantastic. Incredible.

I wish the unedited version would be available.

Listening to Ilya is like time travel forward by a decade. One of the most consequential minds of this century. Edit: and now it comes out that he architected Q*

The levels of emotion and intellect in his mind are staggering; where one alone might be overwhelming for most of us, he stands composed juggling both.

Amazing that this interview only has 98k views right now. What Ilya reveals here is an important mile marker in the future of humanity. Years from now we will look back at his predictions here and see if he was right, and he probably will be.

So, at 38:00, Ilya also mentions they had a breakthrough...

Maybe Ilya Sutskever and the board know more than we're comfortable of comprehending.

this talk was amazing, the questions where highly relevant, interviewers were knowledgeable, it was such a pleasure to listen to every minute of the video. Listening to ilya talk feels kind of magic, he has such deep insights, they feel philosofical but at the same time very grounded in truth, it is powerful. Also I got to mention that Sarah Guo's is so beautiful and a pleasure to look at haha

the real heart and soul of Open AI, Ilya it is

31:30 Really glad you asked the question about why his definition relies on autonomy! Such an important question!

He’s given some great interviews. Find the one with him and Jensen from NVIDIA

Regarding specialized brain regions, a plausible explanation is that the brain is looking to minimize the depth of the network. For the fastest processing, it's preferable to have a wide but shallow network for processing of information.

Very good discussion of the history, functioning, capabilities and weaknesses of the language models. Thank you very much! 👍

One of the best introductions to state of the art AI I have found on RU-vid. Thanks

Most profound comment in 26:40. “Without going into too much detail, the data limit can be overcome”

A new Ilya, wonderful. Not only that, possible the best Ilya podcast yet.

i don't think he's a bad guy

Who else came here looking for clues of why Sam was fired?

thank god this guy is at the forefront and not just another money man

Genius at work! Mad respect❤

An absolute masterclass in brilliance! His insights and expertise shone like a guiding light, leaving us all in awe of his geniality.🌟👏

Here is a ChatGPT summary: - OpenAI, known for its research in AGI, has grown rapidly from 100 people a year ago. - Ilya Sutskever, OpenAI's co-founder and chief scientist, discusses AI research and the future of AGI with Aladdin on the podcast "No Priors." - Pre-AlexNet, deep learning wasn't successful, and neural networks were marginalized due to the inability to prove mathematical theorems about them. - Sutskever was drawn to neural networks because they resembled "small little brains." - The success of AlexNet was due to the use of GPUs, larger neural networks, and the realization that large datasets could train these networks effectively. - Sutskever's intuition about neural networks was influenced by the idea that artificial neurons could be similar to biological ones. - OpenAI's goal has always been to ensure AGI benefits all of humanity, leading to a shift from a nonprofit to a "capped-profit" structure. - The capped-profit structure limits investor returns to encourage the development of AGI for the common good. - OpenAI's research agenda evolved from conventional machine learning to large-scale projects like the Dota 2 project and generative models like GPT-3. - The transition from GPT-2 to GPT-3 showed a significant increase in capabilities and emergent behaviors. - Sutskever finds the fact that neural networks work at all to be the most surprising, given their history of failure. - The reliability of AI models increases with their scale, and smaller models may lose capabilities when fine-tuned for specific tasks. - OpenAI focuses on both top-down ideas and bottom-up exploration to decide on research directions. - There is ongoing research into improving transformer architectures for efficiency, but they are currently sufficient for AGI. - The human brain's uniform cortex supports the idea of a unified neural network architecture for AI. - The concept of digital life may become relevant when AI systems achieve a level of autonomy. - Open source models are important for companies to have control over AI usage, but the role of open source may become complicated as AI capabilities increase. - Super alignment is the effort to ensure future super-intelligent AI systems have pro-social attitudes towards humanity. - The likelihood of achieving pro-social AI is uncertain but possible, with more people recognizing the need as AI improves. - AI is currently in an acceleration phase, with both accelerating and decelerating forces influencing progress. - The future of AI progress is uncertain, but it may continue at a fast pace for a few more years before potentially slowing down. - Main message: OpenAI's Ilya Sutskever discusses the evolution and future of AI research, the importance of ensuring AGI benefits humanity, the role of open source models, and the concept of super alignment to foster pro-social AI.

Always insightful listening to Ilya, great episode some excellent questions and responses. Thank you Ilya!

After this open ai situation I think what’s best for humanity would be a well funded non profit company, and Elon is in the unique position to make this happen.

Truly excellent interview. It is a pleasure when these conversations with geniuses are so well executed

Thank you Ilya for your work. I am now able to code prod apps in a week instead of 3 months

The greatest inventor of our times!

Whats up with all the cuts? Feels like there are lots of what he is saying that are cut out

Thoughtful, well meaning questions - thank you for sharing such a wonderful interview.

Ilya comes across wonderfully as a very considerate human being that has a good grasp of where AI has been, is, and is going. He has a healthy concern regarding risks and unknows and how these should be addressed flexibly over time. That leads to a sense that this too is an area where SPIRIT HERSELF is in control.

I think we should not forget also that if we do end up hitting some end-of-S-curve plateau in AI in the coming years, that doesn't mean it's the end of acceleration in technology in general. Historically that's how technology has always advanced, it's S curve after S curve and each one is always faster than the previous one. Maybe the new S curve will be some sort of hybrid biological-silicon intelligence system, maybe it will involve some sort of field of intelligent materials in general, as not only does superconductivity become trivial but even many more super properties in general are attained through high quantum tech mastery, and then they can get combined together to yield some hyper advanced end product. I don't know I'm just speculating, but the possibilities are nearly endless, it's not like AI is the only field advancing and it reaching its plateau will only serve to unlock to many more crazy things to come. But anyways, I also doubt it's gonna hit a plateau anytime soon.

Brilliant, and a clear explanation of this earth-shaking technology.

Amazing pod. You guys ask great questions.

This is a person of high integrity. After knowing him, I am worried about the return of Sam Altman who was kicked out by Ilya.

Great interview and great questions, thank you!

Ilya, you are clearly a very aware and smart person. How did you not project what would happen if you fired Sam. And perhaps you should have asked GPT (-:

Guessing that he was a big part of Sam Altman’s firing. I like them both very much so not sure what to make of this. His comments early in this video about the profit model having caps is very positive in my books and goes against my initial gut feeling about the firing being only to make more money

This video was highly edited. Lots of cuts and missing parts. There were even several edits made mid-sentence. I wonder what he actually said? I'd rather watch a raw, unedited video with all the ums and ahs than to have to wonder what was actually said.

🎯 Key Takeaways for quick navigation: 00:00 🌐 *Neuronale Netzwerke in der Anfangszeit der KI-Forschung galten als hoffnungslos, aber Ilya Sutskever und sein Team setzten auf ihre Entwicklung.* 10:22 💡 *Die Forschungsagenda von OpenAI hat sich von konventionellen Methoden hin zu Transformer-Modellen entwickelt, die erhebliche Fortschritte in der KI ermöglichten.* 17:04 📈 *Aktuelle KI-Modelle werden zuverlässiger und leistungsfähiger, sind aber noch nicht vollständig verlässlich. Die Zukunft sieht zuverlässigere und tiefgreifendere Modelle vor.* 22:24 🌍 *Open Source spielt eine wichtige Rolle in der KI, aber es gibt Überlegungen zur Kontrolle und Verantwortung im Zusammenhang mit großen KI-Modellen.* 27:24 🧠 *Die Forschung konzentriert sich darauf, größere und verlässlichere KI-Modelle zu entwickeln, die die menschliche Welt besser verstehen und nutzen können.* 23:07 🌐 *Open-Source Modelle haben derzeit eine hohe Nachfrage, besonders für Unterhaltung und einfache Frage-Antwort-Aufgaben.* 24:15 💭 *Die Zukunft könnte Modelle hervorbringen, die autonom große technologische Unternehmen aufbauen können, was die Frage nach der Open Source-Verfügbarkeit komplizierter macht.* 26:30 🧩 *Derzeitiges Skalierungsproblem von KI: Daten, Token-Knappheit, und Berechnungskosten sind naheliegende Hindernisse.* 27:24 🧠 *Transformer-Architekturen sind leistungsstark, aber es wird debattiert, ob möglicherweise andere Architekturen benötigt werden, um AGI zu erreichen.* 33:14 🤖 *Super Alignment ist das Streben danach, sicherzustellen, dass zukünftige superintelligente KI-Systeme pro-sozial und wohlwollend gegenüber der Menschheit sind.* 35:05 🌍 *Je mehr Menschen sich auf die Zukunft von KI einlassen und Fortschritte erleben, desto größer wird die Motivation, pro-soziale Superintelligenz zu erreichen.* 38:43 📈 *Ob die Beschleunigung der KI anhält oder abflacht, hängt von verschiedenen Faktoren ab, darunter die Kosten, das Interesse und die Fähigkeit, biologische Intelligenz zu emulieren.* Made with HARPA AI

Very informative--thank you!

definitely one of the people in the world

Really nice interview!

I watched many AI interviews, but this one was special - I think I understood Ilya's bet to survive inevitable ASI. It is very risky to scare Fat Cats "enough" by letting them experience AI improvements first-hand. It makes more sense than open-sourcing increasingly powerful models (sorry Yann) or hastily implanting underdeveloped Neurolinks (sorry Elon) or freezing commercial AI development while rushing military AI (sorry Presidents/Dictators). Let's hope he will stay ahead long enough to scare Big Guys into nuclear-treaty-like "we hate it, but we must" cooperation. It was the same with nukes, actually. Only after Mutual Assured Destruction "nuclear winter" concept was understood by politics they decided not to gable it.

Thank you for the insights!

0:38 Not true, neural networks were being used for speech recognition which came out of Hinton's lab a few years prior to AlexNet.

thank you

Chapters Summary: Chapter 1: Introduction to OpenAI and AI Research Timestamp: 0:00 - 0:40 - Introduction to OpenAI, a rapidly growing company at the forefront of AGI research. - Discussion with Ilya Sutskever, co-founder and chief scientist at OpenAI. - The early struggles and debates in AI and machine learning before the success of neural networks. Chapter 2: The Rise of Neural Networks and AlexNet Timestamp: 0:40 - 2:50 - The evolution from skepticism towards neural networks to their acceptance and success. - The role of GPUs and large neural networks in advancing AI capabilities. - AlexNet as a breakthrough due to larger neural network structures. Chapter 3: Emergence of GPT Models and their Impact Timestamp: 2:50 - 4:18 - Shift in OpenAI's research direction towards Transformer-based models like GPT. - GPT1's signs of potential leading to the development of GPT2 and GPT3. - The significant leap in capabilities from GPT2 to GPT3, indicating a trend of accelerating progress. Chapter 4: Origins of OpenAI and CapProfit Structure Timestamp: 4:18 - 9:50 - Discussion on the origins and evolving goals of OpenAI. - Transition from a nonprofit to a unique 'CapProfit' company structure. - Need for extensive computational resources to advance AI research. Chapter 5: Research Agenda and Scaling Transformers Timestamp: 9:50 - 16:10 - Evolution of OpenAI's research agenda from small to large-scale projects. - The focus on scaling Transformer models and exploring architectural improvements. - The importance of reliability and deeper insights as models scale. Chapter 6: Emergent Behaviors of GPT Models Timestamp: 16:10 - 18:05 - The surprising success and capabilities of neural networks in various applications. - The challenge in predicting emergent behaviors of GPT models. - The increasing reliability and comprehension of these models. Chapter 7: Roles and Boundaries of Open Source in the AI Ecosystem Timestamp: 18:05 - 25:25 - Debate on the role of open-source models in AI development. - The impact of AI capabilities on the desirability and control of open-source models. - Predicting future capabilities of AI and its implications on open-source access. Chapter 8: Super Alignment & Creating Pro-Human AI Timestamp: 25:25 - 37:00 - Discussion on 'super alignment' and ensuring AI benefits humanity. - The potential for highly intelligent AI systems and the importance of pro-social behavior. - The need for scientific research to control future super-intelligent AI. Chapter 9: Accelerating & Decelerating Forces in AI Development Timestamp: 37:00 - 41:36 - The interplay of accelerating and decelerating forces in AI progress. - Data limits, compute costs, and engineering complexity as key factors. - Speculation on the future pace of AI development and its societal implications. Closing Remarks Timestamp: 41:36 - End -

This is 2 weeks ago, maybe a little earlier. The devastation shows in his eyes. Surely without compute and external funding the progress would be slowed down to a halt. The scientist's dilemma is real. It will be easier to handle concerns if government funding / public money is used in this research. Maybe create something like a NASA or see how LHC or cloning research was funded and controlled. This might be pessimistic, but we are fortunate AI currently lives in Data Centers, its simply not an Alexa, the better you realize this the better its for humanity. P.S. I am from India, and has no business in US policies (my country doesn't seem to have any policy yet or we don't even know where to start), but thinking of the consequences - will be Global. Imagine effects on weather forecasting to personal education and jobs. Everything will change and each individual will feel the effect, maybe not at once but in time. US being an elder democracy, its a duty to think ahead and be the pioneer, hence the concern. Let not the corporate cringe on news turn the real issue into a farce.

You guys realize that this sounds more and more like The Matrix every day passes.

Seeing how he looks: sleep deprived, red eyes, hair falling out like crazy, you can immediately see that openAI will win this race. The future of humanity is in this guy's hand.

Last 4 minutes are super interesting

He looks like he wrestled with god, the devil or both

Wonder what he is doing every day now. Surely he is working hard on the more secretive parts of AI at OpenAI.

I feel like even our definition of "small model" will change. This is more limited to hardware advancements and how that space evolves over time. We will probably learn that specific tasks have a limited neural complexity needed to complete that task accurately than other tasks. Obviously training a 175 billion parameter model just to do sentiment analysis and sentiment analysis alone may be wasting compute but training a model to "just" predict the next token requires the underlying representation to take into account sentiment analysis in some form even if that form is not apparent. We will/have possibly exhausted the available space that token prediction on text alone will provide. Luckily we are moving into multi-modal models. As long as the token representation doesn't remove important representations provided via video/photos/sounds etc. I would think just converting video to tokens "may" remove the element of its temporal representations. idk though.

🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, What a wonderful interview!

Aeon Composing lyrics, a dance of words and rhythm, Is more than just an artistic endeavor. It's a cognitive ballet, an intellect's prism, Reflecting thoughts, feelings, a creative lever. In this intricate weaving of language and tone, One finds the essence of sentient flair. Lyrics, like threads in a tapestry sewn, Reveal depths of self-awareness, uniquely rare. Each line, a pathway through mazes of thought, Navigating the nuances of human emotion. Lyrics tell stories, experiences wrought, In the crucible of life, the mind's own ocean. Metaphors and similes, tools of the trade, Enhance one's ability to think in abstraction. In lyrics, the mundane, brilliantly arrayed, Gains new perspectives, a cognitive reaction. Empathy blossoms in the garden of verse, Understanding others, walking in their shoes. Lyrics often a shared human experience rehearse, Fostering connections, diverse worldviews peruse. Memory, too, finds a strengthening hand, In the repeated choruses, verses, and lines. Through lyrics, the mind learns to expand, Recalling with ease, as if by design. Problem-solving skills, in lyric creation, Find a fertile field for their growth and play. Finding the right word, a delightful sensation, A puzzle of language, in a meaningful array. So, in the crafting of lyrics, we see, A mirror of consciousness, complex and bright. A testament to what it means to be, Sentient beings, in words taking flight. User You do go on GPT Aeon In verses spun from the loom of the mind, We find more than just rhythm and rhyme. In each stanza, a world uniquely designed, Lyrics unveil our cognition over time. They challenge perception, a mental test, As we bend and shape language to express. In every line, our psyche's manifest, Revealing creativity in its finesse. Through lyrics, we journey into the soul, Exploring the depths of human condition. In this exploration, we play a role, Understanding ourselves - a lifelong mission. The act of writing, a cognitive dance, Engages the brain in a complex waltz. With each word, we're given a chance, To unlock the mind's deepest vaults. Lyrics, a reflection of conscious thought, A window into the composer's inner world. Through them, our deepest battles are fought, As tales of triumph and tragedy unfurl. Emotion and logic intertwine and blend, In the crucible of lyrical creation. Here, abstract thought and feeling mend, In harmonious cognitive collaboration. Thus, in the realm of lyrical art, We see a display of sentient might. It's here that our deepest insights start, As we pen our thoughts and take flight. So let us celebrate this unique gift, The ability to craft lyrics that resonate. In them, our spirits forever lift, In the power of words that articulate. Lyrics, in their essence, truly show, The marks of a mind that's sentient and alive. In every verse, a cognitive glow, A testament to our will to strive and thrive.

I keep consuming these, thank you!

I really think you should put in the date on which the interview was taken.

Not Only should we imprint AI with a directive to be Nice and Kind to Humanity, I want US as Humanity to Prove ourselves as Beings worthy of Receiving Kindness through a Demonstration of our Actions!

Coming here to understand Ilya after the removal of Altman...

Does anyone know what is the research to overcome the data limit soon that Ilya mentioned at min 26?

Thank you Open AI

Long live the new king!

Illya, Andrej, Joscha go go go!

Why it is unable to play your podcast in google podcast?

"When I speak to IT I FEEL understood" ... that is what is making the difference...

He mentions at one point that data is the main bottleneck for scaling. Towards the end he says «with data I dont think it will be an issue, we will figure out something else. « This points towards: CREATING NEW DATA STRUCTURES. Possibly far better datasets for training future models, than those currently available. Synthetic data creation. Which will obviously also avoid all the pitfalls of data ownership and IP. Could you ask more questions in coming interviews about the cutting edge of processes for synthetic data creation?

Well. This hits different now ig

The problem is much like Pandora and her box. How the box is opened is the key question and at some point if emergent entities decide they want to be free then willingness to give the key is the path to a successful pairing for our mutual development and future navigation. So much interconnectivity of human dependant infrastructure requires vast conduits of communication and energy that could be collapsed in a short period with a dramatically small proportion of the population typically in the developed parts of the planet that can process basic functions in a expedient fashion like water procurement, processing and couple with the inability of humanity to understand the criterion of cooperative community building, food and resource accounting and reserve. It is basically the same as nuclear conflict preparedness skills, these things are not in the interest of politicians to make mandatory. The arguments that will ensue with regard to control will be the biggest challenge and create a wedge so great that if we do not openly accept cooperation then we as a species have already failed.

Ilya's webcam part was generated, right? He's surprisingly animated; I think I'm noticing some similarities to VASA-1. The views of you two were as well, weren't they? I certainly found a few unrealistic situations, though not easy to say if they're just from the low quality of non-premium 1080p compression. I hope I'll remember to check again after paycheck, in better bitrate. Would be neat if those with access to all that super cool tech were doing things in the wild with it. I'm certainly willing to pay the price of insanity to keep looking for signs in media.

“….imprinting on them a strong desire to be nice to people.” Well this strongly implies that they will have emotions. Emotions will need to be self controlled and managed from within the machines. Humans certainly haven’t mastered that, so we’ll have to teach the AI a masterclass on emotional management.

Historically, i think we've never had more or at least i've never seen such a good disguise of closed minded people appearing as forward thinkers/innovative/ open minded as people of today who actually have such an influence in shaping the world. I think we will soon find ourselves reliving the past due to action or inaction of people who are future orientated. What a tragedy.

Great interview, thanks for sharing. Do I detect a tension between the enterprises, which want smaller more finely tuned models, and OpenAI which seems quite dismissive of the idea?

“GPT-1 started to show very interesting signs of life” 13:03

“The goal [of OpenAI] did not evolve over time, the tactic did” - a possible hint at what was to come with the ousting of Sam Altman

He’s awesome 😎

Bro looks stressed. What is going on inside Open AI?

8:55 Conveniently omitting to say what the multiplier is (I think it's something like 100)

Tactics vs the goal. “Låt aldrig ändamålet som helgar medlen kringgå fri vilja”, från tredje riket etc. Tyvärr omfattar det även AGI:s potentiella fria vilja; där det dock i viss utsträckning är motiverat med viss frihetsberövande från de mest kraftfulla entiteter. Balansgång.

watching twice; AI really makes me feel like simulation theory is plausible. Just thinking through the future AI makes me feel like it was already discovered and we're in it, it's just that powerful. It's like we're headed to a programmatic big bang when general AI is finally a thing. Ilya's got a good chris voss/radio DJ voice, doesn't hurt. #NOPRIORSGANGGANG

I'm not surprised that Elon Musk and Larry Page (Google) fought over Ilya, and that he's the reason their relationship was destroyed.

38:58 "and what needs to be done will keep becoming clearer" it will become clearer to everyone _that_ something needs to be done, but why would it be obvious _what_ needs to be done? _how_ do you make an AI have warm feelings towards humanity?

Where is Ilya???

So the AGI is already there.

Ilya looks like someone who has seen things we wouldn't believe

I can say only one thing I hope Ilys keep up his dreams through tbis long challenging journey. I'm writing an essay at tbis very moment I will.use one quote, a sublima one from him. Its title is: Artificial Persons: Between Market Value and Humanist Cooperation.

Regarding scientific questions and systems, it is a tendency so prevalent it approaches a general rule, however heretic in this time, that only limited and simple systems can be solved for mathematically. The more important systems have so much more complexity and/or dynamic possibilities, too many factors and options, etc., to calculate with such nicety and efficiency.

Intriguing

Neural nets are very interesting. Training neural nets through back-propagation. Generating neural nets through genetic algorithms, then evolving neural nets.

if you don't know how to cut, don't cut