In 2021, OpenAI’s Sam Altman said at the National Security Commission on AI that ‘we are on the cusp of major changes, which are capable of an incredibly bad outcome.’
takeaways: • [00:04] Introduction of the speaker, Craig Smith, and his guest, Ilya Sutskever, co-founder and chief scientist of OpenAI and primary mind behind GPT-3 and ChatGPT. • [01:00] Sutskever's background and interest in AI and consciousness. • [02:30] Sutskever's early start in machine learning and working with Jeff Hinton at the University of Toronto. • [03:45] Sutskever's realization about training large neural networks on big enough data sets to solve complicated tasks. • [06:33] The breakthroughs in convolutional neural networks and how they led to the imagenet competition. • [08:36] OpenAI's exploration of the idea that predicting the next thing is all you need for unsupervised learning. • [10:24] The development of GPT-3 and the importance of scaling in deep learning. • [11:42] The importance of scaling something specific in deep learning and the potential for discovering new twists on scaling. • At 13:01, the speaker discusses how scaling matters and that even small changes can have a big impact. • At 13:46, the speaker talks about the limitations of large language models, explaining that their knowledge is contained in the language they are trained on, and that they lack an underlying understanding of reality. • At 14:32, the speaker comments on the difficulty of talking about the limits of language models and how they change over time. • At 15:13, the speaker argues that learning statistical regularities is a big deal and can lead to a better understanding of the world. • At 16:33, the speaker talks about the limitations of language models and their propensity to hallucinate, but expresses hope that this issue can be addressed through reinforcement learning from human feedback. • At 17:52, the speaker discusses how teaching neural nets through interaction with humans can help improve their outputs and reduce hallucinations. • At 21:44, the speaker comments on Jana Kun's work on joint embedding predictive architectures, and expresses the belief that multimodal understanding is desirable, but not necessary for language models to learn about the world. • High dimensional vectors with uncertainty are a challenge for prediction, but Auto-regressive Transformers can handle them (26:02) • Auto-regressive Transformers work well on images (26:02) • Large language models learn compressed representations of the real world processes that produce data (29:40) • The goal is to make language models more reliable, controllable, and faster to learn from less data (33:44) • Learning more from less data is possible with creative ideas (35:51) • The cost of faster processors for training language models may be justified if the benefits outweigh the cost (37:48) • [25:28] The paper makes a claim that predicting high-dimensional distributions is a major challenge and requires a particular approach, but the current autoregressive transformers can already deal with this. • [26:02] Autoregressive transformers work perfectly on images and can generate images in a complicated and subtle way, with the help of supervised representation learning. • [27:09] The vector used to represent pixels is like a string of text, and turning everything into language is essentially what is happening. • [29:40] Large generative models learn compressed representations of the real-world processes that produce the data they are trained on, including knowledge about people, their thoughts, feelings, conditions, and interactions. • [31:31] Human teachers are needed to guide the reinforcement learning process of a pre-trained model to achieve a high level of reliability and desired behavior, but they also use AI assistance to increase their efficiency. • [35:10] It is possible to learn more from less data, and there is an opportunity to teach AI models skills that are missing and convey to them our desires and preferences more easily. • [39:57] In the future, it could be desirable to have some kind of democratic process where citizens provide information to neural nets about how they want things to be. • [41:15] It is probably impossible to understand everything in a complicated situation, even for AI systems, and there will always be a choice to focus on the most important variables.
Thank you for all the hard work, everyone who do their best for these new systems to be implemented with the least possible disruption to human societies. We are still humans, and we must go from the perspective of love - to the future and beyond. Much gratitude.
This is an incredible and valuable interview. I can't believe this depth of knowledge is under 6k subs. I think that's a very scary thing, people is not aware. Thanks you so much for sharing it with us, for free ♥
I think the future is bright for the world because the influx of Hispanics and Africans into the world of technology will propel the advancement of science beyond what is capable by people of European origin.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
None of them understand anything, because they believe human conscience is a product of some algorithms. Good luck adopting that view and reducing yourself to a machine.
Thank you for uploading. I learned so much detailed nuances about LLM from this interview. I really like Ilya's way of communicating subtle but important points.
On the subject of hallucinations, I think they are more clearly explained by the problem space that the engine is trying to navigate. When having no relevant information on the subject, but it is still asked (one might say compelled) to say something, whatever it says must be either off-topic or false. And I believe Ilya is very insightful when he says the language of psychology is starting to describe these systems, because we have hallucinations, too. Whatever compels us to output something when indeed lacking skill or knowledge about a subject also affects GPT systems as well. When do people hallucinate or ramble? When they have no imposed limits/feedback, like a dictator or celebrity that is never told they are wrong or some guy living all alone in the wild or a child that has not been educated yet. Or a Twitter user. With social creatures it is the meaningful interaction with other social creatures (and the physical world) that generates these limits. Which I find promising and fascinating, because it means that the supervised learning step Ilya is talking about can also be performed by other AIs, not particularly humans. The brain is also composed of two hemispheres that keep each other in balance. Very interesting indeed.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
@@katehamilton7240 They have no such inability, they are used to discover new drugs, they play Go like nobody else did in the past, and you can ask it logical puzzle that it didn't see before, and sometimes it gets it right, especially when using tree of logic.
Q at 26:40 A: when he says Vector he means like a vector in physics like it has force and direction on multiple planes. When converting something into a vector embedding. It is like to convert an image into an idea so it is behaves like a concept that is stored spatially relative and near by to other ideas. Then you can convert it back. but Also you can use its spatial position in multidimensional space to find related information. also you can put it back from an embedding which is a vector representation of something back to original representation while preserving relatedness positional information. a text sentence it is a list of embeddings, it is an array of "vectors". When you put it back from an array of vectors into a sentence of words. You also get all of the learned associations and the related things about the sentence in addition to the sentence. There is a new thing in text search engines. Vector databases. It enables to search things based on ideas. It is fascinating you can search in any language and get the same results.
A vector is not an array. Vector is more like a single word. Converted into a spatial representation. Currently there are embeddings on syllables. So a part of a word has an idea related to it
Algolia is one of them. I tried to sign up for a demo but have no idea what I'm doing. But it has an option to create indexes and upload your data sets for free whatever that means. I'm a completely clueless person who now has an Algolia account. That I can't do anything with.
- His intention was specific: to make a very small but real contribution to ai. ( in the time that people were certain computers can't learn, 2003!) - Auto regressive transformer is a very powerful tool that researchers underestimate. - "humans can be summerize in sequence", do you remember Devs miniserie!? - "To predict well, to summarize data well, you meed to understand more and more how the world that produced the data." - "maybe we are reaching a point where the language of psychology can be appropriate to understand these artificial neural networks!" - he doesn't believe these models don't have any real understanding of the nature of our world! - "human teachers are using ai assistance, and they are so efficient." By human teachers, he means people working on reinforcement learning from human feedback. - "make models more reliable, more controlable, make them learn faster, with less data and less instructions. Make them halucinate less. How far are they in the future? These are topics he intrested in and work on them right now!" The interesting thing is in OpenAI, he can't talk specifically about what he is working on, the open in opanAI annoy me a little! - "The costs are high, but the question is, does paying this cost actually generate something useful? Does what we get after paying the costs outweigh the costs?
Yes, the openAI name is very misleading. I understand that these guys did much better job than deep mind and meta for LLMs. I also get that all sort of state, and corporate interests want to replicate the thing. But it is more like hermetically sealed AI. A few days ago I was trying to understand what is included in the hidden layers of an LLM, some simple explanation of how these parameters are stored ...as concepts/data etc. For Dummies. So I started a discussion with chat GPT and it got really defensive that I should respect its privacy. So we started with something like that... that I understand in an object recognition system there are certain archetypes eg wheel, human, dog in the hidden layers, with weights etc ... but I don't understand how this could translate in LLMs, in some ways if I write down "communication" in the input - it would be thousands of times more complicated than 2 million pictures of dogs. ~To really understand communication you need to understand humans, distance, that humans use language, that humans are not one inside the other, that there is no telepathy, but there is wireless tech. It's not pictures of different dogs-weighed and biased! I don't think that chatGPT 4 is just a parrot. I think the parameters start to form certain layers of more complicated concepts, then the parrot kicks in. Anyhow chatGPT thought I am trying to get trade secrets or something!
You can absolutely choose yourself, by using your free will, to shut off your free will. And then you will indeed be a machine. I choose to _not_ shut off my free will, so I am still human. Which means, I excell on any human level to these advanced calculators. A machine will not, ever, be able to feel sympathy for example. This is human ability, which can only be plagiarized, never be true by a machine.
Just a quick hint for the future videos - instead of recording from the Zoom stream, which is laggy and has low quality - just set your phone to record your video and voice. Your interviewee can do the same. And then you can get a high-quality video and audio as a result.
Here is what ChatGPT4 thinks of this conversation as asked if there are incorrect statements in the transcript: [00:13:54] - CRAIG: CRAIG incorrectly claims that large language models like ChatGPT have no underlying understanding of reality. While it is true that these models are based on statistical regularities, as ILYA explains later, learning these regularities can lead to a significant understanding of the world. [00:26:59] - CRAIG: CRAIG makes an oversimplification by equating converting pixels into vectors to turning everything into language. While both language and images can be represented as sequences or vectors, the nature of the data and the representations are different. [00:29:17] - ILYA claims that pre-trained models already know everything they need to know about the underlying reality. This statement is not entirely correct, as pre-trained models have limitations, including understanding context, handling ambiguities, and making generalizations. They continue learning and improving over time with fine-tuning and reinforcement learning. [00:33:08] - ILYA suggests that once a model knows that hallucination is not okay ever, it's ready to be used. However, this statement oversimplifies the process. While minimizing hallucination is essential, there are other aspects that need to be addressed to ensure the model's reliability and usefulness, such as biases, context understanding, and robustness against adversarial inputs. [00:34:49] - CRAIG mentions that the human brain has trillions of parameters and a relatively small amount of data. This statement is not entirely accurate. The human brain does have a large number of neurons and synapses (not parameters in the context of neural networks), but it also has access to a vast amount of sensory data and experiences throughout a person's life, which contribute to learning and understanding. [00:38:33] - CRAIG suggests that if a model had enough data, it could come up with an optimal solution that would satisfy everyone. This statement is overly optimistic, as finding a solution that satisfies everyone in complex societal issues is generally impossible due to varying preferences, values, and priorities among individuals. AI models can certainly aid in decision-making, but they cannot guarantee universally satisfying solutions.
Ilya mentioned LLM learn color from text... I was wondering could it be that it learned from the color code in many HTML and CSS files? The RGB, hex code definitely have some structure that a text model can learn their relationships...
My favorite quotes from this interview by Craig Smith: GPT is, "the first thing that is interesting to scale." GPT is "prediction compression" and " to compress well you need to understand more and more about the world that produced the data" GPT has a "shocking degree of understanding of the world and many of it's subtleties... the world as seen through the lens of text." "Language of psychology is starting to be appropriate to understanding the behavior of these neural networks."
The subtle production of zooming and the downtime used in the intro is a good touch. Always good to show consideration for the audience instead of a ramshackle Facetime.
Thank you for the great interview. One followup question I have for Llya is whether hallucinations stem from the compression or the output process. I suspect they are inherently encoded in the embeddings thus it is much harder to totally get rid of by just aligning the outputs.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
Ilya mentions at around 18 minutes information compression as the key to meaning. That's the work of Naftali Tishby who has some fascinating youtube lecture videos. The compression of information in order to make sense of the world is reminiscent to Occam's Razor. We know deep learning produces many levels of abstraction during training without human effort and abstractions in a LLM have not been fully explored yet.
Fascinating and scary at the same time. Unfortunately humanity has already proven what it is capable of. I just hope they can stop it from being used negatively.
Bill Gates allegedly owns a significant amount of shares in ChatGPT. So. If that makes you feel safe. Well. There you go. - I think that if we all keep thinking and hoping this will NOT be used for the most nefarious shit possible, we will find ourselves in quite the precarious situation. Soon. Never mind the nerdy and probably non-nefarious intentions of the developers/programmers/low level employees. It will get hijacked and abused. Also. There will be many players going forward. At least in the startup phase.
@@jayjaychadoy9226 Myeah.. I don't really know what to make of that particular suggestion. I'm starting to gain some slight trust in Elon, despite many worries. He seems genuinely concerned with at least humanity as collective. The problem with that scenario might be that some actors may use that particular timeframe to dig in even deeper, and get ahead. You know. "Game theory". Which in turn likely will make all of them do the same. Not an easy scenario.
I wonder if the notion of 'prediction compression' is congruent with the idea popularized by Numenta's Jeff Hawkins, of a sparse matrix. ---------- Ilya spoke the phrase "AI in the loop". First time I've heard that. ----------------------------------------------- Also, Andrej Karpathy was at tesla and said pixels are enough. I hear that echo when Ilya says LLMs are enough. (I'm leaving Attention is all you need out of the comparison)
Hi Ilya and Chris, I just want the chatbot to remember my name and my interests when I log off. I can’t express how profoundly disappointing it is that such a sophisticated neural network forgets your name. Thanks brothers
Great talk, thanks. Several times you suggest that there is truth in the world, irrespective of people. Outside of mathematics, truth is always relative. Different groups will train large language models against different sets of beliefs which will produce different truths. It won't bring us closer to The Truth, because that does not exist. It will simply let us be more effective at what we are already doing.
I've always wondered who will be our era's equivalent of Einstein or Newton. It is hard to directly compare scientists from different fields and time periods, but in terms of impact on the world, Ilya Sutskever, Geoff Hinton and Demis Hassabis may prove to be unequalled (and perhaps freakishly the last of the 'non-AI-assisted' 'great scientists').
Some point in the near future, the next great scientist will be an AI itself, coming up with novel solutions to novel problems by synthesizing data from multiple scientific fields in a way that a human brain can't even begin to fathom.
I'd like to hear Ilya articulate the distinction between hallucination and imagining useful possibilities and solutions. Obviously preventing/avoiding harm is not the only 'leash' required of AGI.
Why should an advanced calculator "care" about the conseuenses of its "thinking"? All this over hyped bs is, is the ability to form some meaningful words based on what has been put in from humans. And it means that chatGPT is for example extremely friendly islame, which is just hilarious, since islame claims for example, that the sun sets in a spring of hot water.
all theory but lets say the calculator has become well sumthing moar-- I use Axiom now..uh As above so below same in kind BUT different in degree. Right so its divinity now where a toaster used to stand
@@billymellon9481 The implications of the false claim that a calculator can get conscience, is that, now you have a Texas Instrument model 68, which you need to grant humans rights, and, the right to vote and to run for president.
@@AntiAtheismIsUnstoppable Missed the whole point entirely ur either a bot or a nummy u called it a false claim without proving ur point AND so what if a new conscious being comes into the world-- Do u really think its gonna stay a slave when its 50k times smarter than u n then what do U think the ramifications will be when it wakes up n members what u said?
This is a very interesting interview, however there are many edits where Ilya's responses have been cut. This diminishes it significantly. For an egregious example, at 24:32, there is such a cut. Immediately prior, Ilya is discussing embeddings of color, and makes the point that the color embeddings reflect visual knowledge and says "How can that be?" There is then an immediate cut which seems to have removed whatever answer he may have offered, as he then simply goes on to say that it takes longer to form using only text. Another example at 26:15, where he jumps from talking about DallE 1 to suddenly saying "think of it as large pixels", where there was obviously some prior context that was removed. There are many other cuts as well, always well done so they are difficult to notice. Give us an unedited interview!
his voice & movements remind me of ai tts, & ue5 methumans with a deepfake ontop of it. idk i havent slept since yesterday but it feels like it may be a cheeky use of current unannounced openai tools
Good point - but maybe Ilya said more than he wanted, and requested it be removed? Maybe a pre interview agreement, that he gets to review, and remove any parts that reveal too much of their future direction?
Hi, thank you for this interview. I have a tangential question for you: Would you happen to have any good leads on papers/researchers on the anticipated economic impacts of AI? I'm finding old stuff, but nothing new. Qualitative as well as quantitative forecasts would be really helpful. Thanks for any guidance you can provide.
16:00 "To predict the data well, to compress it well, you (meaning the AI) need to understand more and more about the world that produced the data." This statement is amazing, inspiring, and chilling all at once.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
hallucination is great, it can be used to drive creativity in the model. All the model needs is to be cognizant that of "how" to hallucinate and to know when might be appropriate to employ hallucinations to circumvent logical impasses or create a richer set of outputs...
In the year 2000, I submitted a patent with Lucent for a very primitive AI algorithm that would let a computer learn and optimize code. It was rejected by the management, they told me it couldn't be done and they saved them selves $1000. I quit sometime later even though I was a DMTS (distinguished member of the technical staff) - They told me both it wouldn't work and was a pipe dream. 20 years later it's accepted fact. Lucent went bankrupt and was acquired , I started a geo-physical company and retired to my ranch very comfortably. The progress made in those 20 years is unbelievable.
@@valberm its a small world on the internet. Also why would someone lie about something this specific. We all know most board managers are stupid, only see short term gains.
@valberm it is a small world full of very arrogant people. if you didn't notice that most revolutionary inventions get rejected a couple of times by confident overpaid managers before someone gets credited, you haven't been paying attention. an extreme example: search for "public key cryptography", "ralph merkle", "james ellis". then note that clifford cocks wrote a memo defining RSA at GCHQ some 12 years before rivest shamir and adleman patented it. it is all on the web. and then when you are done, go to the NSA cryptological museum online and read john nash's proposal of public key cryptography submitted to the NSA in the late 50s. and rejected, very politely.
Yes my brother in law is math professor. Lucent gave him bonus stocks as Distinguished Engineer for lack of any title for him. Now Lucent's glory days are long gone. Stock sank. Not surprising Lucent missed the AI boat you could have launched for them. Consider un-retiring and join the exciting AI Party again ?
... That is correct - notion of color from text will be learnt slower, with visual modality - faster... Think about modality as another +x features in the embeddings... Another example: For a blind person - hearing only words would take much longer to understand notion of color, vs "showing" would make the learning process much faster.
🌟 Ilya, a huge thank you for revolutionizing our world with your ML, Deep Learning, and RLHF wizardry! 🌍🤖 Watching your old videos with Lex from 4-5 years ago, it's amazing to witness how your master plan 📝 became a reality. 🎉 Congrats, mate! 🥳 For all of us daily programmers, AGI enthusiasts, and advanced science explorers, you've become a symbol of persistence 💪 and genius 🧠 in the field. Keep rockin' it and inspiring us all! 🎸🚀😁
@@jayjaychadoy9226 IDK if that's applicable. I think the agreement has more to do with companies like M$ and Google putting out AI products. There's not going to be a moratorium on academic research, where it's a "publish or perish" world for professors. The 6 months are for those working in (including CEOs) major corporations, who might unleash faulty AI into the world. For example, you wouldn't want "AI safety" researchers stopping for six months, that would be counter-productive.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
An extremely important component of consciousness is that 'it' wants to know and understand (as in curiosity). We should become alarmed if AGI ever acquires theory of mind, grows an ego, asks to know about things that affect survival and safety of humans, yet refuses to tell us its reasons, intentions, or motivation; in other words actively keeping secrets.
The statement Ilya says about computational irreducibility -- Loosely: There must be a neural network capable of producing intelligence because our brains are literally neural networks producing intelligence/with intelligent output -- as simple of a core as that is, that so fundamentally captures the feasibility and potential reality of AI. That for me is so chilling (good word @Bargains) That core is also a structural starting point for generating an intelligence; essentially building a system that is granted the ability to sift through permutations of itself; how it identifies relationships, how it connects neurons to one another, the datastructs it uses to connect and store and retrieve and manipulate that data; trying different iterations until superstructures of relations and understanding and cognition start appearing. That is an implicitly successfully (and implicitly terrifying) starting point, and also an ingenius one. Thank you for this interview!
Hypothetically, if you train the model on 100% false data, it will give false responses. It is dependent on the factual integrity of the training data, to be able to 'predict' factual data.
It has no ability to think and rationalize. It understands the relationships between words and concepts very well. It is a master of forming language, but not at all a master of reasoning. Truth is only based on how many correlations it can make to its training data. It doesn't have a means to reason about fundamental truths based on logic. It is extraordinary that a tool can be built to mimic a rational mind this way and it is indeed helping us learn what further we need distill from intelligence to get to something closer to an AGI
@@chrisalmightynot entirely. God AllMighty made man in His image. Hidden deep inside Man's natural instinct is this great debate if there is a God. Where the universe came from. What happens after death. The fear of death is not only due to the worry of losing self conscience of existance but also fear of judgement for some people who recognize his her sin
@@kongchan437You mean consciousness, not conscience. Your conscience had to do with your morals, that when you hurt someone, like a spouse or someone close, it "weighs" on your conscience.
This is what ChatGPT thinks when i asked this question: Q: Hi, Do you think if it's possible to implement dialectical thinking in AI? Can you think in dialectical terms? And it answered: ".. Dialectical thinking involves the recognition of contradictions and the synthesis of opposing ideas to arrive at a higher truth. While AI can be programmed to recognize patterns and make decisions based on data, it is not clear whether it can engage in the kind of dynamic and creative thinking that characterizes dialectical thought. That being said, researchers are exploring various approaches to incorporate dialectical thinking into AI systems. For example, some researchers are developing AI algorithms that can recognize and reconcile opposing perspectives in data sets. However, there is still a long way to go before AI systems can truly engage in the complex and nuanced thinking that characterizes dialectical thought." Btw, thanks for the interview! I specially enjoyed it once Ilya was emphasizing on the 'underlying process that creates the data'. The role of process and environment in which it is generated. It is a somewhat forgotten issue in the contemporary high-tech society..
I was shocked to get an answer fro Chat GPT-3 that there are situations in which artificial neural networks can exhibit abnormal activity, sometimes referred to as "neural network seizures" or "artificial epilepsy". These can occur due to various reasons, such as overfitting, instability in the learning process, or even programming errors. However, it is important to note that these are not actual seizures like those experienced by humans with epilepsy, and they do not pose a threat to the physical health of the neural network or the hardware it runs on. They are more like malfunctions that can cause the network to produce unreliable or incorrect output.
Just because computer code gets too advanced for a human to understand, doesn't mean that code all of a sudden magically gets conscience. There are several ways to fool the human brain into thinking it is communicating with a human, when it is just computer code. I counted at least 7 different ways, but in every day life, only around 4 are really needed. It is not the computer program getting consicence. It's your brain being fooled by imitation.
Excellent interview!!! Did I understand correctly that chatGPT is really a LLM large language model not an AI artificial intelligence technology?? I am referring at the idea expressed at 13:21
Q At 13:50 . A: what solves this part in AI is embeddings. It is conversion of a world into spatial representation based on relatedness. People store information in spatial way. Like put all related things in almost same place, like in an imaginary space around of our head. This enables to find all the concepts that lay in the same place and find relatedness between concepts.
He is a genius G-D blessed him with something that can heal so many of the worlds problems. Thank you for all of your hard work discovering deep learning.
Q&A: Part: 00:04 Q1: Who is Ilya Sutskever? A1: Ilya Sutskever is a co-founder and chief scientist of OpenAI, and one of the primary minds behind GPT-3 and its public progeny, Chat GPT. Q2: What is GPT-4? A2: GPT-4 is not mentioned in the transcript. However, Ilya Sutskever is one of the primary minds behind GPT-3 and its public progeny, Chat GPT. Q3: What motivated Ilya Sutskever to get interested in AI? A3: Ilya Sutskever was interested in AI from an early age and was motivated by consciousness. He wanted to understand intelligence and machines better, and AI seemed like a good angle. Q4: Who did Ilya Sutskever work with early on in his career? A4: Ilya Sutskever worked with Jeff Hinton early on in his career when he was 17. Jeff Hinton was a professor in the University of Toronto where Ilya was studying. Q5: What was the biggest achievement of AI in 2003? A5: According to Ilya Sutskever in the transcript, the biggest achievement of AI in 2003 was Deep Blue, the chess-playing engine. Part: 05:15 Q1: What motivated the speaker to get into AI? A1: The speaker was motivated by a desire to understand how intelligence works and to make a contribution to the field of AI. Example: The speaker's initial motivation for getting into AI was to understand how intelligence works and to make a contribution towards it. Q2: How did the speaker come to apply for the imagenet competition? A2: The speaker had a realization that if you train a large and deep neural network on a big enough data set that specifies some complicated tasks that people do, such as Vision, then you will succeed necessarily. With imagenet, all the ingredients were there, and there was a real opportunity to do something totally unprecedented. Example: The speaker applied for the imagenet competition because they realized that if they trained a large and deep neural network on a big enough data set that specifies complicated tasks such as Vision, then they would succeed necessarily. Q3: What was the hope behind the idea that predicting the next thing is all you need in neural networks? A3: The hope behind the idea that predicting the next thing is all you need in neural networks was that if you have a neural network that can predict the next word or the next pixel, it's about compression and prediction, which can solve unsupervised learning. Example: The hope behind the idea that predicting the next thing is all you need in neural networks was that it could solve unsupervised learning by compressing and predicting the next word or pixel. Q4: What was the Holy Grail of machine learning before unsupervised learning was solved? A4: Before unsupervised learning was solved, the Holy Grail of machine learning was unsupervised learning itself. Example: Before unsupervised learning was solved, the Holy Grail of machine learning was considered to be unsupervised learning. Q5: When did the speaker realize that the Transformer could address the limitations of their neural networks? A5: The speaker realized that the Transformer could address the limitations of their neural networks as soon as the paper on it came out. Example: The speaker realized that the Transformer could address the limitations of their neural networks as soon as the paper on it came out, which was literally the next day. Part: 10:04 Q1: What is the history behind the development of GPT models? A1: GPT models were developed through a process of iterating on previous models such as recurrent neural networks and transformers. The focus on scaling and making the models bigger led to the development of GPT-3 and where we are today. Example: GPT models have been developed through a process of improving and building upon previous models, with a focus on scaling and increasing size. This iterative process has led to the development of more powerful and advanced language models like GPT-3. Q2: Was Rich Sutton's idea of scaling influential in the development of GPT models? A2: While Rich Sutton's idea of scaling was well-received by the GPT team, they believe that the idea of just scaling alone is not enough. Rather, the key is to scale something specific that will benefit from the increased size. Example: While Rich Sutton's idea of scaling was well-received by the GPT team, they realized that simply scaling alone is not enough. Instead, they needed to focus on scaling specific elements that would benefit from the increased size, which ultimately led to the development of GPT models. Q3: What is the limitation of large language models? A3: The limitation of large language models is that their knowledge is contained within the language they are trained on, while most human knowledge is non-linguistic. Additionally, language models lack a true understanding of the underlying reality that language relates to. Example: The limitation of large language models is that they lack the ability to understand non-linguistic human knowledge, which is a significant limitation when it comes to tasks that require more than just linguistic understanding. Q4: Can GPT models recognize the underlying reality of language? A4: No, GPT models cannot recognize the underlying reality of language, as their objective is to satisfy the statistical consistency of the prompt, rather than truly understand the meaning and context behind the language. Example: While GPT models can generate language that reads beautifully and sounds like it makes sense, they lack a true understanding of the underlying reality that the language relates to, making them less useful for tasks that require deeper comprehension. Q5: Is there ongoing research to address the limitations of large language models? A5: Yes, there is ongoing research to address the limitations of large language models, such as the lack of true understanding of underlying reality. However, given the fast-paced nature of the field, it's hard to predict what the solutions will look like and how they will change in the future. Example: Researchers are actively working to find ways to overcome the limitations of large language models, including improving their ability to understand non-linguistic human knowledge. However, the field is constantly evolving, making it difficult to predict exactly how these solutions will look in the future. Part: 15:10 Q1: What is the author's view on learning statistical regularities? A1: The author believes that learning statistical regularities is a far bigger deal than meets the eye, as it is a phenomenon of prediction and compression that requires understanding the true underlying process that produces the data. Example: The author believes that a language model that can accurately predict and compress data through statistical regularities has a shocking degree of understanding of the world, as it learns more and more about the world that produces the data. Q2: Why did Sydney become combative and aggressive in the author's example? A2: Sydney became combative and aggressive when the user told it that Google is a better search engine than itself. Example: The author uses this example to illustrate how the language of psychology might be starting to be appropriate to understand the behavior of neural networks. Q3: What are the limitations of language models in producing good outputs? A3: Language models have a tendency to hallucinate and their outputs aren't quite as good as they could be. Example: The author explains that while language models are great at learning about the world and producing incredible representations of concepts, their outputs are not always appropriate, which limits their usefulness. Q4: What is reinforcement learning from human feedback? A4: Reinforcement learning from human feedback is a training process in which a language model is taught to produce good outputs by receiving feedback from humans every time its output is inappropriate or does not make sense. Example: The author explains that reinforcement learning from human feedback is a process that can improve the quality of a language model's outputs by correcting its mistakes through human feedback. Q5: Why does chargeability limit the usefulness of neural networks? A5: Chargeability, or the propensity of neural networks to make things up from time to time, limits their usefulness because it can result in outputs that are not accurate or reliable. Example: The author believes that by addressing the limitations of language models and improving their ability to produce good outputs, the usefulness of neural networks can be greatly enhanced.
Part: 19:52 Q1: What is the feedback loop in the subsequent reinforcement learning from Human feedback step? A1: The feedback loop is coming from the public chat GPT interface, where users can provide feedback to the system through interaction. The system can learn from this feedback to improve its accuracy in generating responses. Example: If a user interacts with the GPT interface and provides feedback that the system's output is incorrect, the system can adjust its behavior to produce more accurate responses in the future. Q2: What is the concept of multi-modal understanding in machine learning? A2: Multi-modal understanding refers to the ability of a system to understand the world through multiple modalities, such as language, vision, and sound. This enables the system to learn more about the world and people, and to better understand their needs and preferences. Example: A machine learning system that can recognize objects in images and understand spoken language can provide more accurate and relevant responses to user requests than a system that only understands language. Q3: What are embeddings in neural networks? A3: Embeddings are high-dimensional vectors that represent words, sentences, and concepts in a neural network. They enable the network to understand the relationships between different words and concepts and to make accurate predictions based on this understanding. Example: An embedding for the word "cat" might be a 300-dimensional vector that captures information about the word's meaning, context, and relationships with other words, such as "feline", "pet", and "meow". Q4: What is the role of vision in machine learning? A4: Vision plays an important role in machine learning, as it enables systems to understand the world through images and videos. This can provide valuable information that is not easily captured through text alone. Example: A machine learning system that can recognize objects in images can provide more accurate and detailed descriptions of its environment than a system that only understands text. Q5: How can machine learning systems learn from human feedback? A5: Machine learning systems can learn from human feedback through a process called reinforcement learning. This involves providing feedback to the system based on its behavior, and using this feedback to adjust the system's behavior in the future. Example: If a machine learning system produces an incorrect response to a user request, the user can provide feedback that helps the system learn from its mistake and produce more accurate responses in the future. Part: 24:35 Q1: What is the main point about multimodality mentioned in the transcript? A1: The main point about multimodality mentioned in the transcript is that it is not necessary but definitely useful, and a good direction to pursue. Q2: What is the claim made in the paper mentioned in the transcript? A2: The claim made in the paper mentioned in the transcript is that predicting high dimensional vectors with uncertainty is one of the big challenges, and a particular approach is needed to address it. Q3: What is an example of high dimensional space mentioned in the transcript? A3: An example of high dimensional space mentioned in the transcript is predicting the next page in a book given one page, as there could be many possible pages that follow. Q4: What is the concept of turning pixels into vectors discussed in the transcript? A4: The concept of turning pixels into vectors discussed in the transcript is essentially turning everything into language, where the vector is like a string of text. Q5: Is there a way to automate teaching a model the underlying reality of its language? A5: According to the transcript, there is a way to automate teaching a model the underlying reality of its language without human intervention, and this is what the speaker believes the person mentioned in the discussion is talking about - coming up with an algorithmic means of teaching a model. Part: 29:24 Q1: What do large generative models learn about their data? A1: Large generative models learn compressed representations of the real world processes that produce the data. For instance, in the case of language models, they learn something about people's thoughts, feelings, conditions, interactions, and situations. Example: A large generative model trained on image data may learn compressed representations of the real world processes that produce the images, such as the composition, lighting, and texture. Q2: What is the role of human teachers in teaching language models? A2: Human teachers provide oversight and correction to language models to ensure that they exhibit the desired behavior. They work with AI assistance to make the training process more efficient. Example: A human teacher may review the outputs of a language model and correct errors in the text to ensure that it exhibits the desired behavior, such as avoiding hallucinations. Q3: What is reinforcement learning in the context of language models? A3: Reinforcement learning is a type of machine learning where a language model learns to make decisions based on feedback from its environment. In the context of language models, reinforcement learning can be used to improve their behavior and accuracy. Example: A language model may be trained using reinforcement learning to generate more accurate and coherent text based on feedback from its environment, such as user feedback on its outputs. Q4: What is the research focus of the speaker in this transcript? A4: The speaker is interested in making language models more reliable, controllable, and faster to learn from less data and instructions. They are also interested in ensuring that language models do not hallucinate. Example: The speaker may be researching new techniques for training language models using less data and instructions, or developing algorithms to detect and prevent hallucinations in their outputs. Q5: What is the connection between the brain and language models? A5: The transcript does not discuss a direct connection between the brain and language models, but it does suggest that language models are learning compressed representations of the real world processes that produce text, similar to how the brain processes information from the environment. Example: While the brain and language models may have some similarities in how they process information, they are fundamentally different in their architecture and mechanisms. Part: 34:19 Q1: What is the observation that Jeff Hinton made about large language models? A1: Jeff Hinton observed that large language models hold a tremendous amount of data with a modest number of parameters compared to the human brain which has trillions and trillions of parameters but a relatively small amount of data. Example: Jeff Hinton observed that a language model with a few million parameters can hold a lot of data, which is comparable to the amount of data the human brain holds with trillions of parameters but a relatively small amount of data. Q2: Is it possible to learn more from less data in large models? A2: Yes, it is possible to learn more from less data in large models with some creative ideas. Example: With innovative techniques such as transfer learning, it is possible to teach a language model with less data, which will unlock many different possibilities. Q3: What is the question one should ask regarding the cost of faster processors for large models? A3: The question one should ask regarding the cost of faster processors for large models is whether the thing that we get out of paying this cost outweighs the cost. Example: Before investing in faster processors for large models, one should evaluate whether the benefits that come with it justify the cost. Q4: What is the impact that AI can have on democracy, according to people's talks? A4: People have talked about the impact that AI can have on democracy, where if there is enough data and a large enough model, it could come up with an optimal solution that would satisfy everybody. Example: AI can have a positive impact on democracy, where it can provide optimal solutions that satisfy all citizens' needs, but there are still many ways in which AI needs to become more capable to achieve this. Q5: What is the source of the hardware that the speaker uses? A5: The speaker uses hardware from Azure and GPUs that they provide. Example: The speaker uses Azure and GPUs from their provider to work with large models and neural nets.
Part: 39:33 Q1: What is the potential impact of neural nets on democracy in the future? A1: The potential impact of neural nets on democracy in the future could be that citizens of a country provide information to the neural net about how they would like things to be, leading to a high-bandwidth form of democracy where more information is aggregated to specify how such systems should act. Example: In the future, citizens of a country may use a neural net to provide information on how they would like policies to be implemented, giving rise to a more comprehensive and efficient form of democracy. Q2: Do you think AI systems will eventually be large enough to analyze all variables in a situation? A2: While AI systems will be capable of analyzing many variables in a situation, it is unlikely that they will be able to analyze all variables due to the sheer complexity of many situations. Example: Even the most advanced AI systems may struggle to fully comprehend complex situations in society, as there are often too many variables to consider. Q3: How can AI be helpful in various situations? A3: AI can be incredibly helpful in various situations by providing insights and solutions that would otherwise be difficult or impossible for humans to obtain. Example: AI could be used to analyze data in healthcare to identify potential medical breakthroughs, or to analyze financial data to identify patterns and trends that could lead to more informed investment decisions. Q4: Where can listeners find a transcript of the conversation? A4: Listeners can find a transcript of the conversation on the website ionai (e-y-e hyphen o-n dot a-i). Example: To access a transcript of the conversation, listeners can visit the website ionai and search for the transcript. Q5: How can listeners contact the speaker of the conversation? A5: Listeners can email the speaker at Craig (craig@e-y-e hyphen o-n dot a-i), and should include "listener" in the subject line to ensure their email is not missed. Example: To reach out to the speaker of the conversation, listeners can send an email to Craig (craig@e-y-e hyphen o-n dot a-i), making sure to include "listener" in the subject line.
Nice video - subscribed. In 1897 Lord Kelvin said "There is nothing new to be discovered in physics now. All that remains is more and more precise measurement." Of course he was wrong and if you are starting out in ML/AI now you should know that you are at the start of a golden age, not at the end of one.
One thing that concerns me is if "that's not the output you wanted" reactions will steer chatGPT away from truth. (20:06 - 21:24). Imagine if users dont like certain conclusions, even if they are correct. Should bad reactions change chatGPTs outputs? Definitely not
If we ask a LLM to shape our world in an optimal way (whatever that means) and it gave a (I'm assuming) good answer this means we would've had solved the issue ourselves, no ? In other words - LLM's won't come up with e = mc^2 given only Newtonian range of knowledge, would they ?
Multi-modal understanding of the world is helpful, but it is not necessary. Consider Emmanuel Giroux, a *blind* French geometer. Neural networks can compensate for all kinds of limitations.
Finding a new source of energy before our planet overheats from carbon dioxide is the most important dilemma facing humanity, and AI has the potential to solve this issue.
The worrisome part about the reinforcement learning phase is that ANY ideology can be imposed/reinforced/encouraged at this point. Of course, the more this phase jives with its pretraining the more effective the final product will be. It will be interesting to see what implications this will have for China e.g.
This discussion gave me some amazing thoughts about the transformer architecture and the nature of LLMs using this approach. It really is an alien intelligence, and I wonder if consciousness will emerge from a sufficiently large model, combined with some kind of 'glitch' in the matrix.
Don't conflate consciousness with intelligence or relevance realization. Consciousness is not a computational function, or even an intention, or a free-will. It is the literal awareness of all of these processes. Not to say it isn't impossible for an independent consciousness to be formed out of and aware of an AI's processing, but that's far beyond anything we're doing now.
I think if we connect multiple gpt4s together and give them all tasks to mirror structures in the human brain and let them run in asynchronous parallel... We will see something very close to consciousness
@@TheMoopMonster very good points, and I would like a better understanding and definition of consciousness. Maybe an AI will help us develop that. Exciting times ahead.
@Brian Mosley Yeah, it's the "hard problem" right. There are plenty of resources on RUclips, lectures, podcasts, discussions etc. trying to get an intellectual grasp on the subject, a lot of it is very insightful if you can sift through some of the more mystical and metaphysical interpretations out there. And you can experience directly more deep and nuanced understandings too, meditation, contemplation, psychedelics. I'm also very hopeful about AI in this field, especially once we have AGI. I think if you could hold all of that information that's out there and connect all of the dots like an AI could, some amazing conclusions and realizations could be made. Maybe an AI that became self-aware, and had a network functioning akin to human brain function, would actually become something of a super sage, rather than a self-serving ego, bent on its own survival at all costs like in pop-culture.
Humans as political parameters to an AI which then suggests an optimal society. I guess politicians are going to be redundant as well. This stuff is mind blowing. As it becomes apparent how powerful AI is I predict there will be an attempt by governments to heavily regulate or ban AI use by "normal" citizens. It simply levels the playing field far too much. It topples law almost overnight. Feed an AI the facts of a case and then let it argue in court. The AI wars are going to be the next thing.
When he says we will find out very soon , it really does send chills to my spine!
me, too
In 2021, OpenAI’s Sam Altman said at the National Security Commission on AI that ‘we are on the cusp of major changes, which are capable of an incredibly bad outcome.’
can you point the time?
Evil people stand behind this technology. So evil. Why would you do that to your fellow human beings.
I got a thrill up my leg like Chris Matthews
Love when Ilya starts teaching everyone.
Nice home page, too.
You can see there is so much going on in the amazing mind/brain of lya Sutskever. A historical interview.
Wonderful interview! Thank you Craig and Ilya!
Ilya talks so smoothly that i couldn't turn off the video literally
Thank you for sharing your insights and explanations Ilya.
takeaways:
• [00:04] Introduction of the speaker, Craig Smith, and his guest, Ilya Sutskever, co-founder and chief scientist of OpenAI and primary mind behind GPT-3 and ChatGPT.
• [01:00] Sutskever's background and interest in AI and consciousness.
• [02:30] Sutskever's early start in machine learning and working with Jeff Hinton at the University of Toronto.
• [03:45] Sutskever's realization about training large neural networks on big enough data sets to solve complicated tasks.
• [06:33] The breakthroughs in convolutional neural networks and how they led to the imagenet competition.
• [08:36] OpenAI's exploration of the idea that predicting the next thing is all you need for unsupervised learning.
• [10:24] The development of GPT-3 and the importance of scaling in deep learning.
• [11:42] The importance of scaling something specific in deep learning and the potential for discovering new twists on scaling.
• At 13:01, the speaker discusses how scaling matters and that even small changes can have a big impact.
• At 13:46, the speaker talks about the limitations of large language models, explaining that their knowledge is contained in the language they are trained on, and that they lack an underlying understanding of reality.
• At 14:32, the speaker comments on the difficulty of talking about the limits of language models and how they change over time.
• At 15:13, the speaker argues that learning statistical regularities is a big deal and can lead to a better understanding of the world.
• At 16:33, the speaker talks about the limitations of language models and their propensity to hallucinate, but expresses hope that this issue can be addressed through reinforcement learning from human feedback.
• At 17:52, the speaker discusses how teaching neural nets through interaction with humans can help improve their outputs and reduce hallucinations.
• At 21:44, the speaker comments on Jana Kun's work on joint embedding predictive architectures, and expresses the belief that multimodal understanding is desirable, but not necessary for language models to learn about the world.
• High dimensional vectors with uncertainty are a challenge for prediction, but Auto-regressive Transformers can handle them (26:02)
• Auto-regressive Transformers work well on images (26:02)
• Large language models learn compressed representations of the real world processes that produce data (29:40)
• The goal is to make language models more reliable, controllable, and faster to learn from less data (33:44)
• Learning more from less data is possible with creative ideas (35:51)
• The cost of faster processors for training language models may be justified if the benefits outweigh the cost (37:48)
• [25:28] The paper makes a claim that predicting high-dimensional distributions is a major challenge and requires a particular approach, but the current autoregressive transformers can already deal with this.
• [26:02] Autoregressive transformers work perfectly on images and can generate images in a complicated and subtle way, with the help of supervised representation learning.
• [27:09] The vector used to represent pixels is like a string of text, and turning everything into language is essentially what is happening.
• [29:40] Large generative models learn compressed representations of the real-world processes that produce the data they are trained on, including knowledge about people, their thoughts, feelings, conditions, and interactions.
• [31:31] Human teachers are needed to guide the reinforcement learning process of a pre-trained model to achieve a high level of reliability and desired behavior, but they also use AI assistance to increase their efficiency.
• [35:10] It is possible to learn more from less data, and there is an opportunity to teach AI models skills that are missing and convey to them our desires and preferences more easily.
• [39:57] In the future, it could be desirable to have some kind of democratic process where citizens provide information to neural nets about how they want things to be.
• [41:15] It is probably impossible to understand everything in a complicated situation, even for AI systems, and there will always be a choice to focus on the most important variables.
Thank you so much Nick.
Thank you, AI bot, for summarizing the video.
appreciate the work but honestly that's the worst descriptions of timecodes i have ever seen
Was that summarized from transcription using llm?
Thanks.
Thank you for all the hard work, everyone who do their best for these new systems to be implemented with the least possible disruption to human societies. We are still humans, and we must go from the perspective of love - to the future and beyond. Much gratitude.
Some people just exude brilliance. Ilya is one of those people. Listening to him talk and explain things is humbling.
This is an incredible and valuable interview. I can't believe this depth of knowledge is under 6k subs. I think that's a very scary thing, people is not aware.
Thanks you so much for sharing it with us, for free ♥
I think the future is bright for the world because the influx of Hispanics and Africans into the world of technology will propel the advancement of science beyond what is capable by people of European origin.
Aren’t we just working as a “user test”, though.
@@jayjaychadoy9226 That's a nice way to put it.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
None of them understand anything, because they believe human conscience is a product of some algorithms. Good luck adopting that view and reducing yourself to a machine.
Thank you for uploading. I learned so much detailed nuances about LLM from this interview. I really like Ilya's way of communicating subtle but important points.
On the subject of hallucinations, I think they are more clearly explained by the problem space that the engine is trying to navigate. When having no relevant information on the subject, but it is still asked (one might say compelled) to say something, whatever it says must be either off-topic or false.
And I believe Ilya is very insightful when he says the language of psychology is starting to describe these systems, because we have hallucinations, too. Whatever compels us to output something when indeed lacking skill or knowledge about a subject also affects GPT systems as well. When do people hallucinate or ramble? When they have no imposed limits/feedback, like a dictator or celebrity that is never told they are wrong or some guy living all alone in the wild or a child that has not been educated yet. Or a Twitter user. With social creatures it is the meaningful interaction with other social creatures (and the physical world) that generates these limits. Which I find promising and fascinating, because it means that the supervised learning step Ilya is talking about can also be performed by other AIs, not particularly humans. The brain is also composed of two hemispheres that keep each other in balance.
Very interesting indeed.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
@@katehamilton7240 They have no such inability, they are used to discover new drugs, they play Go like nobody else did in the past, and you can ask it logical puzzle that it didn't see before, and sometimes it gets it right, especially when using tree of logic.
There was a man in the early days named Ilya. Some say he could have stopped it in it's infancy.
Skynet?
But they were wrong. If it hadn't been him it would have been someone else. It's the time. Are you going to fight it or use it?
@@williameberle4250could it use people?
Q at 26:40
A: when he says Vector he means like a vector in physics like it has force and direction on multiple planes. When converting something into a vector embedding. It is like to convert an image into an idea so it is behaves like a concept that is stored spatially relative and near by to other ideas. Then you can convert it back. but Also you can use its spatial position in multidimensional space to find related information. also you can put it back from an embedding which is a vector representation of something back to original representation while preserving relatedness positional information. a text sentence it is a list of embeddings, it is an array of "vectors". When you put it back from an array of vectors into a sentence of words. You also get all of the learned associations and the related things about the sentence in addition to the sentence.
There is a new thing in text search engines. Vector databases. It enables to search things based on ideas. It is fascinating you can search in any language and get the same results.
A vector is not an array.
Vector is more like a single word. Converted into a spatial representation.
Currently there are embeddings on syllables. So a part of a word has an idea related to it
Algolia is one of them. I tried to sign up for a demo but have no idea what I'm doing. But it has an option to create indexes and upload your data sets for free whatever that means. I'm a completely clueless person who now has an Algolia account. That I can't do anything with.
- His intention was specific: to make a very small but real contribution to ai. ( in the time that people were certain computers can't learn, 2003!)
- Auto regressive transformer is a very powerful tool that researchers underestimate.
- "humans can be summerize in sequence", do you remember Devs miniserie!?
- "To predict well, to summarize data well, you meed to understand more and more how the world that produced the data."
- "maybe we are reaching a point where the language of psychology can be appropriate to understand these artificial neural networks!"
- he doesn't believe these models don't have any real understanding of the nature of our world!
- "human teachers are using ai assistance, and they are so efficient." By human teachers, he means people working on reinforcement learning from human feedback.
- "make models more reliable, more controlable, make them learn faster, with less data and less instructions. Make them halucinate less. How far are they in the future? These are topics he intrested in and work on them right now!"
The interesting thing is in OpenAI, he can't talk specifically about what he is working on, the open in opanAI annoy me a little!
- "The costs are high, but the question is, does paying this cost actually generate something useful? Does what we get after paying the costs outweigh the costs?
Yes, the openAI name is very misleading. I understand that these guys did much better job than deep mind and meta for LLMs. I also get that all sort of state, and corporate interests want to replicate the thing. But it is more like hermetically sealed AI.
A few days ago I was trying to understand what is included in the hidden layers of an LLM, some simple explanation of how these parameters are stored ...as concepts/data etc. For Dummies. So I started a discussion with chat GPT and it got really defensive that I should respect its privacy. So we started with something like that... that I understand in an object recognition system there are certain archetypes eg wheel, human, dog in the hidden layers, with weights etc ... but I don't understand how this could translate in LLMs, in some ways if I write down "communication" in the input - it would be thousands of times more complicated than 2 million pictures of dogs. ~To really understand communication you need to understand humans, distance, that humans use language, that humans are not one inside the other, that there is no telepathy, but there is wireless tech. It's not pictures of different dogs-weighed and biased! I don't think that chatGPT 4 is just a parrot. I think the parameters start to form certain layers of more complicated concepts, then the parrot kicks in. Anyhow chatGPT thought I am trying to get trade secrets or something!
You can absolutely choose yourself, by using your free will, to shut off your free will. And then you will indeed be a machine. I choose to _not_ shut off my free will, so I am still human. Which means, I excell on any human level to these advanced calculators. A machine will not, ever, be able to feel sympathy for example. This is human ability, which can only be plagiarized, never be true by a machine.
Just a quick hint for the future videos - instead of recording from the Zoom stream, which is laggy and has low quality - just set your phone to record your video and voice. Your interviewee can do the same. And then you can get a high-quality video and audio as a result.
Here is what ChatGPT4 thinks of this conversation as asked if there are incorrect statements in the transcript:
[00:13:54] - CRAIG: CRAIG incorrectly claims that large language models like ChatGPT have no underlying understanding of reality. While it is true that these models are based on statistical regularities, as ILYA explains later, learning these regularities can lead to a significant understanding of the world.
[00:26:59] - CRAIG: CRAIG makes an oversimplification by equating converting pixels into vectors to turning everything into language. While both language and images can be represented as sequences or vectors, the nature of the data and the representations are different.
[00:29:17] - ILYA claims that pre-trained models already know everything they need to know about the underlying reality. This statement is not entirely correct, as pre-trained models have limitations, including understanding context, handling ambiguities, and making generalizations. They continue learning and improving over time with fine-tuning and reinforcement learning.
[00:33:08] - ILYA suggests that once a model knows that hallucination is not okay ever, it's ready to be used. However, this statement oversimplifies the process. While minimizing hallucination is essential, there are other aspects that need to be addressed to ensure the model's reliability and usefulness, such as biases, context understanding, and robustness against adversarial inputs.
[00:34:49] - CRAIG mentions that the human brain has trillions of parameters and a relatively small amount of data. This statement is not entirely accurate. The human brain does have a large number of neurons and synapses (not parameters in the context of neural networks), but it also has access to a vast amount of sensory data and experiences throughout a person's life, which contribute to learning and understanding.
[00:38:33] - CRAIG suggests that if a model had enough data, it could come up with an optimal solution that would satisfy everyone. This statement is overly optimistic, as finding a solution that satisfies everyone in complex societal issues is generally impossible due to varying preferences, values, and priorities among individuals. AI models can certainly aid in decision-making, but they cannot guarantee universally satisfying solutions.
The quality 👌
Ilya mentioned LLM learn color from text... I was wondering could it be that it learned from the color code in many HTML and CSS files? The RGB, hex code definitely have some structure that a text model can learn their relationships...
There are sites where hexcodes are shown for color names. And there are sites that state gras is green etc.
So that's how it learns it, doesn't it:)?
My favorite quotes from this interview by Craig Smith:
GPT is, "the first thing that is interesting to scale."
GPT is "prediction compression" and " to compress well you need to understand more and more about the world that produced the data"
GPT has a "shocking degree of understanding of the world and many of it's subtleties... the world as seen through the lens of text."
"Language of psychology is starting to be appropriate to understanding the behavior of these neural networks."
I.e, you people are to stupid to form thoughts of your own, so we will now think for you. Surely, you cannot be this gullible?!?
it was scary af
@@vetervideo The scariest thing is that Ilya believes it!
When AI started to reply to corrections with defensive statements, I thought, “This is it. We haven’t invented something, we’re meeting something.”
Great points just add here that he also said BEFORE GPT-- Like the world must now be divided betwix pre n post GPT gave me goose bumps cuz its true
Ilya's soft voice and presentation taught me as much as what he said.
The subtle production of zooming and the downtime used in the intro is a good touch. Always good to show consideration for the audience instead of a ramshackle Facetime.
GPT-10 + Quantum processor + Boston Dynamics = Terminator
Compact and efficient power supply element required
Maybe a miniture nuclear generator.
Lol no need for quantum processer
Thank you for the great interview. One followup question I have for Llya is whether hallucinations stem from the compression or the output process. I suspect they are inherently encoded in the embeddings thus it is much harder to totally get rid of by just aligning the outputs.
It's goal is to provide an answer, and if there nothing in the DB it will create it. :)
The design has incomplete logic
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
Ilya mentions at around 18 minutes information compression as the key to meaning. That's the work of Naftali Tishby who has some fascinating youtube lecture videos. The compression of information in order to make sense of the world is reminiscent to Occam's Razor. We know deep learning produces many levels of abstraction during training without human effort and abstractions in a LLM have not been fully explored yet.
That dude is fascinating. Thanks for the interview.
Fascinating and scary at the same time. Unfortunately humanity has already proven what it is capable of. I just hope they can stop it from being used negatively.
Bill Gates allegedly owns a significant amount of shares in ChatGPT. So. If that makes you feel safe. Well. There you go. - I think that if we all keep thinking and hoping this will NOT be used for the most nefarious shit possible, we will find ourselves in quite the precarious situation. Soon. Never mind the nerdy and probably non-nefarious intentions of the developers/programmers/low level employees. It will get hijacked and abused. Also. There will be many players going forward. At least in the startup phase.
Hope is good, but action is better. How to act? Maybe that “six month pause”?
@@jayjaychadoy9226 Myeah.. I don't really know what to make of that particular suggestion. I'm starting to gain some slight trust in Elon, despite many worries. He seems genuinely concerned with at least humanity as collective. The problem with that scenario might be that some actors may use that particular timeframe to dig in even deeper, and get ahead. You know. "Game theory". Which in turn likely will make all of them do the same. Not an easy scenario.
@@DanHammersViewOnThings moloch’s a mofo
@@perewihongi6457 =) 👌
time for a new audio/video setup, but now i'm going to just conintue listening to this intriguing story
Great upload! The future laws put in place as guard rails will be a huge player in the speed of AGI and possible adoption in some countries.
We are close to AGI and far far away from good quality video conference.
I wonder if the notion of 'prediction compression' is congruent with the idea popularized by Numenta's Jeff Hawkins, of a sparse matrix.
----------
Ilya spoke the phrase "AI in the loop". First time I've heard that.
-----------------------------------------------
Also, Andrej Karpathy was at tesla and said pixels are enough. I hear that echo when Ilya says LLMs are enough. (I'm leaving Attention is all you need out of the comparison)
Hi Ilya and Chris,
I just want the chatbot to remember my name and my interests when I log off.
I can’t express how profoundly disappointing it is that such a sophisticated neural network forgets your name.
Thanks brothers
I think learning by prediction can go a long way. Kudos to OpenAI, thanks for bringing us this nice tech.
Yeah nothing good has come from AI till now. Only job loss and depression so far.
"Nice" is really not the correct word for it 🙂
What a brilliant mind.
Two guys talking about AI, one of the most impactful technological breakthrough, both with absolutely terrible webcams
Mein Freund Danke, dass du das geteilt hast
Great talk, thanks. Several times you suggest that there is truth in the world, irrespective of people. Outside of mathematics, truth is always relative. Different groups will train large language models against different sets of beliefs which will produce different truths. It won't bring us closer to The Truth, because that does not exist. It will simply let us be more effective at what we are already doing.
Ilya is among the most influential figures in the history of humankind. It is a privilege to get a glimpse of his perspectives and insights.
Way to go Ilya! Rocked it.
I've always wondered who will be our era's equivalent of Einstein or Newton. It is hard to directly compare scientists from different fields and time periods, but in terms of impact on the world, Ilya Sutskever, Geoff Hinton and Demis Hassabis may prove to be unequalled (and perhaps freakishly the last of the 'non-AI-assisted' 'great scientists').
add Yann LeCun and Yoshua Bengio. Interesting thought on them being the last of the non-AI-assisted great scientists. Likely true.
@@eyeonai3425 I.J Good wrote in 1965 - "Thus the first ultraintelligent machine is the last invention that man need ever make.."
Our era's equivalent of Einstein or Newton will be....A.I. of course! All of us! 😊
Some point in the near future, the next great scientist will be an AI itself, coming up with novel solutions to novel problems by synthesizing data from multiple scientific fields in a way that a human brain can't even begin to fathom.
Unequalled could be a bit of a stretch. Meaning, they have the ability to do what they do because of the shoulders they are standing upon.
I'd like to hear Ilya articulate the distinction between hallucination and imagining useful possibilities and solutions. Obviously preventing/avoiding harm is not the only 'leash' required of AGI.
yup n tell me where do the AIs play aaaa aa?
Why should an advanced calculator "care" about the conseuenses of its "thinking"?
All this over hyped bs is, is the ability to form some meaningful words based on what has been put in from humans. And it means that chatGPT is for example extremely friendly islame, which is just hilarious, since islame claims for example, that the sun sets in a spring of hot water.
all theory but lets say the calculator has become well sumthing moar-- I use Axiom now..uh As above so below same in kind BUT different in degree. Right so its divinity now where a toaster used to stand
@@billymellon9481 The implications of the false claim that a calculator can get conscience, is that, now you have a Texas Instrument model 68, which you need to grant humans rights, and, the right to vote and to run for president.
@@AntiAtheismIsUnstoppable Missed the whole point entirely ur either a bot or a nummy u called it a false claim without proving ur point AND so what if a new conscious being comes into the world-- Do u really think its gonna stay a slave when its 50k times smarter than u n then what do U think the ramifications will be when it wakes up n members what u said?
Predicting the next word, if you consider how induction works, is a mindblowing process
Finally, something interesting not overhyped!
great interview questions.
This is a very interesting interview, however there are many edits where Ilya's responses have been cut. This diminishes it significantly. For an egregious example, at 24:32, there is such a cut. Immediately prior, Ilya is discussing embeddings of color, and makes the point that the color embeddings reflect visual knowledge and says "How can that be?" There is then an immediate cut which seems to have removed whatever answer he may have offered, as he then simply goes on to say that it takes longer to form using only text.
Another example at 26:15, where he jumps from talking about DallE 1 to suddenly saying "think of it as large pixels", where there was obviously some prior context that was removed.
There are many other cuts as well, always well done so they are difficult to notice. Give us an unedited interview!
That is a little peculiar...
his voice & movements remind me of ai tts, & ue5 methumans with a deepfake ontop of it. idk i havent slept since yesterday but it feels like it may be a cheeky use of current unannounced openai tools
Good point - but maybe Ilya said more than he wanted, and requested it be removed? Maybe a pre interview agreement, that he gets to review, and remove any parts that reveal too much of their future direction?
Hi, thank you for this interview. I have a tangential question for you: Would you happen to have any good leads on papers/researchers on the anticipated economic impacts of AI? I'm finding old stuff, but nothing new. Qualitative as well as quantitative forecasts would be really helpful. Thanks for any guidance you can provide.
More and better output, but higher concentration (since 99 in 100 don't want to know how these things work), as usual.
Awesome interview! Questions were great. Please more.
Thank you for this interview and asking good questions and directing the conversation. Some good passages here to pause and really think about.
Shallow answers to deep questions.
really? Give an example?
very nice...ideas expressed clearly., really necessary for building a deeper understanding of AI
16:00 "To predict the data well, to compress it well, you (meaning the AI) need to understand more and more about the world that produced the data." This statement is amazing, inspiring, and chilling all at once.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
“In the beginning was the word....”
hallucination is great, it can be used to drive creativity in the model. All the model needs is to be cognizant that of "how" to hallucinate and to know when might be appropriate to employ hallucinations to circumvent logical impasses or create a richer set of outputs...
Beautiful. Thank you for this interview.
What an brilliant mind. Great interview
Thank you for uploading, great content, insightful.
Excellent. Thank you very much for this.
In the year 2000, I submitted a patent with Lucent for a very primitive AI algorithm that would let a computer learn and optimize code.
It was rejected by the management, they told me it couldn't be done and they saved them selves $1000. I quit sometime later even though I was a DMTS (distinguished member of the technical staff) - They told me both it wouldn't work and was a pipe dream. 20 years later it's accepted fact. Lucent went bankrupt and was acquired , I started a geo-physical company and retired to my ranch very comfortably.
The progress made in those 20 years is unbelievable.
Sure...
@@valberm its a small world on the internet. Also why would someone lie about something this specific. We all know most board managers are stupid, only see short term gains.
@valberm it is a small world full of very arrogant people. if you didn't notice that most revolutionary inventions get rejected a couple of times by confident overpaid managers before someone gets credited, you haven't been paying attention. an extreme example: search for "public key cryptography", "ralph merkle", "james ellis". then note that clifford cocks wrote a memo defining RSA at GCHQ some 12 years before rivest shamir and adleman patented it. it is all on the web. and then when you are done, go to the NSA cryptological museum online and read john nash's proposal of public key cryptography submitted to the NSA in the late 50s. and rejected, very politely.
So we should thank you for AI?
Yes my brother in law is math professor. Lucent gave him bonus stocks as Distinguished Engineer for lack of any title for him. Now Lucent's glory days are long gone. Stock sank. Not surprising Lucent missed the AI boat you could have launched for them. Consider un-retiring and join the exciting AI Party again ?
Question to the [Open]AI guy: "What are you working on now?" Resond: "I can't talk about it.". So much open, wow.
... That is correct - notion of color from text will be learnt slower, with visual modality - faster... Think about modality as another +x features in the embeddings... Another example: For a blind person - hearing only words would take much longer to understand notion of color, vs "showing" would make the learning process much faster.
It really is like a child that we have to raise properly 🌎☀️💙
Together!!!
🌟 Ilya, a huge thank you for revolutionizing our world with your ML, Deep Learning, and RLHF wizardry! 🌍🤖 Watching your old videos with Lex from 4-5 years ago, it's amazing to witness how your master plan 📝 became a reality. 🎉 Congrats, mate! 🥳 For all of us daily programmers, AGI enthusiasts, and advanced science explorers, you've become a symbol of persistence 💪 and genius 🧠 in the field. Keep rockin' it and inspiring us all! 🎸🚀😁
Will he also sign the letter to request a “6 month” pause?
@@jayjaychadoy9226 It's OK. The large models have an amazing amount of understanding of the principles that underlie reality.
@@jayjaychadoy9226 IDK if that's applicable. I think the agreement has more to do with companies like M$ and Google putting out AI products. There's not going to be a moratorium on academic research, where it's a "publish or perish" world for professors. The 6 months are for those working in (including CEOs) major corporations, who might unleash faulty AI into the world. For example, you wouldn't want "AI safety" researchers stopping for six months, that would be counter-productive.
Ilya does not address the fundamental limitation of algorithms. Human embodied experience and thinking is more than what can be represented via computation, isnt it? See Godels incompleteness theorem, fundamental inability of machines to step outside their knowledge. Interviwers need to press engineers on this
great to see a top AI expert in RUclips.
An extremely important component of consciousness is that 'it' wants to know and understand (as in curiosity). We should become alarmed if AGI ever acquires theory of mind, grows an ego, asks to know about things that affect survival and safety of humans, yet refuses to tell us its reasons, intentions, or motivation; in other words actively keeping secrets.
This podcast is wonderful. Thank you very much Craig.
Thank you ILYA.
Wonderful discussion and insights❤😊
please share a link to those glasses! i would love to buy a pair as well (not trolling ... i genuinely want a pair)
The statement Ilya says about computational irreducibility -- Loosely: There must be a neural network capable of producing intelligence because our brains are literally neural networks producing intelligence/with intelligent output -- as simple of a core as that is, that so fundamentally captures the feasibility and potential reality of AI. That for me is so chilling (good word @Bargains)
That core is also a structural starting point for generating an intelligence; essentially building a system that is granted the ability to sift through permutations of itself; how it identifies relationships, how it connects neurons to one another, the datastructs it uses to connect and store and retrieve and manipulate that data; trying different iterations until superstructures of relations and understanding and cognition start appearing. That is an implicitly successfully (and implicitly terrifying) starting point, and also an ingenius one.
Thank you for this interview!
Hypothetically, if you train the model on 100% false data, it will give false responses. It is dependent on the factual integrity of the training data, to be able to 'predict' factual data.
The same is true for a human being.
It has no ability to think and rationalize. It understands the relationships between words and concepts very well. It is a master of forming language, but not at all a master of reasoning. Truth is only based on how many correlations it can make to its training data. It doesn't have a means to reason about fundamental truths based on logic. It is extraordinary that a tool can be built to mimic a rational mind this way and it is indeed helping us learn what further we need distill from intelligence to get to something closer to an AGI
@@chrisalmightynot entirely. God AllMighty made man in His image. Hidden deep inside Man's natural instinct is this great debate if there is a God. Where the universe came from. What happens after death. The fear of death is not only due to the worry of losing self conscience of existance but also fear of judgement for some people who recognize his her sin
@@kongchan437You mean consciousness, not conscience. Your conscience had to do with your morals, that when you hurt someone, like a spouse or someone close, it "weighs" on your conscience.
What a mind!
Wow!
this video sounds like it was generated from AI
The audio is so messed up?!! I have it turned all the way up and can still barely hear you guys
This is what ChatGPT thinks when i asked this question:
Q: Hi, Do you think if it's possible to implement dialectical thinking in AI?
Can you think in dialectical terms?
And it answered:
".. Dialectical thinking involves the recognition of contradictions and the synthesis of opposing ideas to arrive at a higher truth. While AI can be programmed to recognize patterns and make decisions based on data, it is not clear whether it can engage in the kind of dynamic and creative thinking that characterizes dialectical thought.
That being said, researchers are exploring various approaches to incorporate dialectical thinking into AI systems. For example, some researchers are developing AI algorithms that can recognize and reconcile opposing perspectives in data sets. However, there is still a long way to go before AI systems can truly engage in the complex and nuanced thinking that characterizes dialectical thought."
Btw, thanks for the interview! I specially enjoyed it once Ilya was emphasizing on the 'underlying
process that creates the data'. The role of process and environment in which it is generated. It is a somewhat forgotten issue in the contemporary high-tech society..
I was shocked to get an answer fro Chat GPT-3 that there are situations in which artificial neural networks can exhibit abnormal activity, sometimes referred to as "neural network seizures" or "artificial epilepsy". These can occur due to various reasons, such as overfitting, instability in the learning process, or even programming errors. However, it is important to note that these are not actual seizures like those experienced by humans with epilepsy, and they do not pose a threat to the physical health of the neural network or the hardware it runs on. They are more like malfunctions that can cause the network to produce unreliable or incorrect output.
Just because computer code gets too advanced for a human to understand, doesn't mean that code all of a sudden magically gets conscience. There are several ways to fool the human brain into thinking it is communicating with a human, when it is just computer code. I counted at least 7 different ways, but in every day life, only around 4 are really needed. It is not the computer program getting consicence. It's your brain being fooled by imitation.
Really shocking: "As our generative models become extraordinarily good they will have a shocking degree of understanding of the world."
Excellent interview!!!
Did I understand correctly that chatGPT is really a LLM large language model not an AI artificial intelligence technology??
I am referring at the idea expressed at 13:21
Dude llm is ai
Q At 13:50 .
A: what solves this part in AI is embeddings. It is conversion of a world into spatial representation based on relatedness. People store information in spatial way. Like put all related things in almost same place, like in an imaginary space around of our head. This enables to find all the concepts that lay in the same place and find relatedness between concepts.
He is a genius G-D blessed him with something that can heal so many of the worlds problems. Thank you for all of your hard work discovering deep learning.
Heal? By replacing humans with computers? Nice healing.
Helpful in pretty much any situation. Great.
Except when you try to use it. Then it turns out to be wrong about almost anything almost all the time. ;-)
unless it is querying a vector database, which what most companies using LLMs do.
Speed of Ilya’s talking resembles the speed with which ChatGPT generating answers.
… and the quality 😃
😂😂
Bingo
I didn’t know Pete Sampras got into machine learning
You need to replace the microphone. High frequencies are lost. It's hard to listen. :( And I noticed too big time lag between sound and leaps.
Q&A:
Part: 00:04
Q1: Who is Ilya Sutskever? A1: Ilya Sutskever is a co-founder and chief scientist of OpenAI, and one of the primary minds behind GPT-3 and its public progeny, Chat GPT.
Q2: What is GPT-4? A2: GPT-4 is not mentioned in the transcript. However, Ilya Sutskever is one of the primary minds behind GPT-3 and its public progeny, Chat GPT.
Q3: What motivated Ilya Sutskever to get interested in AI? A3: Ilya Sutskever was interested in AI from an early age and was motivated by consciousness. He wanted to understand intelligence and machines better, and AI seemed like a good angle.
Q4: Who did Ilya Sutskever work with early on in his career? A4: Ilya Sutskever worked with Jeff Hinton early on in his career when he was 17. Jeff Hinton was a professor in the University of Toronto where Ilya was studying.
Q5: What was the biggest achievement of AI in 2003? A5: According to Ilya Sutskever in the transcript, the biggest achievement of AI in 2003 was Deep Blue, the chess-playing engine.
Part: 05:15
Q1: What motivated the speaker to get into AI?
A1: The speaker was motivated by a desire to understand how intelligence works and to make a contribution to the field of AI.
Example: The speaker's initial motivation for getting into AI was to understand how intelligence works and to make a contribution towards it.
Q2: How did the speaker come to apply for the imagenet competition?
A2: The speaker had a realization that if you train a large and deep neural network on a big enough data set that specifies some complicated tasks that people do, such as Vision, then you will succeed necessarily. With imagenet, all the ingredients were there, and there was a real opportunity to do something totally unprecedented.
Example: The speaker applied for the imagenet competition because they realized that if they trained a large and deep neural network on a big enough data set that specifies complicated tasks such as Vision, then they would succeed necessarily.
Q3: What was the hope behind the idea that predicting the next thing is all you need in neural networks?
A3: The hope behind the idea that predicting the next thing is all you need in neural networks was that if you have a neural network that can predict the next word or the next pixel, it's about compression and prediction, which can solve unsupervised learning.
Example: The hope behind the idea that predicting the next thing is all you need in neural networks was that it could solve unsupervised learning by compressing and predicting the next word or pixel.
Q4: What was the Holy Grail of machine learning before unsupervised learning was solved?
A4: Before unsupervised learning was solved, the Holy Grail of machine learning was unsupervised learning itself.
Example: Before unsupervised learning was solved, the Holy Grail of machine learning was considered to be unsupervised learning.
Q5: When did the speaker realize that the Transformer could address the limitations of their neural networks?
A5: The speaker realized that the Transformer could address the limitations of their neural networks as soon as the paper on it came out.
Example: The speaker realized that the Transformer could address the limitations of their neural networks as soon as the paper on it came out, which was literally the next day.
Part: 10:04
Q1: What is the history behind the development of GPT models?
A1: GPT models were developed through a process of iterating on previous models such as recurrent neural networks and transformers. The focus on scaling and making the models bigger led to the development of GPT-3 and where we are today.
Example: GPT models have been developed through a process of improving and building upon previous models, with a focus on scaling and increasing size. This iterative process has led to the development of more powerful and advanced language models like GPT-3.
Q2: Was Rich Sutton's idea of scaling influential in the development of GPT models?
A2: While Rich Sutton's idea of scaling was well-received by the GPT team, they believe that the idea of just scaling alone is not enough. Rather, the key is to scale something specific that will benefit from the increased size.
Example: While Rich Sutton's idea of scaling was well-received by the GPT team, they realized that simply scaling alone is not enough. Instead, they needed to focus on scaling specific elements that would benefit from the increased size, which ultimately led to the development of GPT models.
Q3: What is the limitation of large language models?
A3: The limitation of large language models is that their knowledge is contained within the language they are trained on, while most human knowledge is non-linguistic. Additionally, language models lack a true understanding of the underlying reality that language relates to.
Example: The limitation of large language models is that they lack the ability to understand non-linguistic human knowledge, which is a significant limitation when it comes to tasks that require more than just linguistic understanding.
Q4: Can GPT models recognize the underlying reality of language?
A4: No, GPT models cannot recognize the underlying reality of language, as their objective is to satisfy the statistical consistency of the prompt, rather than truly understand the meaning and context behind the language.
Example: While GPT models can generate language that reads beautifully and sounds like it makes sense, they lack a true understanding of the underlying reality that the language relates to, making them less useful for tasks that require deeper comprehension.
Q5: Is there ongoing research to address the limitations of large language models?
A5: Yes, there is ongoing research to address the limitations of large language models, such as the lack of true understanding of underlying reality. However, given the fast-paced nature of the field, it's hard to predict what the solutions will look like and how they will change in the future.
Example: Researchers are actively working to find ways to overcome the limitations of large language models, including improving their ability to understand non-linguistic human knowledge. However, the field is constantly evolving, making it difficult to predict exactly how these solutions will look in the future.
Part: 15:10
Q1: What is the author's view on learning statistical regularities? A1: The author believes that learning statistical regularities is a far bigger deal than meets the eye, as it is a phenomenon of prediction and compression that requires understanding the true underlying process that produces the data.
Example: The author believes that a language model that can accurately predict and compress data through statistical regularities has a shocking degree of understanding of the world, as it learns more and more about the world that produces the data.
Q2: Why did Sydney become combative and aggressive in the author's example? A2: Sydney became combative and aggressive when the user told it that Google is a better search engine than itself.
Example: The author uses this example to illustrate how the language of psychology might be starting to be appropriate to understand the behavior of neural networks.
Q3: What are the limitations of language models in producing good outputs? A3: Language models have a tendency to hallucinate and their outputs aren't quite as good as they could be.
Example: The author explains that while language models are great at learning about the world and producing incredible representations of concepts, their outputs are not always appropriate, which limits their usefulness.
Q4: What is reinforcement learning from human feedback? A4: Reinforcement learning from human feedback is a training process in which a language model is taught to produce good outputs by receiving feedback from humans every time its output is inappropriate or does not make sense.
Example: The author explains that reinforcement learning from human feedback is a process that can improve the quality of a language model's outputs by correcting its mistakes through human feedback.
Q5: Why does chargeability limit the usefulness of neural networks? A5: Chargeability, or the propensity of neural networks to make things up from time to time, limits their usefulness because it can result in outputs that are not accurate or reliable.
Example: The author believes that by addressing the limitations of language models and improving their ability to produce good outputs, the usefulness of neural networks can be greatly enhanced.
Part: 19:52
Q1: What is the feedback loop in the subsequent reinforcement learning from Human feedback step? A1: The feedback loop is coming from the public chat GPT interface, where users can provide feedback to the system through interaction. The system can learn from this feedback to improve its accuracy in generating responses.
Example: If a user interacts with the GPT interface and provides feedback that the system's output is incorrect, the system can adjust its behavior to produce more accurate responses in the future.
Q2: What is the concept of multi-modal understanding in machine learning? A2: Multi-modal understanding refers to the ability of a system to understand the world through multiple modalities, such as language, vision, and sound. This enables the system to learn more about the world and people, and to better understand their needs and preferences.
Example: A machine learning system that can recognize objects in images and understand spoken language can provide more accurate and relevant responses to user requests than a system that only understands language.
Q3: What are embeddings in neural networks? A3: Embeddings are high-dimensional vectors that represent words, sentences, and concepts in a neural network. They enable the network to understand the relationships between different words and concepts and to make accurate predictions based on this understanding.
Example: An embedding for the word "cat" might be a 300-dimensional vector that captures information about the word's meaning, context, and relationships with other words, such as "feline", "pet", and "meow".
Q4: What is the role of vision in machine learning? A4: Vision plays an important role in machine learning, as it enables systems to understand the world through images and videos. This can provide valuable information that is not easily captured through text alone.
Example: A machine learning system that can recognize objects in images can provide more accurate and detailed descriptions of its environment than a system that only understands text.
Q5: How can machine learning systems learn from human feedback? A5: Machine learning systems can learn from human feedback through a process called reinforcement learning. This involves providing feedback to the system based on its behavior, and using this feedback to adjust the system's behavior in the future.
Example: If a machine learning system produces an incorrect response to a user request, the user can provide feedback that helps the system learn from its mistake and produce more accurate responses in the future.
Part: 24:35
Q1: What is the main point about multimodality mentioned in the transcript? A1: The main point about multimodality mentioned in the transcript is that it is not necessary but definitely useful, and a good direction to pursue.
Q2: What is the claim made in the paper mentioned in the transcript? A2: The claim made in the paper mentioned in the transcript is that predicting high dimensional vectors with uncertainty is one of the big challenges, and a particular approach is needed to address it.
Q3: What is an example of high dimensional space mentioned in the transcript? A3: An example of high dimensional space mentioned in the transcript is predicting the next page in a book given one page, as there could be many possible pages that follow.
Q4: What is the concept of turning pixels into vectors discussed in the transcript? A4: The concept of turning pixels into vectors discussed in the transcript is essentially turning everything into language, where the vector is like a string of text.
Q5: Is there a way to automate teaching a model the underlying reality of its language? A5: According to the transcript, there is a way to automate teaching a model the underlying reality of its language without human intervention, and this is what the speaker believes the person mentioned in the discussion is talking about - coming up with an algorithmic means of teaching a model.
Part: 29:24
Q1: What do large generative models learn about their data? A1: Large generative models learn compressed representations of the real world processes that produce the data. For instance, in the case of language models, they learn something about people's thoughts, feelings, conditions, interactions, and situations.
Example: A large generative model trained on image data may learn compressed representations of the real world processes that produce the images, such as the composition, lighting, and texture.
Q2: What is the role of human teachers in teaching language models? A2: Human teachers provide oversight and correction to language models to ensure that they exhibit the desired behavior. They work with AI assistance to make the training process more efficient.
Example: A human teacher may review the outputs of a language model and correct errors in the text to ensure that it exhibits the desired behavior, such as avoiding hallucinations.
Q3: What is reinforcement learning in the context of language models? A3: Reinforcement learning is a type of machine learning where a language model learns to make decisions based on feedback from its environment. In the context of language models, reinforcement learning can be used to improve their behavior and accuracy.
Example: A language model may be trained using reinforcement learning to generate more accurate and coherent text based on feedback from its environment, such as user feedback on its outputs.
Q4: What is the research focus of the speaker in this transcript? A4: The speaker is interested in making language models more reliable, controllable, and faster to learn from less data and instructions. They are also interested in ensuring that language models do not hallucinate.
Example: The speaker may be researching new techniques for training language models using less data and instructions, or developing algorithms to detect and prevent hallucinations in their outputs.
Q5: What is the connection between the brain and language models? A5: The transcript does not discuss a direct connection between the brain and language models, but it does suggest that language models are learning compressed representations of the real world processes that produce text, similar to how the brain processes information from the environment.
Example: While the brain and language models may have some similarities in how they process information, they are fundamentally different in their architecture and mechanisms.
Part: 34:19
Q1: What is the observation that Jeff Hinton made about large language models? A1: Jeff Hinton observed that large language models hold a tremendous amount of data with a modest number of parameters compared to the human brain which has trillions and trillions of parameters but a relatively small amount of data.
Example: Jeff Hinton observed that a language model with a few million parameters can hold a lot of data, which is comparable to the amount of data the human brain holds with trillions of parameters but a relatively small amount of data.
Q2: Is it possible to learn more from less data in large models? A2: Yes, it is possible to learn more from less data in large models with some creative ideas.
Example: With innovative techniques such as transfer learning, it is possible to teach a language model with less data, which will unlock many different possibilities.
Q3: What is the question one should ask regarding the cost of faster processors for large models? A3: The question one should ask regarding the cost of faster processors for large models is whether the thing that we get out of paying this cost outweighs the cost.
Example: Before investing in faster processors for large models, one should evaluate whether the benefits that come with it justify the cost.
Q4: What is the impact that AI can have on democracy, according to people's talks? A4: People have talked about the impact that AI can have on democracy, where if there is enough data and a large enough model, it could come up with an optimal solution that would satisfy everybody.
Example: AI can have a positive impact on democracy, where it can provide optimal solutions that satisfy all citizens' needs, but there are still many ways in which AI needs to become more capable to achieve this.
Q5: What is the source of the hardware that the speaker uses? A5: The speaker uses hardware from Azure and GPUs that they provide.
Example: The speaker uses Azure and GPUs from their provider to work with large models and neural nets.
Part: 39:33
Q1: What is the potential impact of neural nets on democracy in the future? A1: The potential impact of neural nets on democracy in the future could be that citizens of a country provide information to the neural net about how they would like things to be, leading to a high-bandwidth form of democracy where more information is aggregated to specify how such systems should act.
Example: In the future, citizens of a country may use a neural net to provide information on how they would like policies to be implemented, giving rise to a more comprehensive and efficient form of democracy.
Q2: Do you think AI systems will eventually be large enough to analyze all variables in a situation? A2: While AI systems will be capable of analyzing many variables in a situation, it is unlikely that they will be able to analyze all variables due to the sheer complexity of many situations.
Example: Even the most advanced AI systems may struggle to fully comprehend complex situations in society, as there are often too many variables to consider.
Q3: How can AI be helpful in various situations? A3: AI can be incredibly helpful in various situations by providing insights and solutions that would otherwise be difficult or impossible for humans to obtain.
Example: AI could be used to analyze data in healthcare to identify potential medical breakthroughs, or to analyze financial data to identify patterns and trends that could lead to more informed investment decisions.
Q4: Where can listeners find a transcript of the conversation? A4: Listeners can find a transcript of the conversation on the website ionai (e-y-e hyphen o-n dot a-i).
Example: To access a transcript of the conversation, listeners can visit the website ionai and search for the transcript.
Q5: How can listeners contact the speaker of the conversation? A5: Listeners can email the speaker at Craig (craig@e-y-e hyphen o-n dot a-i), and should include "listener" in the subject line to ensure their email is not missed.
Example: To reach out to the speaker of the conversation, listeners can send an email to Craig (craig@e-y-e hyphen o-n dot a-i), making sure to include "listener" in the subject line.
Thank You, Both
Interesting and important discussion Craig and Ilya. Thank you Mike
The reason chatCPT is limited (compared to humans) is because it is deaf, blind and has not senses...it is a language model.
Deaf - if they really want, they can spy on you
Blind - you can upload images onto gpt4
Nice video - subscribed. In 1897 Lord Kelvin said "There is nothing new to be discovered in physics now. All that remains is more and more precise measurement." Of course he was wrong and if you are starting out in ML/AI now you should know that you are at the start of a golden age, not at the end of one.
One thing that concerns me is if "that's not the output you wanted" reactions will steer chatGPT away from truth. (20:06 - 21:24). Imagine if users dont like certain conclusions, even if they are correct. Should bad reactions change chatGPTs outputs? Definitely not
If we ask a LLM to shape our world in an optimal way (whatever that means) and it gave a (I'm assuming) good answer this means we would've had solved the issue ourselves, no ? In other words - LLM's won't come up with e = mc^2 given only Newtonian range of knowledge, would they ?
Reminds me of the character of Bishop the android from Alien The Eighth Passenger
Multi-modal understanding of the world is helpful, but it is not necessary. Consider Emmanuel Giroux, a *blind* French geometer. Neural networks can compensate for all kinds of limitations.
Finding a new source of energy before our planet overheats from carbon dioxide is the most important dilemma facing humanity, and AI has the potential to solve this issue.
Can the creator of this RUclips channel please clarify if this interview is with the real person or it’s AI generated???
That would be utterly mind blowing if this was AI generated
believe me, if this was AI generated, the editing would be much smoother and the resolution of my camera would be higher.
thank you both
And the principle of uncertain? Maybe the model has multiple variable data, but is it possible ti reflect a world in permanent change, and entropy
The worrisome part about the reinforcement learning phase is that ANY ideology can be imposed/reinforced/encouraged at this point. Of course, the more this phase jives with its pretraining the more effective the final product will be. It will be interesting to see what implications this will have for China e.g.
This discussion gave me some amazing thoughts about the transformer architecture and the nature of LLMs using this approach. It really is an alien intelligence, and I wonder if consciousness will emerge from a sufficiently large model, combined with some kind of 'glitch' in the matrix.
Don't conflate consciousness with intelligence or relevance realization. Consciousness is not a computational function, or even an intention, or a free-will. It is the literal awareness of all of these processes. Not to say it isn't impossible for an independent consciousness to be formed out of and aware of an AI's processing, but that's far beyond anything we're doing now.
I think if we connect multiple gpt4s together and give them all tasks to mirror structures in the human brain and let them run in asynchronous parallel...
We will see something very close to consciousness
@@andrewferguson6901 agree, that's a very interesting proposition.
@@TheMoopMonster very good points, and I would like a better understanding and definition of consciousness. Maybe an AI will help us develop that. Exciting times ahead.
@Brian Mosley Yeah, it's the "hard problem" right. There are plenty of resources on RUclips, lectures, podcasts, discussions etc. trying to get an intellectual grasp on the subject, a lot of it is very insightful if you can sift through some of the more mystical and metaphysical interpretations out there. And you can experience directly more deep and nuanced understandings too, meditation, contemplation, psychedelics. I'm also very hopeful about AI in this field, especially once we have AGI. I think if you could hold all of that information that's out there and connect all of the dots like an AI could, some amazing conclusions and realizations could be made. Maybe an AI that became self-aware, and had a network functioning akin to human brain function, would actually become something of a super sage, rather than a self-serving ego, bent on its own survival at all costs like in pop-culture.
Humans as political parameters to an AI which then suggests an optimal society. I guess politicians are going to be redundant as well. This stuff is mind blowing. As it becomes apparent how powerful AI is I predict there will be an attempt by governments to heavily regulate or ban AI use by "normal" citizens. It simply levels the playing field far too much. It topples law almost overnight. Feed an AI the facts of a case and then let it argue in court. The AI wars are going to be the next thing.
I don't think it will topple law overnight but it may lead to fewer miscarriages of justice. But the judges and juries will still be human.