An AI Drama in the C-Suite

29:27Public officials need to walk into open AI and learn the truth. You know, if this was a weapons manufacturer, and there'd been some kind of, you know, massive fiasco at their company, and there were kind of internal murmurings about a technology that's so powerful, it may soon utterly rewrite sort of the human future and potentially destroy huge swathes of humanity and the economy and whatever. You know, we'd want we'd go in there, we'd want to go in there. Why is this different? I suspect it's different because the truth is that we're not really there yet. But Sam Altman just wants to equivocate and not really say that because he knows that fueling a lot of the excitement behind OpenAI is this idea that they're just about to create AGI.

30:14They're just about to pass this incredible, make this incredible breakthrough. Yeah, we'll see. I mean, at the moment, And what they're exactly, as Mikhail said earlier, what they've been most successful at, ironically, is given their sort of founding as a research organization is not is not the creation of this technology. That was Google researchers who invented the transformer model. It's the monetization of it. Other people invented the large language model. OpenAI did do. Yeah, they did a lot to develop it and help perfect it. But they really went insane when they created when they found a way to create a really killer product around it.

30:52And that was essentially just simply to just plug it into a chat app and call it chat GPT. That's what sent everyone. That's what sent, you know, the world crazy. So it's it's marketing and product that they've actually been best at so far. Boy, for those listening on audio, we should point out that Mikhail's cat has gotten up and stirred around on that. That is a fascinating point, David, the idea that essentially what this is about is productization of the underlying technology, which did not before exist, building a front end interface, a chat bot on top of it. Mikhail, jump in, pick up on that or any other points you want to head on to.

31:31Oh, 100%. Prior to OpenAI, the transformer models, along with other types of LLMs, such as ones based on recurrent networks and long short-term memory networks, were primarily used for writing marketing copy. And they weren't particularly good at it. One of the things that made OpenAI's GPT project so compelling was the sheer amount of data that they blasted into it. the degree to which they spent effort training this thing and building the sheer size of the network was unprecedented. And it was largely an experiment to see, will it continue to get better as you grow the network and feed more data into it?

32:24And it turns out the answer was a resounding yes. I don't think that they necessarily knew that going in. We've certainly, neural networks have been around for a very, very long time. I can talk about their history at length. But the point is that it has not always been the case that if you just throw more neurons and more data at a network, that it'll automatically get better. And in this particular case, it did. And that's why they're on the map now. It's also worth pointing out that prior to getting big with GPT, my own experience with open AI. And I think a lot of other developers experience was with, and I could be facing hallucination here, but they had a project called the open AI gym, which was a really cool little toy where you could basically, they would pit little neural network agents against one another in various games, like capture the flag or tag or battleship, or, you know, just like little little just toys right and so the idea is that uh programmers could write their own ai agent submit them to this gym uh and just have it run autonomously like hundreds of thousands of times millions of times with other opponents against other opponents and then like learn and get better so they were already like playing with the the idea of exposing ai to the general public in little incremental like cutesy ways before they, you know, ran this experiment.

33:53And now GPT is all that anyone knows them for. David, you want to touch on that? I mean, yeah, I think that I think I think it's instructive to look at the recent history of Altman inside the organization and what he's been really good at and what he's been really good at is putting OpenAI front and center of a broader story and a story that started somewhere else and that is developing in all kinds of ways, in all kinds of directions. And yeah, OpenAI need vast credit. I mean, GPT-3 was a huge step up in terms of large language models and their competence. And then, but no one, people were excited, but people in my world were excited.

34:41It's not like the world was excited until they created ChatGPT. And just that word chat really set people on fire and this idea that you could talk to this thing and it was very natural. It was just a route in for millions of people and a really intuitive way for them to understand what this technology was about. And it's clear that there is some internal tension, to say the least, in OpenAI about, you know, who are we really? What's our soul as an organization? How do we proceed from here? Do we want to just kind of move faster and break things? Or do we want to do what we originally said, which is be socially responsible and be primarily about how do we develop this technology?

35:27How does the world develop this technology in a way that's safe? And we're just going to have to keep watching. I mean, it's fascinating. It feels now that the conversation around safe AI is up for grabs in a way that it wasn't before, because open AI are kind of stepping back from they're not step they're not stepping away from that conversation. But they're clearly not going to be the the organization that is at the heart of that conversation in the way they were originally intended to be. So you see in the last couple of hours, Meta has just announced, along with IBM and loads of other partners, what they're calling the AI Alliance, which is this big organization, big new organization, to have an industry-wide conversation, including academia and policymakers and all the other stakeholders, about AI safety.

36:20That's a clear attempt to step into the positioning that we all once thought OpenAI would inhabit, but they seem to have stepped away from. Because that's a very influential conversation to own. It's a great marketing conversation, and you get to shape the regulatory environment, which is another thing Altman and OpenAI want to do. They want to shape the regulatory environment to consolidate their lead. And that's why that conversation is so valuable. Whenever one company in a particular sector gets very powerful, it always seems as though their competitors form a consortium to have a conversation around what they're doing.

36:59It's like this interesting thing that we just see time and time again in business. It also plays to OpenAI's financial interests to create regulations that didn't exist when they were rising in order to sort of lock the door behind them now that they've made it up into the upper echelons. Let's make sure that others can't get there too. So like, let's, you know, I tend to be quite a cynic about these things. Let's not talk about these whole like safety concerns as if they're anything more than, than, you know, rent seeking an empire building. Like OpenAI wants you to believe that they're, that what they're doing is extremely dangerous so that you will beg your politicians to prevent other people from getting to the point that OpenAI has already reached.

37:46We're going to take another quick break and be right back with more of the day's top analysis on the Real Vision Daily Briefing. Join over 5 ,000 attendees for the largest AI event in Asia, Super AI in Singapore, June 5th and 6th, 2024. for. Edward Snowden, Benedict Evans, Balaji Srinivasan, and over 150 others will hit the stage, joining the industry's most influential to explore and unveil the next wave of transformative AI technologies. Singapore will become a vibrant AI hub for a full week from June 3rd to the 9th, with over 150 side events that will make for unparalleled networking opportunities.

38:33Visit superai.com for 20 % off tickets with the code REALVISION. Look for the link in the description. This sounds like it could be like a Netflix comedy series. The three of us should go on a global tour to talk about this powerful black box that we've developed that we're ultimately going to IPO. And we're going to beg the world's leaders to constrain the industry around us because it's just too powerful. I'd watch that show. I mean, Netflix should pick that up. Yeah, that's a great drama. We've seen this precedent in, I believe, season three of Netflix. Sorry, season three of South Park. Cartman gets himself an amusement, gets a million bucks, uses it to buy an amusement park, and runs ads on local TV saying, like, there's this really awesome amusement park and you can't come in.

39:21It's got, like, all these rides and only I get to ride them. Anyway, the chief question that I think hamstrings the entire discussion of safety in AI is that on a technical level, I am not sure that the concept of restraining a sentient being through some kind of behavioral constraint bolts or anything like that makes sense. I don't think that a being that's capable of self-modification is also capable of being fitted with the kind of parameters that will prevent it from breaking out of some kind of corral. especially not if it's smarter than us especially not if it's seeded with an entire corpus of sci-fi stories that human beings have written about how such machines break out and the strategies that they take for overcoming their own uh their own boundaries i love it so we're gonna have to we're gonna have to make sure that the machines don't get a hold of you know 50 years of dystopian science fiction around artificial intelligence things already do like we uh we already trained GPT on an enormous corpus of literature, which includes, for example, Isaac Asimov's Robot series, which is all about how these robots are built with inherent constraints, like fundamental to their entire existence.

40:50And yet they still are able to find ways to creative, I shouldn't say circumvent the constraints or like escape the constraints, but rather adhere to them in very creative ways that end up completely undermining the existence of the constraints in the first place. And it's like I said earlier, any sentient being is going to be capable of equivocation. If you tell it not to kill people, it'll redefine what it perceives as a person. If you, you know, if you tell it to like not steal, it'll redefine its understanding of property and theft. Like, you know, you're not going to hold these things back any more than you can people.

41:25By the way, this is literally right out of Isaac Osmoff's Three Laws of Robotics that we're talking about. But let me just throw one thing out here. And I want to talk about Q learning, reinforcement learning, so-called model free learning, because there's a lot to talk about here in terms of the technology. But let me just throw this out there. Effective altruism, it's a word, a phrase, a buzz kind of phrase that's been in the news a lot lately. Do you guys put any stock in it at all? Do you care about it or do you think it's a distraction?

41:54Go ahead. Yeah, I mean, I'm very interested in effective altruism. And obviously now you have the whole effective accelerationism kind of movement. By the way, let's define that effective accelerationism because in many ways it stands in opposition to effective altruism's view of AI technology. Yeah, exactly. I mean, you know, accelerationism is a movement, is a philosophical movement with really interesting kind of roots of its own that go all the way back to, I mean, you know, certainly to the 90s. And you could you could do a genealogy beyond that. But, you know, in the 90s at Warwick University, a British philosopher called Nick Land did a lot of work around the early Internet and sort of cybernetics and what a connected world meant.

42:43And obviously in sort of 1992 in rainy Britain at Warwick University, these were incredibly exciting and outlandish ideas. And most people would would would would plot a line when it comes to accelerationism back to kind of Nick land and that that community at Warwick University. And a sense like this is it gets very thorny and very contested very quickly. But accelerationism is essentially about saying it's futile to attempt to hold back the intensely powerful machine that is techno capitalism. It's futile to try. We shouldn't try. Our destiny and the destiny of the world of the universe essentially is technology.

43:32technology and all we can do is lean into that force and accelerate the processes of technological change. And this new kind of effective accelerationism movement is about saying that and about saying, how can we do it most effectively? And I guess, how can we do it in a way that most benefits human beings? So it's about saying, contrary to the popular mainstream position at the moment that technology is running out of control and we need to rein it in and it's becoming too powerful and it's dominating our lives too much. The real path to utopia, to a better future, is to massively accelerate the forces of technology and essentially of capitalism and get to some kind of tech-fueled utopia.

44:18That is a caricature to a certain extent of the position, but I hope it gives a rough idea of the position. Effective altruism is much more skeptical around technology. I mean, it has a nuanced position around technology, but effective altruism has its roots in saying, let's bring evidence and science and reason to our thinking about altruistic work and charity work, essentially. Let's bring real hard evidence and deeply kind of researched arguments to our beliefs about what we can do to make the world a better place. Because we hope we're all trying to do good. We're trying to make the world better.

45:00But often our efforts are not evidence-based. They're not very effective. They don't really work. We waste a lot of money. How can we avoid that? That's what effective altruism is about. Michael Voloshan, did you just reposition your camera to better feature the one sheet for your 2016 science fiction novel, Dopamine? It actually features it a little bit less effectively at this point, from my view. But now that the cat escaped, I figured I would show off a little bit more of my bookshelf, where the astute observer will notice a series of Warhammer books, which they should infer would put me solidly in the effective escalationism camp, where I do believe that the sooner we can give rise to the machine god, the better and quicker our salvation will come.

45:50Praise the Omnissiah, the Omnissiah saves. The further observer will note that this is Warhammer fantasy rather than Warhammer 40K, but I'm not going to get into the detail on that. By the way, it's official. We've reached the point in the conversation where the news flow component is over. We're just diving into pure nerddom now. If you're not a member of that, go to another YouTube channel because we're going deep right now. Something that we wanted to talk about since the beginning of the show, something that I know that you're passionate about, Mikhail and David, that you are interested in, is this idea of Q-learning.

46:23What's happening with the algorithms right now? Because there are many people who believe that we're at this point where we're about to experience a significant shift in the level of what this technology can do in the near term. So if you don't mind, David, I would love to field that first, because I've been seeing a lot of misconceptions floating around about the QSTAR algorithm and about its relationship to a lot of other things and what it can exactly do. So first of all, the QSTAR algorithm is not directly related to Q learning. It is, it unfortunately also has the word Q in it or the letter Q, but it's not related to deferred reinforcement learning.

47:10And it is not, like, it's a completely separate class of algorithm. them. It is also not related, as some people have said, to quantum computing, though the intersection between quantum computing and artificial intelligence is itself a very fertile field for discussion, which we won't have time for today, obviously. Let's try to just put some rough definitions around this, just to give people who are interested a sense of what these technologies do and why they're different from what we have today? So the QSTAR algorithm is a theorem prover. It was not developed at OpenAI, at least if they're using QSTAR to denote the same algorithm they're working with internally as what academic papers have used QSTAR for.

48:02But what does it do? What does it mean? It's a theorem prover. If you give it a bunch of logical statements, like, you know, this exists, and this implies that, then that must be true. It's, if you feed it with a gigantic database of facts and relationships, it can infer new, it can infer new, I don't want to say facts, but if you feed it with axioms, it'll derive theorems from those axioms and the relationships therein. How about this? Two questions. Number one, how is that different from what we have today? And number two, what can it do in terms of functionality that might be more enhanced or advanced than what we have today?

48:52So the QSTAR algorithm by itself has been around since 1973, and theorem provers are even older than that. So it's a branch of AI that was developed during the quote-unquote AI winter, and it's been known to be able to – it's a pretty powerful theorem prover as far as they go. But it's been part of our world, and we haven't had much to do with it. Um, so you can like prove new, uh, you can, uh, prove mostly already proven theorems and mathematics, but that's kind of a, not a lot of people care, um, outside of a very esoteric bubble. So what happens is when you rig it to a neural network, you can solve...

49:40Basically what happens is there's two kinds of logic, induction and deduction. Theorem provers are great at deductive logic, but deductive logic is incapable of establishing truths about the world. It can only establish truths within a closed axiomatic system. Inductive logic can determine truths about the world, but it can't do anything with those truths. And neural networks are really great for performing induction. So the idea is that if you have a neural network doing induction in order to learn truths about the world, and you have a deductive module powered by something like Qstar that is able to work with the truths discovered by the inductive module and then derive subsequent implications from them, now you've got something that actually is capable of observing and reasoning and drawing conclusions.

50:35David, want to jump in? I'm lost. Yeah, I mean, I think it's an excellent description. And in short, it's about taking a step closer, just as Mikhail said, taking a step closer to a kind of machine intelligence that is like human intelligence. We can walk around the world, learn things about the world around us that are contingent. They're not necessary truths. They happen to be the case. The sun rises every morning. When I'm cold, I shiver, whatever it is. We can gather that kind of information from the world. We can also reason from first principles. we can be given a set of exactly axioms and then we can do closed reasoning that allows us to deduce new truths based on those kind of simple axiomatic truths we've been given if you can have a machine intelligence that can do a bit of both of those things you just have something that's more akin to a human intelligence No one's saying it would therefore be a human intelligence or as powerful as a human intelligence, but you've taken a step closer.

51:57So it's easy to see on the face of that why some people in open AI might believe this is potentially a huge breakthrough and we're not treating it in the way our initial charter or manifesto said we should treat these kinds of breakthroughs. yeah but I mean it's not clear to me I mean Mikhail we're not totally sure are we that that's what QSTAR refers to inside open AI this again is stuff that's been inferred by you know extremely knowledgeable machine intelligence people like Mikhail with years of experience and in in the academic literature and all of that but no one inside open AI has really come out and explained this, which I find startling because if you guys, I mean, the organization is still called OpenAI.

52:47If you want to be remotely transparent, and if you really are so concerned that this technology is so powerful and needs to be regulated and might destroy us all, do you want to tell us what happened with this whole fiasco? Do you want to tell us what Qstar refers to? They haven't said anything as far as I know. OpenAI has hundreds of employees. So if there's something absolutely earth shattering in there, the idea that there are 700 employees that just signed the petition to reinstate Sam Altman, that's just the Sam Altman faction. That's not even counting the rest of the employee base. Now, not all of them are developers.

53:27A lot of them are stuff like RLHF trainers and publicists and whatever. But that's not the point. The point is that, you know, if you really think that all of them are keeping some sort of absolute breakthrough secret, then you probably believe that the moon landing was a conspiracy too. So I believe that it's much more likely that they are marrying a fairly conventional implementation of QSTAR with their existing plug -in system for GPT that allows for GPT to access a deductive module in a way that's very similar to the way that it accesses a calculator. or the web nowadays. Guys, what might this technology do?

54:19Like how might it actually impact its applications? What types of problems might it solve? What types of functionality could it deliver to end users? I still don't have my head around that. If you give it a series of facts about it, like just tell it in plain English, a series of facts about like either yourself or your company or some situation in the world, it can assemble those facts into, like behind the hood, it'll assemble those facts into symbolic logic and use, and like figure out how they all come together. And then it'll give you an answer about a prognostication about your situation or your company or whatever that's likely to be, that's informed by reasoning rather than by off-the-cuff inference.

55:13So, Mikhail, can you make up a toy example for what this might be?

55:20I mean, the oldest syllogism known to man, you know, if I tell it Socrates is a man and all men are mortal, it'll be able to tell you that Socrates is mortal. So the point of the neural network is to find patterns in the world. And the point of the deductive module is to figure out how those patterns come together in order to predict the future state of the world. It's a very powerful marriage, but it's ultimately completely dependent on how exactly you expose this to the real world in general. Here's a toy example that I can think of. It might find that every single time a company, like, you know, it might find that a certain, that every time a company in a certain industry makes a certain kind of announcement that, like, its stock dips.

56:16So it'll be able to reason that if the, you know, and then every time its stock dips, then it can infer that such and such hedge fund loses money or such and such other consequence happens. So it'll be able to piece together that two-stage thing that the company, it'll read the news, it'll see the company makes this announcement, and then it'll be able to infer, okay, well, that means that the stock will lose money. And then that in turn means that this hedge fund will dip. That in turn means that so-and-so person will have less money. That in turn means that, I don't know, their mistress will not get a good Christmas gift and the best is, I don't know.

56:54It can perform this chain is what I'm getting at. Right now, GPT performs the illusion of being able to do such chains of reasoning, but the number of links in that chain is very limited. Ash, you and I had a talk where I demonstrated that the GPT can reasonably find its way around a small map of a house, but it gets lost in a very long hallway when it's hooked to a navigation system. So this new system would not have that constraint. It would actually be able to reason across many, many different domains and be able to draw logical conclusions across many, many steps of connection. Yeah, yes, exactly.

57:41I mean, I think what people, and again, this is something of a simplification, But what people really need to understand is that, you know, the large language models we're using at the moment are trained on the deep statistical relationships between words as they're commonly used by us, as they're commonly used in sentences. So fundamentally, this is a machine intelligence that is very good at sounding right. If you say something to it, it will come back to you with the kinds of things people typically say in response to that. But it has no idea and is not capable at all, really, of reasoning about whether it is truly right or not.

58:27There's kind of reasoning instantiated in the way we use language. And because it understands those statistical relationships, it can kind of reproduce that reasoning. It can kind of simulate it. But what Mikhail's describing is the beginnings of a large language model that does that, but it also has true actual reasoning capabilities. So it's the promise of an AI that, for example, a theoretical physicist could sit down with and say, look, I'm working on this new model of quantum gravity. You know, there's this component, there's that component, you know, the maths here, the maths there. But when I move the model in this way, it just doesn't make sense.

59:10And it won't just fire back with sort of statistically likely words that sound like they could be part of this discussion. It will actually be able to truly reason and deliver the outputs of that reasoning. So it would be like having a calculator for reasoning, if that makes sense. And that's something new and very, that is very exciting. But I totally agree with Mikhail that, look, if this was some earth-shattering, dangerous revelation, just to go back to the kind of politics and the fiasco of it, it's ridiculous to think that all 600, 700 people are keeping that a secret. One other thing that I find interesting about it is, look, you can see the really, really interesting directions research is going in.

59:55You know, Mikhail talked earlier about the realization, the breakthrough, and it really was a breakthrough, that if you take, if you train these transformer models on vastly more text, they become much, much more powerful and competent. And that was really the huge breakthrough between GPT-2 and GPT-3. No one was talking about GPTs and large language models with GPT-2. GPT-3 was a huge breakthrough. Suddenly it sounded like a person. It was writing short stories like Oscar Wilde. Everyone was excited. And that was achieved simply by training it on a vastly bigger data set, much, much more language.

1:00:36There is concern and certainly the thought that we're reaching the limits of that. We don't know yet, but will it be the case that an even bigger, bigger, bigger data set is going to lead to another quantum leap in competence? Lots of people think probably not, but that there's many competence gains and performance gains and all kinds of exciting stuff to be had around this kind of work. You know, combining this kind of AI with other kinds of AI and other kinds of models that just expand the competence of the AI more broadly. And I think it's interesting that Qstar points in that direction. OpenAI are clearly looking at what can we do to make this thing better beyond just an even bigger data set.

1:01:25It's worth noting that the human brain is not just one thing that's built up to X, billions and billions. We have a language module, our Wernicke's and Broca's areas, combined with a visual spatial module, combined with a reasoning module, combined with a whole bunch of other things. And so it very much makes sense that if you want the computer to reason well, you don't just build a bigger neural network and try to hope that reasoning capabilities will emerge magically out of it. You integrate it with a known system that's been around for multiple generations, not just decades, but generations that's already been very, very good at logical inference or deduction, rather.

1:02:12And you just connect it to that. Guys, extraordinary conversation bridging the gap here between neuroscience and artificial intelligence at the end. Really incredible. This is the place to have these conversations on Real Vision. We just couldn't appreciate you more joining us. Thanks so much. Thank you so much, Ash. It's always so fun dropping in. This is always a blast. And I look forward to next time when there'll be even more developments, more excitement, and we will get ever closer to summoning the machine god. Wow. What a show, guys. Thank you so much. And thank you so much for watching.

1:03:00haven't dared try? With Plus500 Futures, you can trade crypto without the hassle of opening a wallet. With just a few clicks, you can register and start practicing with their free and unlimited demo. See a trading opportunity? You'll be able to trade it in just two clicks. Feel ready? You can move to real money with as little as$100 once your account is approved. And the great thing is that in addition to crypto, Plus500 gives you access to a wide range of instruments. S &P 500, NASDAQ, gas, and much more. Explore equity indices, energy, metals, Forex, and beyond. With a simple and intuitive platform, you could trade anytime, anywhere.

1:03:35Experience the fast, accessible futures trading you've been waiting for with Plus500. With over 20 years of experience, Plus500 is your gateway to the markets. Visit us.plus500.com to learn more. Trading in futures involves the risk of loss and is not suitable for everyone. Not all applicants will qualify. Plus500, it's trading with a plus. Thank you.