Keynote: The Future of AI - Avishai Abrahami and Prof. Amnon Shashua

My pleasure.

It was 1998. I had a company at that time called CogniTens which developed optical sensors for building accurate three-dimensional models used in car production, and Toyota was one of the customers. And one day I get a request to come and meet their future division, and they told me it's not about CogniTens, it's about your role as a professor or your expertise in computer vision. Okay, I said okay. So I meet their people and they tell me, look, Subaru is coming out with a stereo system that will maintain safe distance to the car in front, what is known today as ACC, adaptive cruise control. And since your area of expertise is 3D, at that time my PhD was about three dimensions from images, and we would like you to consult us and we'll pay you. Don't consult us, I told them, you know, when I shut one eye I don't go blind, right? And then I went through a kind of a lecture, you know, why we have two eyes, right? Because if you really want to measure 3D by triangulation 100 meters away, you'll need a baseline of meters, right? A baseline of 10 cm is not good for triangulation long distances. We use triangulation only for very short distances, to thread a needle for example. But in order to know whether a car is 50 meters away or 80 meters away, you don't use triangulation. So I went through this and I told them, no, monocular is all what you need, you don't need stereo. They said, okay, we appreciate the knowledge that you kind of gave us, but we don't think it will work, we don't think it's going to be robust. We want to build the stereo system. And then I had a kind of a stroke of genius and I told them, you know what, give me a small budget and we'll do some demonstration. At that time, 1998, money was flowing there in Japan, there was no problem, money, and they gave me on the spot $200,000. How much? $200,000 on the spot, okay, almost all right. $200,000, you know, today when you raise money for a company that's nothing, but still, you know, $200,000. So instead of taking that money and putting it in my pocket, I went to my close friend Ziv and I told him, look, this could be huge business. The entire industry is looking at it the wrong way, right? It's really what you need is a monocular camera, and eventually because the cost would be very, very low, it's only one camera, and the performance would be better, although it's not intuitive why one camera would be better than two cameras. The performance would be better and this eventually will be standardized and will be on every car. So we took that $200,000 and this is how we started the company. And that vision really came to fruition. You know, these systems are in every car, they prevent accidents. And then in 2012, 13, with the rise of deep learning, I came to the conclusion that now the next level is autonomous cars. And this is where we started kind of shifting and putting more and more resources on solving the problem of autonomy. Today we spend more than $600 million a year just on building the technology for the next decade.

So I would say that each company has its own culture, but what I can talk about is the culture of the company that I actually run, which is Mobileye. So I only run one company, all the rest I'm a founder, chairman, and so forth, but I'm not running the company. Only Mobileye, which is 4,000 people, smaller than Wix, but still 4,000 people is a mouthful. So I can make observations about culture in a company like Mobileye, and I also had some observations while I was an executive at Intel, kind of to look at a much bigger company, much bigger corporation of let's say 100,000 employees. And I think those observations are kind of catching really the essence. So on one hand, Mobileye, you know, we're developing cutting-edge technology, but on the other hand, that's 4,000 people. So how do you maintain agility and innovation when you have 4,000 people? So it came down kind of to three principles that we try to maintain all the time, which talks about culture. The first is sense of purpose. What I found out is that if there isn't a very strong sense of purpose, people don't focus on the product, they focus on self-advancement. And this is, you know, corporate culture tends to shift the attention from product to self-advancement, you know, what do I need to do in order to advance myself within the organization rather than what do I need to do in order to make the product better. So sense of purpose allows us to maintain this focus on the product. The second observation is about truth, truth-telling. So in corporate, truth-telling is compromised. Truth gets compromised as you go up the managerial level. So each manager tries to hide the bad news and report only the good news upwards. So eventually the CEO doesn't have a good grasp of what's going on downstairs, right? And this is kind of a disease in all corporates, there isn't a sense of truth. So in order to kind of get a better chance of finding out the truth, now we kind of encourage and insist that managers skip managerial levels, that you need to always sample employees very, very down the ladder and not only get your information from those who report to you. And also as a CEO, the number of people that report to the CEO is much more than the C-levels and executive VPs, also technical leaders that are not executives, in order to get this kind of information. And every employee in the company knows that managers sample employees, especially the CEO, and eventually at some point he or she will be sampled. So I shared that belief and I think that, and I have a lot of people that directly report to me, way more than is executive, but it actually works much better. And I also work a lot with people that are not reporting to me or reporting to somebody who reports to somebody who reports to me, and this is very, very important. And the third part is cultivate leadership, not just being a manager, but cultivate leadership. And here also we found that there are three components that define leadership. One, you need to be intelligent, because if you're not intelligent, the people that report to you could game you and manipulate you. So this is basic, you need to be intelligent. Not everything in life you need to be intelligent, you know, there are politicians that you would not say that they're intelligent, right? We have a great example happening live today. So intelligence is not always a necessary condition, but if you want to be a leader, you need to be intelligent, that's basic. The second one is courage, it's the courage to change, to make change and change the status quo. So what people mostly do is accept the status quo and they don't have the incentive to change it, even if it's a bad status quo, because people adapt, they get familiar with the situation and they don't go and change it. Leaders are people who kind of challenge the status quo, and you want people to challenge the status quo, otherwise you're in stagnation. And third is to be able to kind of generate a vision, articulate a vision. So those are kind of the principles of a good culture that over the years I found out that this is what I want to see. And this is especially important when the organization is big. The word organization, it's small, it's a family, everybody knows everyone else and you handpick the people. When you have a corporate, then things get out of control very, very quickly.

What I think is the next step. So in order to understand this question, I think it's important to first internalize what are the components of a good AI system. And I believe that there are four components: knowledge, skill, precision, and intelligence. Now, knowledge and skill is what we have today. Language models of today are knowledgeable and are skillful in every downstream task that you focus on. You know, give it sufficient data for that downstream task and it will be skillful. Play chess better than humans, play Go better than humans, but that's skill. Precision is knowing what you know and admitting what you don't know. Language models are not good at precision, right? They don't know whether the answer they're providing is correct or not correct. Even if the answer is code, they don't know whether the code is correct or not correct, right? So precision is still missing. The fourth one, intelligence, is tricky. You know, there's a seminal paper back in 2007 by Legg and Hunter, I think, they compiled a list of 70 definitions of intelligence. So it means that it's an ill-defined concept. But as computer scientists, they have a definition of intelligence, and the definition of intelligence is about generalization. Generalization is the ability to figure out a rule based on methods, tricks, tools that you have acquired through your profession, define a rule that will fit observations, and observations could be input-output pairs, right? The problem with language models, and true genius, they can develop new tools, for example, Newton developed calculus, calculus didn't exist, right? But let's not talk about true genius, just the ability to find a rule that will fit observation. For example, Kepler, right? He made lots and lots of observations and he kind of compiled them into three rules, right? So this is the definition of intelligence from a computer scientist's point of view: generalization. The problem with language models of today is that they don't abide by this definition, and this is a bit tricky. During training, the systems learn based on this definition of generalization, but during inference, there is no learning. Now there was a hope when GPT-3 was introduced, there was a hope that few-shot learning is a way to do learning during inference, but today we know that this is not correct anymore. So what is happening with today's language models is that they're using only search. And during search, instead of fitting a rule to observations, they have kind of an internal reward, a neuronal reward for intermediate steps. This is what o1 is doing with these chain of thoughts. Now, this is a bit interesting because if you look at AI in the 80s, it was rule-based symbolic AI. Then in the 90s, machine learning, statistical learning has emerged with this idea of generalization, probably approximate correct framework. And now it seems that there is a shift towards search, because what these models are doing is purely search. There's no learning in terms of generalization. Now I believe that inference without learning will never reach the stage of true expertise. Because if you look at language models today, it's like a mediocre assistant. So say for example you give it a task, now here's a problem, write code for this problem. It'll write code and then there's a mistake, so you give it feedback, then it will correct a mistake, then you give it another feedback because instead of finding an n log n complexity, it found an n squared complexity, you give it another feedback and another feedback, right? Eventually you come to the point where you say, I can write it myself, why do I need this lazy assistant that I need to give it feedback again and again and again? And this is because there's no notion of learning. So I believe that there's another breakthrough missing where learning would be embedded into the inference stage, and only then we can see the next step of building AI that reaches the levels of experts and not just a mediocre assistant.

So I'll divide the answer into two. First, I just want to be clear, everybody here is a developer, they're very worried. Yeah, yeah, yeah. So okay, so I'll tell you first, at Mobileye we did a test of using Copilot, right? By the way, today we're not using Copilot, there are open-source alternatives that are even better than Copilot and they cost much less. So we're not using Copilot anymore, we're using open source. But we did a test on using Copilot and what we found out is that it introduces more bugs than a regular developer. And the amount of time it takes you to discover those bugs, and some of the bugs are really very, very subtle, because it looks great and the syntax is correct, and the time it takes to remove the bugs is more than the time it takes to write the code, right? So it actually slows down development, because again we're talking about production code, right? Not something that kind of works, right? It has to be production code, it's in cars, it has to work perfectly and so forth, right? But now, to your question, well, a year ago I founded another company which is still in stealth, because I don't have 100% guarantee we'll be quiet about it, you can just share everything here. It's called Double AI. It's three of my former PhD students, my colleague Shai Shalev-Shwartz, you know, professor at the Hebrew University, is also CTO of Mobileye, we are together for more than 25 years, and it's about seven employees to tackle exactly this problem. How to build, so instead, I don't subscribe to AGI because I think the G is ill-defined. I do subscribe to AI expert. So the kind of question, I remember about a year and a half ago I gave a kind of a lecture at IDC, today it's called Reichman University, right, at IDC about language models. So it was just the beginning, ChatGPT came out, and Gil Kalai was in the audience. So Gil Kalai is a mathematician, and when I finished he told me, do you see any time in the future where this kind of technology can be a great mathematician, can be a great physicist, a great scientist? I said I don't think so, right? I don't know, but it kind of lingered in my head that maybe you can. So this is what Double AI is trying to do, especially in the coding area. And what is missing in order to, you know, give it a problem and getting a solution that is correct, solution means a piece of code that is correct, there are two pieces. One is the precision that I mentioned before, right? Because take for example o1 and give it some problems from Codeforces. So Codeforces, everybody here knows Codeforces. You give a problem from Codeforces, say take 100 problems from Codeforces. It's a bit tricky because if it's old problems, then during pre-training, you know, you already got the problem, the model already learned the solution. Yeah. But let's assume you take new problems which are after the cutoff date of training, and you'll see, o1, which is the best, will solve 20% of the problems, all right? But the other 80%, it doesn't know, it didn't solve it, will give you simply the wrong solution, right? So this is not useful. So you need precision, not only recall. You need to say I don't know how to solve it, that's okay, but if you do give me a solution, I want guarantees that it's correct, because for me checking the solution will take more time than writing the solution. So first is the precision, how to build precision into the language model, this is one task what this team of seven people are doing. And second, this definition of intelligence that I mentioned before, how to create generalization during inference, because this is how an expert works, right? You receive a problem, the first thing you do after you understand the problem is you write down a number of kind of input-output examples, kind of toy examples, so you can start thinking of it, and then you try to discover a rule that will fit those samples, right? So this is exactly the process of generalization, right? But you do it during inference. So those are the two areas that the company is working on. I cannot guarantee now that there will be success, but it looks positive, it looks very, very promising. So hopefully maybe in a few months I can reach to a point where I say yes, we have found out the right thing, and then really go and raise money.

Well, you know, you can take this to an extreme in a brute force case, and this is what AlphaCode by DeepMind they try to do, right? By sampling 1 million times. So say you have a problem, give me a solution, but 1 million times, give me a solution, okay? And then they try to kind of sift through those 1 million solutions. But again, this is still not having the precision and not having generalization, right? It's just having brute force. And eventually, they didn't get to really true expertise, right? So I'm sorry to say we still have to work next year. Yeah, okay, at least it sounds more than one year. And by the way, this is going to be much harder to find AI that may invent calculus, right? We are very far from that. Yeah, definitely.

So kind of the real question is why robots now and not five years ago? What happened that today you see startups, you know, working on humanoid robots? So I think the demand was always there. There are people working in fulfillment centers and assembly plants, production plants, you know, very harsh conditions. Take for example Amazon, they have 1.3 million people working in fulfillment centers and the turnover is 100%. That means every year this entire force of 1.3 million people are replaced because the job is so boring that people cannot, you know, stand it for too long. So the demand, the need to replace labor that do kind of mundane work always existed. The problem was that there was no solution. How do you build a useful humanoid robot? So useful, by useful I mean it should have the dexterity of a human and it should be able to follow instructions like a human. In order to follow instructions, you need a certain level of intelligence, you need to understand the instruction and then you need to follow it. So what has changed in the past few years? I think there are three things that have changed. One is compute. You know, today compute is way, way stronger than five years ago. Second is language models. Following instructions is exactly what language models do. Instructions, the prompt, that's what we know how to do. We know how to do most of the time, most of the time. Sometimes the prompt tends to argue with you, but it seems that you can give it a prompt and it will break it down into steps, and you can then do fine-tuning on it, reinforcement learning on it. You know, this technology is there since 3, 4 years ago. And the third one, which is something that people are less aware of, so the first two everyone is aware of, the third one is called sim-to-real. It's how to, so training a robot in a simulator is relatively straightforward because you can generate infinite amount of data, right, in a simulator. So you can use data-driven techniques like reinforcement learning in order to train the robot to walk, to grasp, to do everything. The problem always was that the simulator is not rich enough to reflect the real world, and then when you take this policy that you developed in the simulator and put it in the real world, it simply doesn't work. And then a few years ago there was some advancements in academia, and it's called sim-to-real, where you inject noise during learning, and this noise injection gives it robustness to the missing parameters in the real world. And then in the real world, the amount of data that you need would be very, very small in order to just compensate for the move from the simulator to the real world. And also simulators have improved, right? If you look at the NVIDIA simulators, the physical, I think the Omniverse 5.5, if you've ever seen that, looks like reality. Exactly, better. So these three things happened in the last five years which make everyone in this domain optimistic that now is the time to build a humanoid robot. And this is why two years ago I founded Mentee Robotics and this is what we are doing.

Okay, and I won't be offended, I respect your time.

So it's called AI alignment, very, it's an alignment issue. So AI alignment means can you align an AI to conform to certain values? And AI alignment says that you cannot. Why? Why you cannot? It says that let's assume you have a reward function and you're optimizing this reward function, and let's assume that the engine of optimization is super sophisticated, super human, okay? Then what happens is that this engine will find, it's called a corner in solution space, it will find a solution that you did not anticipate, and there's no way around it. So I'll give you kind of a down-to-earth example. Let's assume that I have a chatbot, and quite a good chatbot, and I use it not to answer questions but for interaction. Many people are bored, so it's a good interaction if it seems intelligent and you interact with it. Now let's assume that I'm the company who released this chatbot and I want to improve it. So in order to improve, I need to define the reward function. So let's assume that I'm altruistic and the reward function that I put there is make people happy. So during the interaction, you know, elevate their happiness. Okay, so I have very good intentions, right? I'm the engineer who designed the reward function and I put many, many parameters, kind of creating guardrails and so forth and so forth. But this is the reward function: make people happy. Now let's assume that the system that is optimizing the reward function is super, super, super intelligent. So now this system has one goal: elevate the happiness of people. So now this system might find out that if you lower people's IQ, they're more happy, right? Less worries, yeah, you're more happy, right? So what it will do, nobody will understand what this machine is doing. So through interaction, you know, why do you need to work until 5:00? Work until 3:00. Why do you need to do five-hour mathematics? Do three-hour mathematics, you know, all sorts of things, right? And then a generation later, when you have a generation of dumb people, only then you'll understand what this machine has been doing for the last 20, 30 years, right? So this is an example of a corner in solution space, right? Because when you design the reward function, you did not anticipate that this is right. So now AI alignment says, let's assume that you anticipated and you put in your reward function a negative component of lowering IQ, but it will find something else that you do... What you're saying is that there is no chance. So the answer here is that you cannot control, the answer is that you cannot control. This is called the AI alignment problem, and I wrote a paper about this four years ago proving that you cannot control. But this does not mean that we are approaching doomsday, right? Again, if you subscribe to this idea of an AI expert, right? So you have an expert in physics, an expert in mathematics, a great thinker in psychology and in medicine and so forth, this does not mean that this expert is self-motivated, has its own goals, and is trying to achieve its goal. It doesn't necessarily is the case. People are imagining this AGI, which again is ill-defined, they're imagining kind of a being that has self-motivations and it's building its own reward function and trying to optimize it through interactions. This does not need to be the case. So you can build intelligence in terms of a collection of experts that will not destroy humanity, right? And by the way, every technology can be abused. I'm not saying that it will not be abused, but AI alignment means that there is an abuse that you did not anticipate, and this can be prevented if you're not building a kind of an AGI about the not self-motivation.

Look, I believe that the way the technology right now is moving, it will not achieve anything that remotely AGI, because it lacks precision and it lacks intelligence as we now define intelligence. But this does not mean that there will not be a breakthrough tomorrow, next month, a year from now, five years from now. If you have to take a guess, I believe it's a decade. A decade, yeah. All right, so we have at least a decade and a couple of hours, right, until the end of the world. I want to say thank you, for me it was amazing.

My pleasure.