YouSaid · the spoken record

François Chollet

lines on the record: 161
first: 2020-08-31
most recent: 2020-08-31
sittings or episodes: 1
sources: podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

“These reports, I think, basically sum up the meaning of life. Like in the same way that we are the sum of the interactions between many different reports that came from our past, we are ourselves creating reports that will propagate into the future. And that's why we should be, this seems like perhaps an a thing to say, but we should be kind to others during our time on earth because every act of kindness creates reports. And in reverse, every act of violence also creates reports. And you want to carefully choose which kind of reports you want to create. And you want to propagate into the future.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“A beautiful piece of music, a work of art, a grand theory, a new words maybe, that something is going to become a part of the minds of future humans, essentially forever. So everything we do creates reports that propagate into the future. And that's in a way this is our path to immortality is that as we contribute things to culture, culture in turn becomes future humans. And we keep influencing people thousands of years from now. So our actions today create reports.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“The meaning of life? Yeah, that's a big question And I think I can give you my answer, at least one of my answers. So, you know, the one thing that's very important in understanding who we are is that Everything that makes up ourselves that makes up, we are, even your most personal thoughts is not actually your own, right? Like even your most personal thoughts are expressed in words that you did not invent and are built on concepts and images that you did not invent. We are very much cultural beings. We are made of culture. What makes us different from animals, for instance, right? So we are everything about ourselves is an echo of the past, an echo of people who lived before us, right? That's who we are. And in the same way, if we manage to contribute something to the collective edifice of culture, a new idea may be a”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“So, a good model of the world is going to include all kinds of things that are completely useless, actually, just because just in case. Because you need the vessel in the same way that in your portfolio, you need all kinds of stocks that may not have performed well so far, but you need the vesity. And the reason you need the Vestity is because fundamentally you don't know what you're doing. And the same is true of the human mind is that it needs to behave appropriately in the future and it has no idea what the future is going to be like, but it's not going to be like the past. So compressing the past is not appropriate because the past is not predictive of the future.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“That's why as humans, when we are children, in our education, so a lot of it is driven by play, driven by curiosity, we are not efficiently compressing things. We're actually exploring. We are retaining all kinds of things from our environment that seem to be completely useless because they might turn out to be eventually useful. And that's what cognition is really about and what makes it antagonistic to compression is that it is about hedging for future uncertainty.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Your programs, except you have absolutely no idea what sort of context environment and situation they're going to be running in. And you have to deal with that uncertainty, with that future anomaly. An analogy that you can make is with investing, for instance. If I look at the past 20 years of stock market data and I use a compression algorithm to figure out the best trading strategy, it's going to be you buy Apple stock, then maybe the past few years you buy Tesla stock or something. But is that strategy still going to be true for the next 20 years? Well, actually, probably not, which is why if you're a smart investor, you're not just going to be following the strategy that corresponds to compression of the past, you're going to be Following, you're going to have a balanced portfolio. Because you just don't know what's going to happen”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Child with a compression algorithm essentially would be utterly unable, inappropriate to process the next 70 years in the life of that child. So in the models we build of the world, we are not trying to make them actually optimally compressed. We are using compression as a tool to promote simplicity and efficiency in our models, but they are not perfectly compressed because they need to include things that are seemingly useless today, that have seemingly been useless so far, but that may turn out to be useful in the future because you just don't know the future. That's the fundamental principle that cognition, that intelligence arises from is that you need to be able to run appropriate”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Compression is actually kind of a tool in the human cognitive toolkits that is used in many ways, but it's just a tool. It is a tool for cognition. It is not cognition itself. And the big fundamental difference is that cognition is about being able to operate in future situations that include fundamental uncertainty and novelty. So for instance, consider a child at age 10. And so they have 10 years of life experience. They've gotten pain, pleasure, rewards and punishment a period of time. If you were to generate the shortest behavioral program that would have basically run that child over these 10 years in an optimal way, right? The shortest optimal behavioral program given the expense of that child so far. Well, that program that compressed program, this is what you would get if the mine of”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Recognition is compression. But I can't tell you what's the difference. So it's very easy to believe that cognition and compression are the same thing. So Jeff Hawkins, for instance, says that cognition is prediction. And of course, prediction is basically the same thing as compression, right? It's just including the temporal axis. And it's very easy to believe this because compression is something that we do all the time very naturally. We are constantly compressing information. We are constantly trying, we have this bias towards simplicity. We're constantly trying to organize things in our mind and around us to be more regular, right? So it's a beautiful idea. It's very easy to believe there is a big difference between what we do with our brains and compressions.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“It's a very fun test because it's such a simple idea. Like you're given Wikipedia, basically English Wikipedia, and you must compress it. And so it stems from the idea that cognition is compression, that the brain is basically a compression algorithm. This is a very old idea. It's a very, I think, striking and beautiful idea. I used to bit of it. I eventually had to realize that it was very much a flawed idea. So I no longer believe that compression.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“How efficient was. So one thing that's interesting about this notion of scoring you as how many attempts you need is that you can start producing tasks that are way more ambiguous, right? With the different attempts, you can actually probe that ambiguity, right?”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Basically, yes. And I think so, I love the idea of interactivity. I initially wanted an Arctest that had some amount of interactivity where your score on a task would not be one or zero if you can solve it or not, but would be the number of attempts that you can make before you hit the right solution, which means that now you can start applying the scientific method as you solve our tasks, that you can start formulating hypothesis and probing the system to see whether the hypothesis, the observation will match the hypothesis or not. It would be amazing.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“I think you make a great point. The interactivity is a very good setting to force a system to show adaptation, to show generalization. That said, at the same time, it's not something very scalable because you rely on human judges. It's not something reliable because human judges may not feel.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“As a philosophical device, in a philosophical discussion, I think there is something very interesting about it. I don't think it is. In practical terms, I don't think it's conducive to progress. And one of the reasons why is that I think being very human-like, being indistinguishable from a human is actually the very last step in the creation of machine intelligence, that the first ARs that will show strong generalization that will actually implement human-like broad cognitive abilities they will not actually behave overlook anything like humans. Human likeness is the very last step in that process. And so a good test is a test that points you towards the first step on the ladder, not towards the top of the ladder.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“So I like the Exercise Prize better because it's more pragmatic, it's more practical. It's actually incentivizing developers to create something that's useful. As a human mission interface. So that's slightly better than just the imitation.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“We are biased. Like, we have theory of mind. We are constantly projecting emotions, intentions. Agentness. Agentness is one of our core init priors, right? We are projecting these things on everything around us. Like if you paint a smiley on a rock, the rock becomes happy in our eyes. And because we have this extreme bias that permeates everything we see around us, it's actually pretty easy to trick people. It is a different.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“That's the whole in practice, what it turns out, it is very easy to deceive people in the same way that you can do magic in Vegas. You can actually very easily convince people that they're talking to humans when they're actually talking to algorithms.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Achieve with David Copperfield could achieve it in his show at Vegas, right? And what he's doing is very elaborate, but it's not actually. It's not physics, it's not making any progress in our understanding of the universe.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Mind is an information processing system, and that you could probably encode it into a computer. So another reason why I'm not a fan of this type of test is that the incentives that it creates are incentives that are not conducive to proper scientific research. If your goal is to trick, to convince a panel of human judges that they're talking to human, then you have an incentive to rely on tricks and prestige in the same way that let's say you're doing physics and you want to solve teleportation. And what if the test that you set out to pass is you need to convince a panel of judges that teleportation took place and they're just sitting there and watching what you're doing. And that is something that you can.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“And by the way, we should keep in mind that when Turing proposed the imitation game, it was not meaningful for the imitation game to be an actual goal for the field of AI, an actual test of intelligence. It was using the imitation game as a thought experiment in a philosophical discussion in his 1950 paper. He was trying to argue that theoretically it should be possible for something very much like the human mind indistinguishable from the human mind to be encoded in a Turing machine. And at the time, that was a very daring idea. It was threshing credulity. But nowadays, I think it's fairly well accepted.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“I'm not a fan of the Turing test itself or any of its variants for two reasons. So first of all, it's It's really copying out of trying to define and measure intelligence because it's entirely outsourcing that to a panel of human judges. And these human judges, they may not themselves have any proper methodology. They may not themselves have any proper definition of intelligence. They may not be reliable. So the truth is already failing one of the core psychometrics principles, which is reliability because you have biased human judges. It's also violating the standardization requirement and the freedom from bias requirement. And so it's really a cope out because you are outsourcing everything that matters, which is precisely describing intelligence and finding a standalone test to measure it. You're outsourcing everything to people.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Computers are part of it non human systems probably not contributing much, but AIs are definitely contributing to that. Like Google Search, for instance, is a big part of it.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“And you can scale that externalized cognition far beyond the capability of the human brain. And you could see civilization itself is it has capabilities that are far beyond any individual brain. And we'll keep scaling it because it's not rebound by individual brains. It's a different kind of system.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Not just computers, not just phones and the internet. I mean, all of culture, like language, for instance, is a form of external recognition. Books are obviously externalized recognition.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“I think augmenting human intelligence is a very valid and very powerful avenue, right? And that's what computers are about. In fact, that's what all of culture and civilization is about. Culture is externalized cognition and we rely on culture to think constantly.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Don't believe so. I think the brain is very slow. It speaks the fastest things that happen in the brain are at the level of 50 milliseconds, forming a conscious thought can potentially take entire seconds. And you can already read pretty fast. So I think the speed at which you can take information in and even the speed at which you can add with information can only be very incrementally improved. If you're a very, very fast typer, if you're a very trained typer, the speed at which you can express your thoughts is already the speed at which you can form your thoughts.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“I think it's already close, fairly close to optimal, which is why speed reading, for instance, does not work. The faster you read, the less you understand.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Already has direct access to Wikipedia, it's on your phone, and you have your hands and your eyes and your ears and so on to access that information. And the speed at which you can access it.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes. I think that's a good one, but that's also not about externalizing our intelligence via information pricing systems, the external information processing systems, which is very different from brain-computer interfaces.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“I am fairly skeptical of neural interfaces because they are trying to fix one specific bottleneck in human machine cognition, which is the bandwidths bottleneck, input and output of information in the brain. And my perception of the problem is that bandwidth is not at this time a bottleneck at all, meaning that we already have senses that enable us to take in far more information than what we can actually process.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“General. I think it's quite likely that there is a hard limit to high intelligence and the system can be. But at the same time, I don't think humans are anywhere near that limit.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes. So creating a system that is so general that it could essentially achieve human skill parity over arbitrary tasks and arbitrary domains with the same level of improvisation and adaptation power as humans when it encounters new situations. And it would do so over basically the same range of possible domains and tasks as humans and using essentially the same amount of training experience, of practice as humans would require. That would be human level extreme generalization. I don't actually think humans are anywhere near the optimal intelligence bound if there is such a thing.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“And finally, we'd have extreme generalization, which is basically flexibility, but instead of just considering one specific domain, like driving or domestic robotics, you're considering an open-ended range of possible domains. So a robot would be capable of extreme generalization if let's say it's design and trained for cooking, for instance. And if I buy the robot and if I'm able, if it's able to teach itself gardening in a couple weeks, it would be capable of extreme journalization, for instance.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Distribution that you know through your training data. And a higher degree would be flexibility in machine intelligence. So flexibility would be something like an L5 cell driving car or maybe a robot that can pass the cafe cup test, which is the notion that you'd be given a random kitchen somewhere in the country and you would have to go make a cup of coffee in that kitchen. So flexibility would be the ability to deal with unknown unknowns, so things that could not dimension survivability that could not have been possibly foreseen by the creators of the system within one specific task. So generalizing to the long tail of situations in self-driving, for instance, would be flexibility. So you have robustness, flexibility.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Generalization, degrees of generalization as a spectrum. And the lowest level is what machine learning is trying to do, is the assumption that any new situation is going to be sampled from a static distribution of possible situations and that you already have a representative sample of that distribution. That's your training data. And so in machine learning, you generalize to a new sample from a known distribution. And the ways in which your new sample will be new or different are ways that are already understood by the developers of the system. So you are generalizing to known unknowns for one specific task. That's what you would call robustness. You are robust to things like noise, small variations, and so on. For one, a fixed known”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Right, so generalization is a very old idea. I mean, it's even older than machine learning. In the context of machine learning, you say a system generalizes if it can make sense of an input it has not yet seen. And that's what I would call system-centric generalization. Generalization with respect to novelty for the specific system you're considering. So I think a good test of intelligence should actually deal with developer aware generalization, which is slightly stronger than system-centric generalization. The ability to generalize to novelty or uncertainty that not only the system itself has not access to, but the developer of the system could not have access to either”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes, that said, you know, with the state of the art, it's like at 20%, we're still very, very far from human level, which is closer to 100%. And I do believe that it will take a while until we reach human parity on arc. And that by the time we have human parity, we will have AI systems that are probably pretty close to human level in terms of general fluid intelligence, which is, I mean, they're not going to be necessarily human-like. They're not necessarily, you would not necessarily recognize them as being an AGI, but they would be capable of a degree of generalization that matches the generalization performed by human fluid intelligence.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, exactly. That was the primary reason to Kaggle competition is to check. If some clever person was going to hack the benchmark, that did not happen. Like people who are solving the tasks, essentially doing it, well, in a way, they're actually exploiting some flaws of ARC that we will need to address in the future, especially they're essentially anticipating what sort of tasks may be contained in the test set, right?”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Promise, but it's also not an impossible change. It's not accessible. You can start making progress basically right away. At the same time, we are still very far from having solved it. And that's actually a very positive outcome of the competition is that the competition has proven that there was no obvious shortcut to solve these tasks.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“So, one thing I think we did well with Arc is that it's proving to be a very actionable test in the sense that machine performance on arc started at very much zero initially while humans found actually the tasks very easy. That alone was a big red flashing light saying that something is going on and that we are missing something. And at the same time, machine performance did not stay at zero for very long, actually, within two weeks of the cargo competition. We started having a non-zero number. And now the state of the art is around 20% of the test set solved. And so ARC is actually a challenge where our capabilities start at zero, which indicates the need for”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes, absolutely. So arc itself will keep evolving. So I've talked about crowdsourcing. I think that's a good avenue. Another thing I'm starting is I'll be collaborating with folks from the psychology department at NYU. To do human testing on ARC. And I think there are lots of interesting questions you can start asking, especially as you start correlating machine solutions to arc tasks and the human characteristics of solutions. Like, for instance, you can try to see if there's a relationship between the human perceived difficulty of a task and the”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Absolutely. As you solve our tasks as a human, you will be forced to basically introspect. How you come up with solutions. And that forces you to reflect on the human problem solving process and the way your own mind generates abstract representations of the problems it's exposed to. I think it's due to the fact that the set of core knowledge priors that ARC is built upon is so small. It's all a recombination of a very, very small set of assumptions.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, one thing that kind of surprised me that I wasn't expecting is that Lots of people seem to actually enjoy ARC as a kind of game. And I was really seeing it as a test, as a benchmark of fluid general intelligence. And lots of people, just including kids, just enjoying it as a game. So I think that's encouraging.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes, absolutely. It is imperative that the test set that you're using to actually benchmark algorithms is not accessible to the people developing these algorithms because otherwise what's going to happen is that the human engineers are just going to solve the tas But that again, what you're seeing here is the process of intelligence happening in the mind of the human and then you're just capturing its crystallized output. But that crystallized output is not the same thing as the process it generated. It's not intelligent in itself.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“I'm not saying they've done anywhere near a perfect job at it. There is some amount of redundancy and there are many imperfections in Arc. So that said, you should consider ARC as a work in progress. It is not the definitive state, the arc tasks today are not definitive states of the test. I want to keep refining it in the future. I also think it should be possible to open up the creation of tasks to a broad audience, to do crowdsourcing. That would involve several levels of filtering, obviously, but I think it's possible to apply crowdsourcing to develop a much bigger and much more diverse arg data set that would also be free of potentially some of my own personal biases.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, you need a source of novelty, of unfakable novelty. And one thing I found is that as a human, you are not a very good source of unthinkable novelty. And so you have to pace the creation of these tasks quite a bit. There are only so many unique tasks that you can do in a given day.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Including the test and the private test set. I think it's fairly difficult in the sense that a big requirement is that every task should be novel and unique and unpredictable. You don't want to create Your own little world that is simple enough that it would be possible for a human to reverse and generate and write down an algorithm that could generate every possible arc task and their solution. So instance, that would completely invalidate the test.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“It's very similar to classic IQ tests. Like it's not very original in that sense. The main difference with IQTests is that we make the priors explicit, which is not usually the case in IQTest. So we make it explicit that everything should only be built onto a core knowledge priors. I also think it's generally more diverse than IQ tests in general. And it perhaps requires a bit more manual work to produce solutions because you have to click around on a grid for a while. Sometimes the grids can be as large as 30 by 30 cells.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes, so clarifying my own ideas about abstraction by forcing myself to produce tasks that would require the ability to produce that form of abstraction in order to solve them.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source
“When you're faced with a problem as murky as understanding how to autonomously generate abstraction in a machine, you have to co-evolve the solution and the problem. And so part of the reason why DesignARC was to clarify my ideas about the nature of abstraction. And some of the tasks are actually designed to probe bits of that theory. And there are things that turn out to be very easy for humans to perform, including young kids, right? But turn out to be near impossible for machines.”
2020-08-31 · Lex Fridman Podcast · #120 – François Chollet: Measures of Intelligence · IDENTIFIED FROM THE TRANSCRIPT · source