Одна из участниц нашего курса обучения поделилась этим видео Ilya Sutskever – We're moving from the age of scaling to the age of research с Ильей Суцкевером, сооснователем компании OpenAI, чьим научным руководителем был профессор Джеффри Хинтон - один из основоположников глубокого обучения.

Вопрос был такой: "имеет смысл слушать интервью такого типа [в обучении трендвотчингу], или там заведомо неглубоко?". Вопрос был в том смысле, что можно ли как-то сузить зону поиска и копания, так как тема исследования - особенно, по началу, кажется безграничной.

Ответ: "Я смотрю практически ВСЕ. При этом, фильтрую 90% 🙂 То есть я смотрю на все через призму интересующего меня вопроса, — он как точка сборки, в которую не попадает ненужный мне шум".

Ну и Суцкевера грех не посмотреть. Первых полчаса хватило, чтобы понять главное о развитии искусственного интеллекта сегодня:

Судите сами, - привожу основные, на мой взгляд, фрагменты интервью:

Let's take the case of competitive programming,  since you mentioned that. Suppose you have two   students. One of them decided they want  to be the best competitive programmer, so they will practice 10,000 hours for that domain. They will solve all the problems, memorize all the proof techniques, and be very skilled at quickly  and correctly implementing all the algorithms. 

By doing so, they became one of the best. Student number two thought, «Oh, competitive programming is cool." Maybe they practiced for 100 hours,   much less, and they also did really well. Which one do you think is going to do better in their career later on? The second.  Right. I think that's basically what's going on. The models are much more like the first student, but even more. Because then we say, the model should be good at competitive programming so let's get  every single competitive programming problem ever. 

And then let's do some data augmentation  so we have even more competitive   programming problems, and we train on that. Now you've got this great competitive programmer. 

With this analogy, I think it's more intuitive. Yeah, okay, if it's so well trained, all the different algorithms and all the different  proof techniques are right at its fingertips. 

And it's more intuitive that with this  level of preparation, it would not   necessarily generalize to other things. But then what is the analogy for what the second student is doing before  they do the 100 hours of fine-tuning? 

Somehow a human being, after even 15 years with a tiny fraction of the pre-training  data, they know much less. But whatever they do know, they know much more deeply somehow. Already at that age, you would not make mistakes that our AIs make.

How do you think about emotions? What is the ML analogy for emotions? 

It should be some kind of a value function thing. But I don’t think there is a great ML analogy because right now, value functions don't play a very prominent role in the things people do. 

What I was alluding to with the person  whose emotional center got damaged, it’s more  that maybe what it suggests is that the value function of humans is modulated by emotions in  some important way that's hardcoded by evolution. And maybe that is important for  people to be effective in the world. 

The thing which I think is the most fundamental is that these models somehow just generalize dramatically worse than people. It's super obvious. That seems like a very fundamental thing. So this is the crux: generalization.

There are two sub-questions. There's one which is about sample  efficiency: why should it take so much more data for these models to learn than humans? There's  a second question. Even separate from the amount of data it takes, why is it so hard to teach  the thing we want to a model than to a human? 

For a human, we don't necessarily need a verifiable reward to be able to… You're probably mentoring a bunch of researchers right now, and  you're talking with them, you're showing them your   code, and you're showing them how you think. From that, they're picking up your way of  thinking and how they should do research. You don’t have to set a verifiable reward for   them that's like, "Okay, this is the next part of  the curriculum, and now this is the next part of your curriculum. Oh, this training was unstable."  There's not this schleppy, bespoke process. 

You could actually wonder that one possible explanation for the human sample efficiency  that needs to be considered is evolution. Evolution has given us a small amount  of the most useful information possible. For things like vision, hearing, and  locomotion, I think there's a pretty strong case that evolution has given us a lot. For example, human dexterity far exceeds… I mean robots can become dexterous too if you subject  them to a huge amount of training in simulation. 

But to train a robot in the real world to quickly pick up a new skill like a person does seems very out of reach. Here you could say, "Oh yeah, locomotion. All our ancestors needed  great locomotion, squirrels. 

So with locomotion, maybe we've  got some unbelievable prior."  You could make the same case for vision. I believe Yann LeCun made the point that children learn to drive after 10  hours of practice, which is true. 

You don't get to see that much data as a five-year-old.  You spend most of your time in your parents'  house, so you have very low data diversity. 

But you could say maybe that's evolution too. But in language and math and coding, probably not. 

It still seems better than models. Obviously, models are better than the average human at language, math, and coding. But are they better than the average human at learning?

Oh yeah. Oh yeah, absolutely. What I meant  to say is that language, math, and coding—and  especially math and coding—suggests that whatever it is that makes people good at learning is  probably not so much a complicated prior, but something more, some fundamental thing. I'm not sure I understood. Why   should that be the case?

So consider a skill in which people exhibit some kind of great reliability. If the skill is one that was very useful to our ancestors for many millions of years, hundreds  of millions of years, you could argue that maybe

humans are good at it because of evolution,  because we have a prior, an evolutionary prior that's encoded in some very non-obvious way that somehow makes us so good at it. 

But if people exhibit great ability, reliability,  robustness, and ability to learn in a domain that really did not exist until recently, then  this is more an indication that people might have just better machine learning, period. How should we think about what that is? What is the ML analogy? There are a couple of interesting  things about it. 

It's more unsupervised. A child learning to drive a  car… Children are not learning to drive a car. A teenager learning how to drive a car is not  exactly getting some prebuilt, verifiable reward.

It comes from their interaction with  the machine and with the environment.  

It takes much fewer samples. It seems  more unsupervised. It seems more robust?  Much more robust. The robustness  of people is really staggering. 

Do you have a unified way of thinking about  why all these things are happening at once?  What is the ML analogy that could  realize something like this? 

One of the things that you've been asking about is how can the teenage driver self-correct and learn from their experience without an external teacher? The answer is that they have their value function. 

Whatever the human value function is,  with a few exceptions around addiction, it's actually very, very robust. So for something like a teenager that's learning to drive, they start to drive, and  they already have a sense of how they're driving immediately, how badly they are, how unconfident. And then they see, "Okay." And then, of course, the learning speed of any teenager is so fast. After 10 hours, you're good to go.  It seems like humans have some solution, but I'm curious about how they are doing it and why is it so hard? How do we need to reconceptualize the way  we're training models to make something like this possible? 

That is a great question to ask, and it's  a question I have a lot of opinions about. But unfortunately, we live in a world where  not all machine learning ideas are discussed freely, and this is one of them.

There may be another blocker though, which is that

But regardless, I do think it points to  the existence of some machine learning principle that I have opinions on. But unfortunately, circumstances 

make it hard to discuss in detail…".