Майкл Левин и Дональд Хоффман - направления и перспективы AI в "Новой" Парадигме
В день победы над здравым смыслом вышел очередной выпуск научно-популярно-просветительского подкаста инженера/бизнес консультанта Перри Маршалла (Perry Marshall) Evolution 2.0 https://evo2.org/the-podcast/
Встреча интересна «лобовым» столкновением двух нестандартно мыслящих ученых (причем, ученых в разных областях - один в биологии, другой в теории информации и визуальной интеллигенции) Дональда Хоффмана и Майкла Левина. Они очень часто в «научных» кругах вызывают на себя клеймо «wierdos», чудаков от псевдонаук. Но, тем не менее...
May 9, 2022
The Engineering of Consciousness with Michael Levin and Donald Hoffman
Two leading pioneers in the field of cognition discuss the sea change that is underway in consciousness and evolution: Michael Levin is 10 years ahead of multiple fields in biology, producing extraordinary breakthroughs in limb regeneration, cancer, and bioengineering. Donald Hoffman is a champion of a new model that says...
Ранее про Майкла Левина (Michael Levin):
Вопрос детям: Откуда ты знаешь, что не был создан секунду назад с пакетом внедренных воспоминаний?
Ранее про Дональда Хоффмана (Donald Hoffman)
Иллюзии субъективного восприятия пространства...
Дональд Хоффман vs Илон Маск и Ник Бостром: Живем...
Дональд Хоффман о "Чуде" в основе любой научной...
+ скачать книги Donald D. Hoffman / Дональд Хоффман - Видим ли мы реальность такой, какая она есть? - Собрание Книг [1989-2019, PDF/DjVu/EPUB, ENG]
В течении 75 минутной встречи участники совсем не дебатировали друг с другом, а в основном рассказывали о себе, своих исследованиях и описывали свои взгляды в контексте вопросов, задаваемых ведущим.
На 56 минуте Перри Маршал задал финальный вопрос - предложил участникам пофантазировать, если было бы возможно сдать в утиль старые базовые «редукционистские» теории и начать исследования с новой точки consciousness-based зрения, то куда сейчас нужно копать?
Let's say you could push a big reset button and instead of scientific revolution (taking 50 years) we're going to start a new University in a brand new country on an island somewhere. And instead of clinging to reductionism we can embrace an agent-based or a consciousness-based view of science and we're unfettered to pursue that to the n-th degree. How does that science look different 5 -10-15 years from now as it develops? What kinds of things is it doing that are just not possible in the reductionist view?
Дональд Хоффман:
- Трудно, но необходимо избавиться от иллюзии, что время и пространство базовые условия для зарождения всего, и что путем углубления в меньшие измерения мы можем найти базовые «кирпичики», из которых все создано.
letting go of reductionism is a big move we've all been raised in it so this is a new trick for for most of us and hard to think out of the box but the direction that i'm myself pursuing and i would want to pursue is the physicists have already found structures beyond time they're already there they're finding these polytopes. They don't know what the polytopes are about. All they know is that when you compute the volumes of these polytopes you get the right answer and you see symmetries in the data and you can see how space time gets its features of unitarity and locality from these structures that don't really care about unitarian locality. They're beyond space time.The question then is, I think, this is at the forefront of the current science what in the world are these polytopes about right they're the next step they're clearly the next step beyond space time what are they about
- нужно переосмыслить идею зарождения сознания (consciousness) из мозга.
We have to completely reboot how think about consciousness. Right now 99 of my colleagues are in the reductionist mode - how do microtubule quantum state reductions that are orchestrated in the right way cause specific conscious experiences? How do integrated information structures in physical systems create consciousness? How does a global workspace with the right broadcasting capabilities working memory, how does that create conscious experiences? - It's all in a reductionist framework. If we start with small physical systems and then neurobiological systems how can we boot up consciousness from that? Space time is doomed. Those entities don't even exist when they're not perceived. So to be very very clear - I have no neurons right now. If you looked you would see neurons because that's a data structure you create in your interface when you need it. And as soon as you don't need it you garbage collected. So neurons exist when you perceive them and otherwise they don't exist. So the whole neurobiological reductionist approach and the physical reductionist approach really that's the big step first we have to let go of that. And then whatever we propose is going to have to show where these polytopes come from.
- нужно попробовать рассматривать сознание (consciousness) как базу, из динамики которой и выходит восприятие реальности
So what I'm working on are is a dynamics of consciousness. It's not in space and time. It's a dynamics in an abstract consciousness space. And it looks like the asymptotic behavior. So the long-term behavior of this can be described by things called birkhoff polytopes. And the birkhoff polytopes turn out to be cells in the gr the positive grasmanian that gives rise to the amplitude. So the direction I'm looking at right now (and this is, you know, a moon shot, but you got to do it) - you start with a theory in which consciousness is fundamental. It's a dynamics outside of space and time and the dynamics itself can be stationary (in the technical sense that the entropy does not increase). So in some sense it's a timeless dynamics of consciousness. In the sense there's no entropic time. Entropy is not increasing but you can prove a couple things. It's trivial to prove that any projection of this timeless dynamics by conditional probability will lead to an artifact of entropic time.
- время-пространство артефакты проекции динамической системы, которой самой не нужно ни время ни пространство. Энтропия здесь исключена, ресурсы неисчерпаемы.
In other words the reason why space time isn't fundamental is that time itself is an artifact of a projection of a dynamical system. Which itself need not be time. It need not have any time in the entropic sense of time.You can have a sequence but not entropic time.
- внутри такой структуры идея соревнования, конкуренции за ресурсы исключена.
From that we can then say suppose we have a dynamical system of conscious agents in which there's not competition. And there's not limited resources. It could still look like limited resources like time is an artifact that's a limited resource.
- вся идея Эволюции может быть всего лишь проекцией из этой динамики, в которой нет исчерпаемости ресурсов
Evolution of a natural selection itself, that whole framework, may be an artifact of a projection of a system in which there is no competition. In which there is no limited resource. This would be a complete revolution in our understanding of how the whole universe works.
- но мы, человеческие существа, не способны в настоящее время мыслить вне «временных» рамок, а нужно как-то научиться это делать. Научиться мыслить в формате «неисчерпаемых» ресурсов вне времени-пространства.
Because right now we are not yet a species that knows how to think outside of time and that's what we have to do. It's a new trick but we can do it. We can learn to think about structures outside of time and perhaps outside of limited resources. Once we've done that once we've got so the idea would then be have specific models of conscious agent dynamics that lead to specific connections with like the amplitude and the cosmological polytopes. So that we can then give a dynamical systems understanding of these static polytopes that the physicists have found the polytropes are just sitting there they encode all the information but what are they about you know why is there's this platonic world of politics just sitting there isn't there some like dynamism behind that that we should be understanding even if it's not in tropic time there's some kind of dynamism and that's what I'm working on is a dynamism of conscious agents. Whose goal would be to show that it's asymptotic behavior precisely gives rise to these polytopes. And these polytopes then the physicists show how to take them into space time. So we then see how our interface is built we have perception. And consciousness now being fundamental but scientifically rigorous. and then we show all the way through the asymptotics of that projects into the you know natural trihedron for example which projects into space time we see precisely how our interface is created
- как только мы сможем математически и через физические эксперименты доказать правомерность возникновения пространства-времени через динамику сознающих агентов, нам откроются новые возможности «интерфейса» и воздействия на него. Далекие космические путешествия станут реальностью.
And then we can play with the interface. Once we know what's beyond space and time with mathematical precision we can expect new technologies that would not be mind-blowing and a new understanding of life.
With the technologies right now most of the galaxies that we see we could never get to because they're moving away too fast. So if we try to go through space to them there's all this beautiful release real estate that's waving at us and we can never get to it. But if we don't have to go through space-time if we realize that space-time is just a data structure maybe we can go outside of it. Go around it. We just don't have to go through it. That would be truly stunning.
- другой важный аспект нового взгляда должен заключаться в понимании условности разделения на живая/неживая материя. Конечно, про камень или протон нельзя сказать что они «живые», но это значит лишь что у нас просто нет достаточной информации для заключения об этих их свойствах, потому что у нас свой «человеческий» интерфейс, в котором они представлены как «неживые» символы.
But the other aspect of this is that life itself the distinction we make between living and non-living. We think of it as a principal distinction. (...) We have this idea about what's living and what's not. From the point of view that space-time is not fundamental, is just an interface, the distinction we make between living and non-living is an artifact of the limitations of our interface. It's not an insight into the nature of reality. What we have to do is understand that there's going to be this deeper level. Perhaps - a theory of consciousness, but what that means is not that a rock is conscious or that you know a proton is conscious. It just means that we're dealing with the dynamical system of consciousness (...) but that doesn't mean that the proton exists - it's a symbol. It's an interface symbol and second my statement that is not alive. That just means my symbol is not very informative. All of our assumptions right now in science in cognitive neuroscience trying to understand almost every one of them has to be ripped apart and re-thought from the ground up. Once we let go of space-time being fundamental then we need to use a younger generation that's ready to think out of the box because that's where the the resource the ideas are going to come.
Майкл Левин:
- сначала нужно принять к серьезному рассмотрению идею про «разум во всем» (mind everywhere). Нужно пересмотреть идею, что только одна химия может объяснить всю жизнь.
first thing I would like to do is to (I said this in that paper) the reason it's called a technological approach to mind everywhere. Is that I would like an engineering commitment to these things. Let's not have preconceptions to how things should be that chemistry is the best level of explanation or that there's no way that my thermostat has goals and memories
- не нужно держаться за «свободную волю», бежать в страхе от идей Контроля и Предсказания, нужно применить «инженерный» подход.
and then work out away from there. Let's just go with whatever model is going to give us the best prediction and control - that's what engineers do. (...) I've had people say, - You know, you're going to use prediction and control as a criterion of truth. That's terrible! - Well, I don't know what else you have. I don't know. I'm sure it's a philosophically naive view, fine, but I don't know what else we have.
- нужно изучать биологическую систему, проводить эксперименты по открытию ее возможностей, а не навязывать свои концепции.
To me all of this has to be cast in an observer relative and I think Don was mentioning some of this - "an observer relative" and kind of perspective where not everybody's view of things is not equally valid but we can compare them and we compare them with empirical experiment. We do experiments on all this. We don't have feelings about this.
So if you make a claim that some particular system has zero agency or very low agency or it can't do this or it can't do that let's find out can you train it? Can you communicate with it? What is it capable of?
You will be surprised as we and others have found out that you can look at a gene regulatory network and say, - Ah! This is a paradigmatic example of determinism. It's just a bunch of arrows turning genes turning each other on and off. Well, guess what, they can do six different kinds of learning. You don't know these things until you try. You can't just sort of make decisions about this "from an armchair".
So that's the first thing I would say, - Let's commit to doing experiments and having engineering be the kind of role model for this.
- как можно четко определить в развитии момент появления интеллекта? Не существует в природе такого «судебного» понимания, как например, достижения 18-летнего возраста.
Let's take evolution and developmental biology seriously. What they're telling us is that almost everything of interest is a graded continuum that you can vary smoothly and if you think there's some sort of phase transition that happens - let's zoom in and let's make some chimeras - and you will find out that in your sort of what you thought was a rapid transition, there are in fact lots of in-between cases that are going to give you a headache as far as trying to decide is it cognitive or isn't it. There's no way anybody's going to draw any convincing bright lines here right so all of this is it's kind of like I use the example of - it's kind of like the category of being an adult - it's useful in court and everything but we all know there's nothing that magically happens to you when you turn 18...
- нужно избавиться от навешивания «ярлыка» на все - интеллигентно ли оно, обладает разумом или нет.
people have these categories - Is it a machine or is it alive? Is it a robot? It's not cognitive! I'm cognitive that's just phys none of. These categories are useful. They're not going to survive the next few decades as we make hybrids and cyborgs and the bio-engineered things. That all kinds of pseudo-problems arise by pretending that there are sharp distinctions between these things
- необходимо пересмотреть концепции памяти (генетической и обычной), где она может находиться и храниться.
we need to radically revise about of where is it encoded. So people ask me all the time when we talk about the goals for these states. They say, -Well, where is it encoded? Because they're used to looking at things like "protein sequences". Where, if you ask me, where is it encoded? I know exactly where it's encoded it's encoded in the DNA! That's how protein sequences work!... But most things don't work that way.
- условно говоря, в «платоновом» пространстве где-то находятся неизменяемые возможности, законы, которые проецируются в наш мир через машины, которые и мы сами также можем построить.
somewhere in "platonic" space there are some sort of affordances from the laws of computation the laws of mathematics. The kinds of things that don't change when the laws of the Universe change. These truths of number theory and all this stuff. They are there for you when you make the right machine. When you make a configuration of matter - that somehow resonates with these things.
- в самих генах не заложена конкретная информация о формах, размерах и свойствах.
We need to get away from this simplistic idea that things have to be encoded somewhere. People understand emergence and complexity and all that but it's way more than that. And it's this idea that you don't need to micromanage all this stuff because a lot of it you get for free.
- Эволюция запускает «машины» в общем смысле, а уже их «программное» обеспечение делает всю тяжелую работу по адаптации. Наша задача понять, каким образом пробуждаются те или иные возможности «программы»
What evolution does is tweak the machines. And a lot of the hard work is not done at the level of that hardware at all. It's just pulling down these capacities. The work is going to look like trying to understand. Map out with the shape of that space - what space was it that had a xenebot attractor in it? What else is in that space? What else do these cells know how to do?
- возможность применения новых открытий в регенеративной медицине не с точки зрения молекулярной медицины, генной инженерии, а с точки зрения стимулирования ответной реакции по регенерации. Нужно объяснить клеткам что нужно сделать, а не насильно внедрять в них готовые решения через генетическое конструирование.
Regenerative medicine - where we don't try to micromanage the process and try to operate the way that we do. All molecular medicine today it's all about rewiring the hardware. It's all about changing the genes the protein pathways, protein engineering - all that stuff. That's fine. But let's see if we can go beyond that and take advantage of the intelligence of the system we should be operating with. Proper experimentation as far as what is the IQ level of all the components. What do cells know how to do? What does tissue and organs know how to do? And can we do "behavior shaping" so regenerative medicine to actually rewrite? And we've done this in certain contexts - to rewrite the gold state and let the system do what it knows how to do right engineers.
- биологи должны начать думать как инженеры, Нужно думать о цели, не от малого к большому, а наоборот, и это откроет совершенно новые возможности.
Biologists don't often believe in goals but engineers sure do. Otherwise you couldn't use thermostats. You can only write - "you have a thermostat". Because full well you can depend on this thing. Even though some stuff is going to change it will get the job done. So again, you gain massive amounts of power that way that are just impossible from the bottom up.
- это наш билет в создание искусственного интеллекта.
I think what we're going to gain from this is general artificial intelligence. Because by paying attention to how it is that competent tiny little competent parts scale up into large selves so the scaling of cognition the quad the competition. And I could tell stories about how that might work. Competition and cooperation between these things as they scale to larger goals and more profound stress states and things like that. That is how you build real intelligence that is self developed from the ground up from pieces that don't know what they are until they sort of form themselves.
- возможно «пробудить» сознание в любом правильном создании, неважно, создано оно из металла или полностью биологическое. Нужен правильный софт, и способы его внедрения. Нужно запустить процесс само-создания, используя базовые принципы, ограниченные ресурсы, конкуренцию.
That's when people say, - Ah! Machines can never BE! They're metal! It's not the substrate, it's not the fact that what they're made of. It's not the fact that they came off of a factory instead of evolution. It's not any of that. It's that I think it's the process of self-creation using basic principles of active inference and limited resources and various other things that are very generic.
- и тогда естественным образом возникнет самосознающая система с внутренней мотивацией и прочими нам хорошо известными биологическими вещами. Нужно отказаться от идеи «создателя» с внедренными программами, что и как нужно делать.
And then you have a system that has true preferences like intrinsic motivation and all of this stuff. And it really then you have a real agent as opposed to the kinds of machines and AI we have now which is set by somebody else, by the user. The creator says - "This is what you are! Here's what you can do! Here's the program that you're following!" You're good to go and so on, so I think we can get way beyond that.