Mikhail Burtsev. Artificial Intelligence. How to distinguish a cat from a dog.

Maxim Borisov and Boris Stern discuss Artificial Intelligence with Mikhail Burtsev, a Landau Fellow at the London Institute of Mathematical Sciences 00:47 Stanislav Lem in his novel "Invincible" wrote about "necroevolution" - evolution among non-living devices. You work at the intersection of biology and computer systems. How meaningful are all these parallels between evolution in the case of living and non-living matter, and how naive (or insightful) was Lem in his time? 2:25 As far as we know, computer systems experts are now sharply divided on the question of the possibilities of training models on "secondary" data about the surrounding world (for example, ChatGPT and others deal only with linguistic reality without their own access to the surrounding physical world). Is "real" artificial intelligence possible in the future? Or are all these "Markov chains" and "extended T9" leading to a dead end, and should we try something else? What point of view do you adhere to and why? 7:45 Model of the world. Who builds it, a programmer or a computer? 14:07 All sorts of concerns are expressed about AI, sometimes quite alarmist (computers will take over the world and enslave or destroy humanity), sometimes simply naive, sometimes convoluted (neural networks distort reality, while basing themselves on real human statements and images - after some time, AI products will flood the Internet, and then they will stop relying on anything real and will sharply "become stupid"). It is clear, given that the development of AI is inevitable - if it is banned in some part of the world, then in other places it will continue to develop and provide a clear advantage to those countries that did not impose a ban... 21:38 Artificial intelligence is penetrating many areas of human activity. Will this not lead to the degradation of natural intelligence in these areas? 26:02 What kind of control should AI have - state, supranational - or, as many hope, will the development of neural networks within the framework of open projects by numerous private companies and corporations provide sufficient transparency and allow threats to be stopped in time? 30:32 What are the prospects for Russia in this area, what are its specifics (for example, Sberbank neural networks "don't know" about famous Russian writers and artists just like their English-language counterparts, and they also interpret many Russian words strangely, for example, they draw palm trees instead of hands, which indicates the genesis of all such developments from some purely Western models and subsequent internal translation). Have we reached the world level? What can your laboratory boast about in this regard? 37:15 Where should we expect real breakthroughs in the near future, and which areas are still lagging behind and for what reason? It is obvious that, in general, neural networks generating images have "shot" unexpectedly powerfully over the past year - and they are still developing rapidly. They probably surprise us even more than neural networks working with text. But robots that are tolerably oriented in the surrounding space are still a rarity. Voice communication, speech recognition have been somewhat delayed (although there seem to be no fundamental obstacles), but, apparently, someone is about to seriously surprise us. What else could be so unusual? A "brain - computer" interface? 43:03 Should we expect the emergence of "strong" artificial intelligence? 43:47 Pattern Recognition: How Does a Machine Tell a Cat from a Dog? It is clear that they can do this already, but it seems that here too there are not key breakthroughs, but successes in narrow areas. In any case, systems that recognize images generated by the same neural networks are clearly not able to restore the original request for images, they incorrectly interpret many things and relationships between the depicted characters, the essence of the plots and name incorrectly used styles and those artists to whom they belonged... With retouching, with increasing the resolution of an arbitrary image and other processing, everything is also not very good yet... Or is this only true for publicly available services?