The leading specialist of META* in artificial intelligence, Yann Lekun , dispelled popular misconceptions about the possibilities of modern AI-systems. During the performance at the Hudson forum, he emphasized that the current models are not able to truly think, plan and reason like the human brain.
Lekun entered into a polemic with optimistic colleagues, including the founder of XAI Ilon Musk and co -founder of Google Deepmind, Shein Legg. These experts argue that AI at the human level will appear in the near future. However, according to Lekun, this can happen only after several decades.
Why even the most advanced LLM are far from human intelligence? Modern systems work as one -dimensional predictors who guess the next token – letters or short words. Models for working with images and videos act as two -dimensional predictors, predicting the next pixel. At the same time, none of the systems perceives the three -dimensional world.
The tasks for a person are at an impasse even the most advanced algorithms. The lecturer gave an indicative example: a ten -year -old child learns to remove from the table in a few hours, and by the age of 17 he masters the driving of the car. Artificial intelligence, despite thousands of hours of training, is still very bad in the physical world.
The solution to the problem can be a new type of AI architecture – “World Model”. She will create a mental model of behavior of the world and will be able to predict the consequences of various actions.
The principle of the “world model” can be considered on the example of cleaning in the bedroom. The human brain, seeing the mess, immediately represents the sequence of actions to guide purity. We do not need to sort out different methods or learn to clean the room – we create an action plan on the first attempt.
The concept attracts more and more attention in the industry. Recently, a group of famous AI researchers, headed by Fei Fei Lee, presented a project for 230 million dollars for its startup World Labs. Openai also calls its so far unsettled Sora video generator “world model.”
Lekun described the concept of using “world models” to create AI of the human level in article 2022. The system receives a basic idea of the world through video and memory, then predicts changes based on these data. The model also takes into account the desired result and restrictions, so as not to harm people when reaching the goal.
The research laboratory
is actively working on the creation of advanced technology.
* META and its products are recognized as extremist, their activities are prohibited in the territory of the Russian Federation.