Full podcast episode: ua-cam.com/video/5t1vTLU7s40/v-deo.html Lex Fridman podcast channel: ua-cam.com/users/lexfridman Guest bio: Yann LeCun is the Chief AI Scientist at Meta, professor at NYU, Turing Award winner, and one of the most influential researchers in the history of AI.
00:00:26 Train a world model by observation without relying on gigantic data sets. 00:00:42 Explore innovative ideas that do not necessarily require scaling up. 00:00:49 Implement planning with a learned world model for non-physical systems like the internet or databases. 00:01:13 Develop a system to plan a sequence of actions for problem-solving in various scenarios. 00:01:54 Investigate hierarchical planning to handle complex tasks efficiently. 00:03:00 Learn how to represent hierarchical action plans for robots or intelligent systems. 00:03:13 Train systems to understand hierarchical representations of action plans using deep learning.
Anderson's ACT-R goal stacking maybe. Add time to complete, action wait times/down time, and action ranking. Really want to mimic humans, add a need stack (charging, cleaning sensors, etc.)
No, not just robotics. He’s saying that planning can be used for both non-physical (databases, task scheduling/chip design, etc.) and physical environments, which isn’t limited to just robotics. In fact robots require A LOT of physical data, especially with RL systems. Operating in virtual simulation with planning can help robots navigate in the physical world easier with synthetic data. TLDR he’s just saying that we should use hierarchical planning for all intelligent systems that require a series of complex actions
Usually AI is good at doing one specific task. I think he is talking about programming AI systems that do multiple tasks in sequence, and understanding which is the next best task. In the example of a robot going from a city to another it would be: 1. Get out of the house. 2. Identify the best method of transportation. 3. Request an uber to the airport. Etc
He's saying current AI systems are only capable of doing multi-step (hierarchical) tasks when all the steps (plans) are defined and easily trainable, but how do you get a model to learn a multi-step task without defining the all plans
3:00 - millions of years of evolution. If we come to understand the brain better as a system and be able to directly database adaptations that include such hierarchical understanding and navigation through the world which all animals possess to some degree, it would improve AI as a discipline. Perhaps even render it novel folly in comparison to the super AI 'hardware' we already possess. Nature so often already has solutions to our modern civilized questions.
Just like how babies learn to recognise physical objects when her parent points to the moon and says, "This is the moon". Not by reading a book that says, "We can see the moon at night in the sky".
LeCunn and FAIR are experts at training self supervised systems in different domains. These models learn through simply observing patterns in large amounts of redundant data. So Wav2vec2, is set up in a way that it is just listening, and learning, listening, and learning. I-JEPA/V-JEPA and Dinov2 are viewing, and learning, viewing and learning. Through this type of process, a world model is trained, the world being the observed data. Similar to what @sohamdats said except that example is more of a case of supervised learning, in which the model is not simply observing, but being told. It's more akin to how a baby in its first months, figures out the contrast of light, then the geometry of shapes, and so on. The baby is forming its visual world model, given its observed data.
Full podcast episode: ua-cam.com/video/5t1vTLU7s40/v-deo.html
Lex Fridman podcast channel: ua-cam.com/users/lexfridman
Guest bio: Yann LeCun is the Chief AI Scientist at Meta, professor at NYU, Turing Award winner, and one of the most influential researchers in the history of AI.
00:00:26 Train a world model by observation without relying on gigantic data sets.
00:00:42 Explore innovative ideas that do not necessarily require scaling up.
00:00:49 Implement planning with a learned world model for non-physical systems like the internet or databases.
00:01:13 Develop a system to plan a sequence of actions for problem-solving in various scenarios.
00:01:54 Investigate hierarchical planning to handle complex tasks efficiently.
00:03:00 Learn how to represent hierarchical action plans for robots or intelligent systems.
00:03:13 Train systems to understand hierarchical representations of action plans using deep learning.
wonderful summary thx !!
… This man is so technical that even Lex didn’t know what to say
I noticed this as well😂
Asked for a piece of advice for PhD students, received a research statement as a response.
his alex net is such a historically important paper , i have read it before but it was taught to me today again by my professor.
Anderson's ACT-R goal stacking maybe. Add time to complete, action wait times/down time, and action ranking. Really want to mimic humans, add a need stack (charging, cleaning sensors, etc.)
So basically focus on robotics?
No, not just robotics. He’s saying that planning can be used for both non-physical (databases, task scheduling/chip design, etc.) and physical environments, which isn’t limited to just robotics.
In fact robots require A LOT of physical data, especially with RL systems. Operating in virtual simulation with planning can help robots navigate in the physical world easier with synthetic data.
TLDR he’s just saying that we should use hierarchical planning for all intelligent systems that require a series of complex actions
I got like 50 percent of that... maybe lol.
Haha.
Usually AI is good at doing one specific task. I think he is talking about programming AI systems that do multiple tasks in sequence, and understanding which is the next best task. In the example of a robot going from a city to another it would be: 1. Get out of the house. 2. Identify the best method of transportation. 3. Request an uber to the airport. Etc
He's saying current AI systems are only capable of doing multi-step (hierarchical) tasks when all the steps (plans) are defined and easily trainable, but how do you get a model to learn a multi-step task without defining the all plans
3:00 - millions of years of evolution. If we come to understand the brain better as a system and be able to directly database adaptations that include such hierarchical understanding and navigation through the world which all animals possess to some degree, it would improve AI as a discipline. Perhaps even render it novel folly in comparison to the super AI 'hardware' we already possess. Nature so often already has solutions to our modern civilized questions.
we ARE Natures Solution to advancement , I predict an emergent Species , possibly from genetic manipulation . Because living minds are more efficient
plenty of opportunity for Phd's theses in AI and Robotics --plenty of jobs open afterwards also
Why ignoring the whole work of Options in RL !!!
Hard to get his accent. Thank you for the subtitles
what does he mean by "how to train a world model by observation"
Just like how babies learn to recognise physical objects when her parent points to the moon and says, "This is the moon". Not by reading a book that says, "We can see the moon at night in the sky".
LeCunn and FAIR are experts at training self supervised systems in different domains. These models learn through simply observing patterns in large amounts of redundant data. So Wav2vec2, is set up in a way that it is just listening, and learning, listening, and learning. I-JEPA/V-JEPA and Dinov2 are viewing, and learning, viewing and learning. Through this type of process, a world model is trained, the world being the observed data. Similar to what @sohamdats said except that example is more of a case of supervised learning, in which the model is not simply observing, but being told. It's more akin to how a baby in its first months, figures out the contrast of light, then the geometry of shapes, and so on. The baby is forming its visual world model, given its observed data.
@@sohamdats @simonkotchou9644 Thanks
@@sohamdats Stuff like this inherently requires large scale in terms of computational resources though, right?
Very interesting snippet.
I want a real flesh and bones border collie, who shows me what humility , loyalty, intelligence and unconditional love is. 🐶
and you can eat it if times get tough , but a robot dog might gather food for you !
1:12
End to end Ai Planner like FSD 12?!??
🤔
What is this ?
Lexy PhD is a scam 😂😂😂😂 ask me 😂😂😂😂
Could you collaborate more?
@@Farinata2 Higher Education is Fake People pay money to get the degree same as rest ....so don't get PhD go to work
@@Farinata2 it's a ponzi scheme.
So, is he speaking French or English?
Annnnddd that’s your “ai experts” right there. Lost in the sauce