There have always been debates about whether AI can compete with the human brain or not. Some people think that it can but Yann LeCun, Meta’s chief AI scientist, says that it's not possible. He says that they can make AI come close to the human brain in ten years only by using a method called “world model”. OpenAI has already released a new feature on ChatGPT called “Memories” that makes the AI model remember all of the conversations. OpenAI says that the model they used for it, o1, can also make AI capable of complex reasoning.
Image: Hudson Forum / YT
Even though it seems like we are closer to general artificial intelligence (AGI), LeCun says that we have a long way to go to achieve it. He said that even though there are many AI optimists claiming that human-like AI is closer, it is still going to take a lot of time. He added that the latest large language models like MetaAI and ChatGPT are still far from human-level AI. The reason for this is obvious. Most large language models are one dimensional which means that they can predict things in just one way and cannot understand the three dimensional world.
This is the reason AI models cannot do tasks a human can. AIs are not reliable in the physical world because they cannot even do simple tasks even if they are trained on millions of data. LeCun suggests that building a three dimensional world, Work Models, for AI can be a solution. He described the world model as a mental model of how you perceive a world behaving. Try to understand it with an example of a dirty room. When a human looks at a dirty room, he makes a mental image of what he has to do to clean it. Your brain looks at the space in three dimensions and creates a plan to achieve the goal of cleaning the room.
This is what the AI world model is trying to achieve and a world model can also take more data than LLMs. Many AI labs are chasing this idea and trying to start their work on it. LeCun says that Meta AI research lab, FAIR (Fundamental AI Research), is also working towards this idea. LeCun says that it isn't easy to achieve and there are a lot of complicated things they would have to work with. So, it's probably going to take a lot of years.
Read next: Chip Industry Faces Talent Shortage As It Makes Its Way To 1 Trillion Dollars Revenue
Even though it seems like we are closer to general artificial intelligence (AGI), LeCun says that we have a long way to go to achieve it. He said that even though there are many AI optimists claiming that human-like AI is closer, it is still going to take a lot of time. He added that the latest large language models like MetaAI and ChatGPT are still far from human-level AI. The reason for this is obvious. Most large language models are one dimensional which means that they can predict things in just one way and cannot understand the three dimensional world.
This is the reason AI models cannot do tasks a human can. AIs are not reliable in the physical world because they cannot even do simple tasks even if they are trained on millions of data. LeCun suggests that building a three dimensional world, Work Models, for AI can be a solution. He described the world model as a mental model of how you perceive a world behaving. Try to understand it with an example of a dirty room. When a human looks at a dirty room, he makes a mental image of what he has to do to clean it. Your brain looks at the space in three dimensions and creates a plan to achieve the goal of cleaning the room.
This is what the AI world model is trying to achieve and a world model can also take more data than LLMs. Many AI labs are chasing this idea and trying to start their work on it. LeCun says that Meta AI research lab, FAIR (Fundamental AI Research), is also working towards this idea. LeCun says that it isn't easy to achieve and there are a lot of complicated things they would have to work with. So, it's probably going to take a lot of years.
Read next: Chip Industry Faces Talent Shortage As It Makes Its Way To 1 Trillion Dollars Revenue
