In an interview with Nicolai Tangen, CEO of Norway's wealth fund, Elon Musk said Artificial General Intelligence (AGI) could arrive by 2026.
Meta AI Chief and Turning Award Winner Yann LeCun thinks that is nonsense. "Eventually, machines will surpass human intelligence… it’s gonna take a while though,” he said. “It’s not just around the corner — and it’s certainly not next year like our friend Elon has said.”
LeCun's ideas are featured predominately in my book, The Intelligence Tsunami (https://www.amazon.com/dp/B0CXH5DDRP)
First, LeCun doesn't think AGI is a useful term. He doesn't believe humans have general intelligence. Humans have a wide range of both natural intelligence and training in specialized areas. He doesn't think average human intelligence is a useful concept.
LeCun also thinks that animal intelligence is even more broad. He has noted that orangutans are more intelligent than humans in ways that allow them to survive in their natural environments. Most humans wouldn't live long in an orangutan world. He believes artificial digital agents will likely develop specialized intelligence as humans and animals do. In some ways, digital agents already far surpass human capabilities. In other ways. it will be a long time before they do, if they ever do. LeCun has another issue with Musk's 2026 forecast. LeCun doesn't believe that digital neural networks like those underlying ChatGPT and Google Gemini will ever be truly intelligent. These large language models (LLM) have an amazing ability to manipulate words and trick us into believing they are fluent. But they have no real understanding of what their words mean. Humans judge whether they are accurate or hallucinating. LLMs can't tell the difference. Remember that hallucination first described a human perceiving something that doesn't actually exist. Humans make errors, too.
“Without reasoning, planning, persistent memory, and understanding the physical world (capabilities which are four essential characteristics of human intelligence), that current AI systems can’t do, AI applications remain limited and error-prone,” LeCun said. He observes that LLMs predict the next token, a word or word segment, in a sentence. That is a discrete problem, but there are a limited number of words to choose from in any language. Most human perception is of the physical world, which is a much more complicated problem than predicting the next word. As you pan your eyes around, it is difficult, if not impossible, to predict what you will see next.
LeCun has proposed Hierarchical - Joint Embedding Predictive Architecture (H-JEPA) to achieve these four characteristics of intelligence. His model is based on digital agents trained with self-supervised learning by finding patterns and relationships within the data itself without the need for manually labeled examples.
Human brains don't process all of the sensory information they receive. They would be overwhelmed. Our brain has a world model of reality and is amazing at selecting the important parts of the sensory data received and matching that against the world model to understand the physical environment.
JEPA does something similar. It doesn't attempt to predict full reality, which isn't possible, but rather attempts to predict a compressed representation (embedding) of the data. For example, when you are driving, your brain makes no effort to predict the fluttering of the leaves on the trees beside the road. That would be incredibly complex to attempt to predict, and it's not relevant to your ability to drive.
H-JEPA takes that a step further. When you drive to a store, it is impossible for you to plan every turn of the steering wheel when you leave your driveway. Your brain breaks your trips in chucks, and then those chucks into sub-chucks. The information leading to the action of your turning the steering wheel is in the sub-chuck you are currently processing. Your brain processes through these chunks and sub-chunks as you go. H-JEPA is an effort to replicate this process of our brains as a path to true digital intelligence.
LeCun is the first to acknowledge that there is much work to do before H-JEPA is a reality. Perhaps others will develop better concepts for true intelligence.
LeCun believes what is almost certain, though, is that no one will get to true digital intelligence by 2026, or anywhere close to that.
Let's talk.
Let's inspire your team and your organization to excel.
John Warner
864-561-6609
