Deep learning is a glorified form of curve fitting. Being able to predict the next token/frame is good if your goal is to fake the ability to generate text or v

Deep learning is a glorified form of curve fitting. Being able to predict the next token/frame is good if your goal is to fake the ability to generate text or video, but doesn’t yield a deep understanding of the world as proven by the inability of SOTA models like GPT-4 to do any sort of serious reasoning reliably. The same is true for the latest text to video model of OpenAI called Sora, which is incapable of knowing the basic laws of physic despite being trained on an enormous amount of videos. These models excel at generating human-like text and videos by predicting the next token or frame based on vast datasets they have been trained on. However, their capabilities should not be mistaken for a path toward achieving Artificial General Intelligence (AGI).