Humans don't think logically. We feel, then we justify. That's because our brain runs on shortcuts called heuristics. They save time, but they also trick us. Let's talk about how your brain cheats you every single day. (a thread)

When my cousin was 4 years old, we realized he wasn't like other kids. While most children played with toys, he spent hours flipping through books, literally any book he could find. His parents noticed this early and encouraged him, giving him every resource they could get their hands on. Then something incredible happened. By the time he turned 5, he could speak Marathi fluently. No one in the family spoke Marathi, which made it even stranger. And one day, he sat down at a piano and started playing simple melodies perfectly. That's when his parents realized something profound. Even though each book contained only bits of information, he was subconsciously connecting knowledge across all of them. He wasn't memorizing, he was learning how things relate, how they fit together. He didn't need someone to tell him what to do. He just learned the underlying structure and from there he could do anything: read new languages, play instruments, even make logical guesses about things he'd never seen before. For years, I couldn't stop thinking about it. How could someone trained on nothing but patterns and exposure suddenly start reasoning? Then it hit me. That's exactly what Large Language Models do. LLMs are trained for one simple task - predicting the next word. They are not explicitly trained for text generation, question answering or text summarization. Yet, when prompted, they can perform all these tasks remarkably well. Earlier NLP models were built for specific, narrow tasks like sentiment analysis or translation and couldn't generalize beyond that. But LLMs, despite being trained on a single objective, started performing a wide range of NLP tasks with ease. It was observed that the architecture built for a simple next word prediction task is very effective for many tasks as well. This phenomenon is primarily due to the the vastness and diversity of the data used in training. For instance, GPT-3 was trained on 300B tokens from Reddit submissions, Wikipedia and countless books. So it has access to a wide variety of text styles and information. Though it is only trying to predict the next token, it ends up building an internal model of how the world and language work. To be precise, it now possesses knowledge about grammar, semantics, logic and common sense reasoning. So the remarkable abilities of GPT aren't the result of explicit instruction, they are emergent side effects of scale, data, and architecture.