Recently, I took some time over the weekend to carefully read through an interview with OpenAI founder Ilya. This interview is worth watching a few times. Besides discussing the shift from the scaling era to the research era (where intelligence can’t just be achieved by endlessly expanding computational power), what impressed me the most was his perspective on 'research taste.' This 'taste' is what allows him, during the research process, to navigate highly uncertain matters by using his own taste (beliefs and experience) to verify things top-down. In AI, this belief is fundamentally tied to the anthropomorphism of neural networks (the principles of the human brain). This sense of taste is foundational. When experiments don’t align with beliefs, it might sometimes be due to bugs in the data itself. But if you only focus on the present and the visible data—things you already know—you might miss the truly correct path. This research taste isn’t just applicable to AI LLM research. Whether you’re starting a business, investing, hunting for airdrops, or developing new products, you’re always dealing with highly uncertain situations. Your taste is your fundamental understanding of the essence of things or the basic rules of certain phenomena—those foundational dimensions. For example, if you’re a product manager and you notice a feature that almost no one uses, you might conclude that users don’t need it and decide to cut it. But it’s also possible that the design is flawed, and users simply didn’t notice the feature. If you lack product taste, you might make decisions and choices based solely on the partial information you can see. Thinking back to many years ago, I remember reading technical articles on deep learning on CSDN that explained how to use RNNs to perform addition, subtraction, multiplication, and division. At the time, I just thought it was fascinating. But Ilya’s curiosity had already connected the dots—if RNNs could predict basic arithmetic, they should be able to handle more complex tasks. He also realized that the foundation of neural networks comes from mimicking the structure of the brain as a theoretical basis. These two points were crucial in building Ilya’s foundational taste for exploring LLM intelligence research. https://www.(dwarkesh.com)/p/ilya-sutskever-2