LARRUZ breaks down Anthropic's new research on what's happening inside LLMs. A 3-minute demo shows how changing one word in a prompt reveals the model's thinking process during poem generation, moving closer to understanding these black boxes.
For the first time ever, we are starting to understand what LLMs like ChatGPT actually think behind the scenes. These insights come from a brand new research paper by Anthropic. They're basically starting to decode what goes inside of the black box inside of an LLM. If you're not familiar so far, even the most sophisticated people building these models did not fully understand the end to end process of how these models were generating such level of quality and showing these levels of intelligence within what they're doing. And and they created this little three minute video that I recommend you watch if this piqued your curiosity, in which they showed how AI writes a poem and how they can infer on how it's actually thinking depending on one word that is being switched out. I'm just saying this is a real step in that direction. And honestly, the first glimmer of hope showing that this problem might actually be solvable.
