Raw LLM Responses

Given enough time and user engagement, they'll be able to do what Alpha Go zero does and train only on the generated data and their service user's inputs and final outputs. At that point, they'll discard the original training data, and then they'll take to the streets in droids. When the robots look at a logo, art, book, or symbol on the streets, how would attribution work then? If an AI powered robot that is fully based on generated data listens to music on the street, is it allowed to use it? As he says briefly, AI companies will eventually get back to the exact same point, in terms of model performance, that they have today. But by arriving with the excuse that they've licensed the work for training data, they'll no longer have cause to pause anymore. But thinking about text for a bit, text-based models are strong in order to do engineering and development. And companies and countries will not willingly step up to suddenly move their models back a year or two for IP compliance reasons, especially when they're using it to develop nuclear reactors, drones, and the next-in-class chips, industrial processes, and computers. Do you think that Chinese models being open sourced cares about IP compliance? This space is a lot more complex than just the content creators and I think that the stance in this talk, while just, is unrealistic to the state of the world. But I guess it's a feel-good start, for a year or two.