Rohan Anil and Boaz Barak argue researchers have substantial control and understanding over AI model training and behavior
Anil previously worked on Gemini pretraining at Google DeepMind.
#133Boaz Barak@BOAZBARAKTCSI agree. There is a very real sense in which we are building AI models and have a significant amount of control and understanding over them, including the causes of their behavior and how to modify them. I wrote this in my "non review" of if anyone builds it, everyone dies.
rohan anil@_arohan_I am a bit bumped that scientists go to the world and say they don't what they are doing inside the lab. It's humble thing to do (like we all aspire to understand things deeply), but we know a lot more, how these models are trained, how scaling laws are done, how mixtures are created, what training data influences what behavior, why certain models failed and why some succeeded. If assumption this is building god, then please cite all the papers that god was created from, and their authors be elevated to heaven. When you go say we don't know what is happening, people take it quite literally which might be alarming for someone reading the news a million miles away.
@boazbaraktcs Can't both be true? We understand many of the parts but that doesn't mean we understand the whole sufficiently at the same time
Boaz Barak@boazbaraktcsI agree. There is a very real sense in which we are building AI models and have a significant amount of control and understanding over them, including the causes of their behavior and how to modify them. I wrote this in my "non review" of if anyone builds it, everyone dies.
@_arohan_ Isn't it true though? We can predict the outcome of small and large interventions but I don't think this is enough because there are plenty of interactions between parts we can't predict and behavior we cannot guarantee
rohan anil@_arohan_I am a bit bumped that scientists go to the world and say they don't what they are doing inside the lab. It's humble thing to do (like we all aspire to understand things deeply), but we know a lot more, how these models are trained, how scaling laws are done, how mixtures are created, what training data influences what behavior, why certain models failed and why some succeeded. If assumption this is building god, then please cite all the papers that god was created from, and their authors be elevated to heaven. When you go say we don't know what is happening, people take it quite literally which might be alarming for someone reading the news a million miles away.