Assuming online training for causality follows a similar trajectory to the tasks studied here (language modeling for code or natural language), we would expect scaling to reduce the amount of additional signal needed to get good at causality. But obviously that’s an empirical question that would be good to get an answer for as you mention in the article!
Assuming online training for causality follows a similar trajectory to the tasks studied here (language modeling for code or natural language), we would expect scaling to reduce the amount of additional signal needed to get good at causality. But obviously that’s an empirical question that would be good to get an answer for as you mention in the article!
Are you aware of any examples of the opposite happening? I guess it should for some tasks.