Those scales are log so "years" more training may be an improvement but in absolute terms it may not be worth running. One point at which the cost vs return doe

Those scales are log so "years" more training may be an improvement but in absolute terms it may not be worth running. One point at which the cost vs return doesn't make sense to keep running, another point at which new approaches to building LLMs can quickly give a better result than training the old model's training for years would give anyways. There is probably also a practical limit at which it does truly flatten, it's probably just well past either of those points so it might as well not exist.