I have a question: why do companies that create big LLMs don’t use an algorithm that can change itself? I mean, if a company could create an algorithm that can change itself based on its weight allocation efficiency in solving problems, wouldn’t that be a step closer to AGI?
1 Like
Maybe because its quite a complex task to implement!
they are already using it , but there is not any perfect answer for it , and training an algorithm on big data will take very very long time and is too expensive , so its not something that they can make it perfect in one shoot , so each step they improve their solution but there is not any jump to the perfect solution ,
As Saman is pointing out one thing is the price of retraining and another thing is that you need to ensure convergence in the iterations, so that a better version is guaranteed based on some performance metrics. Read about the issues with RL algorithms and giving them total freedom of choices, you will learn why they are always used in controlled environments.
Best,
Rosa