Problem is context length: How much can one truly learn from their mistakes in 100 thousand tokens, or a million, or 10 million? This quote from Dwarkesh Patel is apt
How do you teach a kid to play a saxophone? You have her try to blow into one, listen to how it sounds, and adjust. Now imagine teaching saxophone this way instead: A student takes one attempt. The moment they make a mistake, you send them away and write detailed instructions about what went wrong. The next student reads your notes and tries to play Charlie Parker cold. When they fail, you refine the instructions for the next student. This just wouldn’t work. No matter how well honed your prompt is, no kid is just going to learn how to play saxophone from just reading your instructions. But this is the only modality we as users have to ‘teach’ LLMs anything.
If your proposal then extends to, “what if we had an infinite context length”, then you’d have an easier time just inventing continuous learning (discussed in the quoted article), which is often discussed as the largest barrier to a truly genius AI!
Problem is context length: How much can one truly learn from their mistakes in 100 thousand tokens, or a million, or 10 million? This quote from Dwarkesh Patel is apt
If your proposal then extends to, “what if we had an infinite context length”, then you’d have an easier time just inventing continuous learning (discussed in the quoted article), which is often discussed as the largest barrier to a truly genius AI!
agreed context is maybe the bottleneck.