I don’t have a strong opinion on hiding nostalgebraist’s post behind a login gate. But as a data point, I’m not affiliated with any major labs and I don’t currently work on LLMs but in other AI fields and I still read the Chinchilla paper before seeing this post (and I wasn’t surprised by its results), so hiding the post wouldn’t have made much of a difference for me.
However, I am very surprised that a report like ELK is publicly available for any web crawl to access. I think that if you query a future LLM that has this report in its training data and prompt it with a query related to hiding its intentions, you’ll get much better results. Is this desirable? It seems to provide a shortcut to a lot of knowledge that an LLM would need to reason about by itself without access.
I don’t have a strong opinion on hiding nostalgebraist’s post behind a login gate. But as a data point, I’m not affiliated with any major labs and I don’t currently work on LLMs but in other AI fields and I still read the Chinchilla paper before seeing this post (and I wasn’t surprised by its results), so hiding the post wouldn’t have made much of a difference for me.
However, I am very surprised that a report like ELK is publicly available for any web crawl to access. I think that if you query a future LLM that has this report in its training data and prompt it with a query related to hiding its intentions, you’ll get much better results. Is this desirable? It seems to provide a shortcut to a lot of knowledge that an LLM would need to reason about by itself without access.
Yeah a few people have also brought up this concern recently. Will think about it.