What a fantastic analysis. It seems the more we peel back the layers, the more we are seeing that a lot of the reasoning these models demonstrate could be, as a previous researcher said, a "mirage", that is being generated from its mimicry of the training data.
I use claude 2 mostly for my research purpose in Humanities. It retains data even in larger pdfs better than gpt 4 used to do. I haven't used the gpt 4 128k yet, i hope 2.1's 200k will also improve data retrieval better.
Nice insights! How could mechanistic interpretability play a role in this? I feel like the analyses we’re making today (like this one) will be resolved in the future by having a good theory of what happens inside NNs
It’s funny, the more we look at the problems the more we find ways to quickly work around some issues. I saw a RAG pattern recently where the result set of a search is spread so the most relevant results are at the top and bottom of the values used to expose the most relevant data to the LLM because of these sorts of limitations Where we find problems there are creative solutions. I think now that we have a better handle on why we’re seeing this, we can better address it with training data or other approaches that may change things fundamentally
A good artificial language should be context-free. If you want to find fault with AI it is the underlying language: English. Take for example the sentence: in the box are apples and oranges or bananas. Ask the AI monkey: will I find an apple when I open the box? No right answer. There you go. What good is context? It just messes up a language.
What a fantastic analysis. It seems the more we peel back the layers, the more we are seeing that a lot of the reasoning these models demonstrate could be, as a previous researcher said, a "mirage", that is being generated from its mimicry of the training data.
Good stuff. Subbed
Thanks for the sub!
I use claude 2 mostly for my research purpose in Humanities. It retains data even in larger pdfs better than gpt 4 used to do. I haven't used the gpt 4 128k yet, i hope 2.1's 200k will also improve data retrieval better.
Nice insights! How could mechanistic interpretability play a role in this? I feel like the analyses we’re making today (like this one) will be resolved in the future by having a good theory of what happens inside NNs
It’s funny, the more we look at the problems the more we find ways to quickly work around some issues. I saw a RAG pattern recently where the result set of a search is spread so the most relevant results are at the top and bottom of the values used to expose the most relevant data to the LLM because of these sorts of limitations
Where we find problems there are creative solutions. I think now that we have a better handle on why we’re seeing this, we can better address it with training data or other approaches that may change things fundamentally
A good artificial language should be context-free. If you want to find fault with AI it is the underlying language: English. Take for example the sentence: in the box are apples and oranges or bananas. Ask the AI monkey: will I find an apple when I open the box? No right answer. There you go. What good is context? It just messes up a language.
makes a good point