It's not in the networks themselves, but the combinatorial ways that we can string them together. Even if the language models themselves don't become much more

It's not in the networks themselves, but the combinatorial ways that we can string them together. Even if the language models themselves don't become much more powerful, there are a lot of really weird things to discover in terms of how we can pipeline them together to accomplish things. For example things like this: https://yoheinakajima.com/task-driven-autonomous-agent-utilizing-gpt-4-pinecone-and-langchain-for-diverse-applications/ https://yoheinakajima.com/task-driven-autonomous-agent-utili... In a way, you can also look at language models as being the new basic operation. Activations and floating point math are replaced by words and symbolic reasoning.