>Meaning legal system will have to bend towards reality and copyright is dead in LLM context and thus many other contexts as well.

I certainly hope so. There are already a few text and data mining exemptions in place in Japan, Israel and Europe. The US has fair use, so we'll see what ends up being covered and what needs to be licensed. In my view just scaling up llms with more data isn't the future either way. We see this with llama 3, Claude 3, coheres model and gpt 4 being all pretty close together on llm leaderboards. If anything progress seems to be slowing down.