Dunno! It could be the case that there just needs to be a trillion parameters to be useful enough outside of highly-constrained scenarios. But I would certainly challenge those who work on LLMs to figure out how to require far less compute for the same outcome.