The lowest hanging fruit aren't even that pie in the sky. The LLM doesn't need to be capable of original thought and research to be worth hundreds of billions, they just need to be smart enough to apply logic to analyze existing human text. It's not only a lot more achievable than a super AI that can control a bunch of lab equipment and run experiments, but also fits the current paradigm of training the LLMs on large text datasets.

The US Code and Code of Federal Regulations are on the order of 100 million tokens each. Court precedent contains at least 1000x as many tokens [1], when the former are already far beyond the ability of any one human to comprehend in a lifetime. Now multiply that by every jurisdiction in the world.

An industry of semi-intelligent agents that can be trusted to do legal research and can be scaled with compute power would be worth hundreds of billions globally just based on legal and regulatory applications alone. Allowing any random employee to ask the bot "Can I legally do X?" is worth a lot of money.

[1] based on the size of the datasets I've downloaded from the Caselaw project.

No amount of intelligence can do this without the experimental data to back it up.