Am I right to conclude all these AI platforms that do things like code review are just increasingly irrelevant wrappers around the top models? No one is actually training or even fine tuning models at this point, right?

I remember when you had to put in these little hacks like to get LLMs to reply to complete the phrase:

me: England

llm: London

me: France

llm: ?

And on every iteration you need less and less clever engineering. At this point I don't even need to specify where to look for things, the LLM will just figure it out. And my overly specified prompts sometimes even hurt me as it's too narrow. I think I read that Harvey for law for instance actually scores worse on legal exams than just using the models out of the box.

These companies that build rails for these system and use clever prompts can't be adding value. And if they are, they'll have to completely re-engineer everything every 6m when a new model comes out.

There is value in rails for sure, but those will likely be general rails like Open Claw and even that will be incorporated into the base layer