This is a common rebuttal. But it falls flat, in my opinion, because LLMs appear to be a compression of language-based knowledge, and that compression is lossy. The result is that "knowledge" retrieved from the language-based storage model is biased towards the "average" idea.

As a concrete example, when you ask Claude Sonnet 5.5 to generate a small-to-medium-sized one-shot Python project, it will always fall into the same code patterns. Same structures of functions. Same file layout. Same package layout. They're not even good patterns, because if you throw a few common linters at it (e.g. flake8, isort, mypy) there are a ton of linting errors to clean up.

The only way that it escapes this mediocrity is through context. If you give it enough guardrails or sample data (e.g. Python code to emulate), it will start to go off in a "novel" direction that is defined by that context.

Now, what's the most likely way that "novelty" will be generated in an automated system like this? Randomly pulling in context and going for a random walk through the compressed language-based knowledge storage. It remains to be seen whether that's an effective way of producing more than just parlor tricks for investors. Human+LLM is overwhelmingly more novel and effective than only LLM.