"Foundational AI companies love this one trick"
It's part of why they love agents and tools like cursor -> turns a problem that could've been one prompt and a few hundred tokens into dozens of prompts and thousands of tokens ;)
It's be nice if I could solve any problem by speccing it out in its entirety and then just implement. In reality, I have to iterate and course correct, as do agentic flows. You're right that the AI labs love it though, iterating like that is expensive.
I love this! My take on it for MCP: https://github.com/kordless/EvolveMCP
The last commit is from April 2023, should this post maybe have a (2023) tag? Two years is eons in this space.
Crazy that OpenAI only launched o1 in September 2024. Some of these ideas have been swirling for a while but it feels like we're in a special moment where they're getting turned into products.
The bigger picture goal here is to explore using prompts to generate new prompts
I see this as the same as a reasoning loop. This is the approach I use to quickly code up pseudo reasoning loops on local projects. Someone had asked in another thread "how can I get the LLM to generate a whole book", well, just like this. If it can keep prompting itself to ask "what would chapter N be?" until "THE END", then you get your book.
^
Excellent fun. Now just to create a prompt to show iterated LLMs are turing complete.
Let's see Paul Allen's prompt.
LLM quine when?
Repeat this sentence exactly.
https://chatgpt.com/share/680567e5-ea94-800d-83fe-ae24ec0045...
I feel that often getting LLMs to do things like mathematical problems or citation is much harder than simply writing software to achieve that same task.