Woven x26. Hit a wall with scala. Turns out, an AI can write scala and do it well, but completely melts on multi-project builds. This is a problem when you've got a single project that needs five API's.
So we ported it back to go. This puts me half a day behind, but it's worth it.
Got all the way into testing Cornerstone (the chat service component with the tool layer) and the system started looping and hallucinating like crazy.
Then, I remembered... we forgot to implement the [🐼: Panda Protocol]. It's a notation and labelling system that keeps them from doing that. It doesn't have to be pandas, but remember AI's love emojis. They give them more attention than regular text. But the rule of thumb here, is that if you would see the context window and be confused by it.... chances are good that the AI would too.
Looping and hallucinations come from bad data stacks, and bad or incomplete process... not faulty LLM's.
The whole thing is 100% preventable, and well documented.
Don't believe anyone who tells you otherwise.