I’d wager they’re attempting to replicate or integrate tools developed by the open source community or which got revealed by Meta’s leak of Llama source code. The problem is, all of those were largely built on the back of Meta’s work or were cludged together solutions made by OSS nerds who banged something together into a specific use case, often without many of the protections that would be required by a company who might be liable for the results of their software since they want to monetize it.

Now, the problem is that Meta’s Llama source code is not based on GPT-4. GPT-4 is having to reverse engineer a lot of those useful traits and tools and retrofit it into their pre-existing code. They’re obviously hitting technical hurdles somewhere in that process, but I couldn’t say exactly where or why.