My thesis is that lots of companies are going to want to enable OpenAI for their teams, but they'll need a way to do it safely/securely. Usage Panda acts as a proxy (either cloud or self-hosted) to enforce policies on these API calls before they're sent to OpenAI, so you can do neat things like cost monitoring, rate limiting, data redaction, etc.
It's not actually an _AI_ product (in that it doesn't use AI itself), but I'm building it to support AI usage, so hopefully that counts!
The idea is I wanted a game where you could just make any concept and the game plays that accordingly. You could play as a Marionette Master, a tourist from the future, a Man-Bat. You could search for secret passages using your elf vision, echolocation, or a spell. You could fight monsters with acid breath or Hamon. The game does handle balance though.
What works: Character generation engine, primitive inventory, conflict resolution, injury tracking.
It is incomplete at two challenges though. I'll add more if anyone gets past the second. It's also quite railroaded for balance, but I've been removing some of the rails gradually and it's turning out far better than expected.
There is an easter egg I tried to hack in within an hour. It worked so extremely well that I'd actually like to make it the main feature. You can type whatever you want at char creation (and only there). Feel free to experiment with prompt injections too.
All not worth the FOMO
https://harmonote.com/songs/nfUjirHCnasBZ9qQrsbzM8-heat-wave...
https://harmonote.com/songs/cDPf5CmkvenrngevyfckNd-alone
https://harmonote.com/songs/EkZnDvV4LFAJAHG34EBwXC-paper-thi...
That said, it's surprisingly tricky to build. GPT4 can use tools in theory, but it'll happily do things like hallucinate tools or tool options that don't actually exist. The non-determinism is a pretty unique problem to have to deal with.
.