Hi all!
The past week after the update with discrete abilities we had several play tests that went really well. I’m starting to feel like we’ve got a core experience that’s really fun, at least for single player, the longterm quest system especially feels like it makes a huge difference on the experience relative to AI Dungeon.
We’re still trying to decide if we’re going to have multiplayer (will have more to discuss on that in the next week) in the first early access version, but regardless it really feels like we’ve crossed a threshold. I’m confident that you’re going to love playing Heroes 😄.
There’s still lots of refining to do, but the core experience is there and it solves a lot of the problems that prevented AI Dungeon from being a fun RPG like we wanted.
With that done, we’re now looking towards the process of productionizing Heroes so that we can let as many people as possible play it. I’ve spent a lot of time the last week thinking and laying the ground work for that, particularly on fleshing out our processes for making the AI cheap enough to run. Our goal is to eventually get it to the point that we can let free players play at least a limited amount, though it may take some time to get there.
This is something that’s much more difficult than most people realize, but we have quite a bit of experience doing this kind of thing with AI Dungeon. And we also have done some early work proving this process out on Heroes, so I have a high degree of confidence we can do this.
Essentially we need to take this AI game engine, that uses several different AI components based on larger models, like GPT-4, and refine it down into much cheaper models. And we need to do all of this while not just maintaining the same accuracy as on the large models, but improving it.
We’ve tested this on a few components already, and were actually able to get a better accuracy than GPT-4 by training custom models, so we’re pretty confident that we can maintain or even improve on the fun of the prototype version.
Essentially at a high level the process looks like this:
- Use the big models to prototype what the different pieces look like.
- Test and refine the instructions for the big models to make sure things are working well.
- Use the big models to gather a bunch of data of those tasks.
- Prune out the bad data where even the big models failed, leaving only the good data (as much as you can)
- Train the small models.
- Keep refining based on feedback till it’s good enough.
- Profit. (or at least break even 😅. These things are expensive!)
One of the things that really makes this possible is all the really great work that all those in the open source model community have done to make better and better open source models. Big shout out to everyone who is contributing to create value that all of us can build on. We probably couldn’t do this if we had to only rely on the big closed source providers, so we’re extremely grateful for all of you. ❤️
– Nick
