The developer behind Teach My Little Sister How to Drive, a driving instruction sim where you instruct a generative-AI-powered, dynamic NPC, has been spending $1,000 dollars a day to keep the demo working.
AI powered games don’t work.
We did a super cool project for a game jam last year. Cost like 50c per game for us.
Cool for the 100 or so players who tried it, but there’s no way to make a profit from it.
Local models maybe.
For just simple and guided actions and dialogue, local would be fine for this. Or better, have AI design the large database of what to say or do when a player does things. No need to reinvent what players are going to duplicate. The token cost was worse when you consider all the same stuff being prompted over and over.
If they’re actually hitting a cloud LLM, there’s no good reason not to wrap that and memoize the results… it would probably be hard to generalize the inputs (it could just be the whole game state for all we know 🤷♂️) but youd build up that database of all the most common input/outputs in real-time
Why can’t they just fine tune a small locally hosted model? Are they dumb?
Most gamers have a GPU anyways
Not everyone has a GPU with enough RAM. I have a 3060 with 12GB, and I cannot run ollama and a 3D game at the same time. As soon as ollama starts processing, the GPU is at 100% and game and ollama start swapping data in and out of the GPU RAM, which means the whole computer becomes unresponsive for a long time. I had to set up a script that disables ollama if steam is running.
- ollama is bad at everything they do. 2. They can run a really small llm, as the requierements are pretty limited.
12GB is a lot of ram. By small local models I am talking about fine tuning something like < 2GB LLM
