about JevTTD (Jev x OpenTTD)
I wanted to see how far I could push Jev as a system one model, it's extremely fast and inexpensive to run, which makes it attractive for agents that need to make lots of decisions. But it also creates an interesting constraint, because Jev is not a large general-purpose reasoning model, and there is not currently a practical way for me to fine-tune it for a specific task. I heard a lot of hackers trying to take advantage of Jev and they let Jev handle a lot of things, but there is currently no specific way to tell, "is my way of using Jev good for production?"
That led to the central idea behind this project, which is: What if, instead of fine-tuning Jev itself, we tune the harness via the SDK, which controls its inputs and typed state beliefs?
The goal is to use a slower but more capable LLM like Astra during an optimization phase to repeatedly improve the harness through which Jev observes the environment and takes actions. The LLM can change what information Jev receives, derive more useful features, reorganize the available actions, or introduce higher-level programmatic abstractions. Jev itself remains frozen.
Each candidate harness is evaluated on the same task. The LLM sees the results and modifies the harness, while the expensive model is used during development to search for a better interface. Then the final deployed agent can continue using the much faster and cheaper Jev with the discoveries and gains.
The hypothesis is that a significant amount of an agent's apparent intelligence may come not only from the model itself, but from the quality of the abstractions exposed to it.
Games provide a natural environment for testing this idea because they are cheap, with clear state, repeatable actions, and transparent reward signals. Instead of subjectively deciding whether one harness is "better," we can run the agent repeatedly and measure whether it actually completes more objectives, earns a higher score, or reaches its goals more efficiently.
OpenTTD is a sim game with players acting as transport company managers. It was an good test bed for this experiment (open source, controllable, sexy visuals in a live demo, and complex enough to see if the harness matters). I tested Jev with both very low-level controls like placing individual pieces of infrastructure and managing vehicles, or much higher-level actions such as identifying and building profitable transport routes.
Log in or sign up for Devpost to join the conversation.