LLMs are so well tuned for optimizing the lighting pathfinding problem. Give them a dataset and tell them to optimize a parameter and they go ham. Big things ahead for the LDK pathfinder and lighting payment success rates.

Replies (11)

Does this mean that llms are also well-tuned to figuring out the likely path of a payment? That reliability might come with a price.
Constant's avatar
Constant 2 days ago
they are a traveling salesman at heart
Not directly, no. But if you have a large dataset to build an algorithm from…..
The same old probing dataset we used to optimize the scoring analysis last time, but with a year and a half more data, a bugfix in the harness, and LLMs trying several totally different approaches to investigate potential directions.
After I land some updates maybe I’ll put up a website and let people submit their results and throw their agents at the problem.
Sadly it really wants to over-tune for nodes that send a lot of probes, which isn’t really ideal, but may be able to push it towards having options for everyone.