Description
Hi Pawel,
I thought you may be interested in collaborating with me and a few other folks (with relevant GitHub repos) to build an AI that can win races. I left Google to work with some famous sailors (Stan Honey, Steve Benjamin, etc.) building out a sailing data platform. We have the largest data set of sailing races on earth and are building a platform to open it up to everyone. It’s billions of positions for hundreds of thousands of races across everything from numerous Vendée Globes to casual dinghy races. It's about 1tb of just position/timestamp data dating back the last 20 years. We also have terabytes of weather and sea state data for those races. I am finishing up the process now of converting the weather data to Zarr and writing routines to use it in Dask. I also have all 40k ORC polars. My vision is this: every race ever, with every polar ever and all weather data, all accessible in a RAPIDS Jupiter lab notebook with Dask accelerated routing and industry standard ML libraries. I am pretty close to that now, but I need help with the next step. What I find very interesting is, once we have this environment, can we build an AI agent who wins races? I’ve also been thinking about how we can integrate this with OpenAI Gym. The goal would be to train the agent and then compete in the races following the decisions of the agent and make everything open source. Anyway, let me know if you’re interested or want to hear more. My organization has aws budget and I can spin something up and give you and others access including possibly a small hpc cluster.
Jon