My tiny model just beat @comma_ai's openpilot on a tight bend. First time ever. I'm trying to get there on my own drives, using distillation. I call it Route42.
11 training runs, 267,000 steps, 40+ hours of training so far.
Distillation means my small model learns by copying comma's big driving model. Along with the road camera, it also gets a small map of the road ahead, a lot like the map on your phone's nav screen, so it can see a bend coming before the camera does.
For training run 8 I didn't change how it learns at all. I started it from my best model so far, training run 4, and gave it 66 new clips from my own drives with real curves, turns and stops.
Before this, I spent three training runs just changing how much each old drive counted, and that went nowhere. More real curves did it.
On that bend it was off by 3.17 m on average from the path I actually drove. Lower is better. Stock openpilot was off by 3.41 and my old best was 5.19.
On drives it never trained on, its error score dropped from 135.4 to 81.8.
It's not a clean sweep. It's a little worse on two other clips, openpilot still wins the full bend clip, and some new clips are probably on the same roads as the test clips.
Next I'll keep running Route42 to see if my small model can beat the big openpilot model everywhere, not just on one bend.
To follow along, the first comment of each post will have the prior post update. π§Ύβ¬οΈ
"Do I think a Tesla / SpaceX merger would be a good idea? Yes, but the shareholders will need to vote yes. And Tesla will need 50% of all shareholders to vote yes, which is a high bar to achieve"
β @elonmusk, Tesla and SpaceX CEO
It's really funny looking back at this video from 2018 where I broke down a little bit of the physics of what it would take to make a Roadster fly. π¨ A bit of cringe warning... old spacesuit Tim alert π¨π
You can install openpilot on a 2019 corolla, an 8 year old car, and it will stop at stoplights for you.
If @comma_ai didn't exist, old cars would almost certainly never get capability like this. And most people would claim a project like openpilot could never exist.
Many cool things only exist because a few people decided it should. You can just do things, and more people should!
The new location will not be in California, because power is criminally expensive here.
We'll need robots to remotely fix things sometimes, also hiring people to make those robots work!
Our GPUs have outgrown our little San Diego office. In 2027 we're starting up a bigger better datacenter with more GPUs. We are hiring engineers to work on this.
If you know linux and like sshing into machines that cost more than a Ferrari, you might be a good fit.