@raysan5 The most vocal people in the community are the complete opposite of those who fork, use, and appreciate OSS.
The majority of people forking and copying your code you never hear from.
There's now ~30k openpilot users who have 55% of their miles driven by AI. So this sounds pretty plausible! How many miles will openpilot be driving in 2036?
@chesterzelaya We sell a single hardware product and don’t price discriminate. So we don’t get different profits out of different customers.
So bad customers are just ones with a bad attitude who make unreasonable demands of the company and support.
@yash1_ I tried porting our on-policy driving model trainer to train with RL (SAC) instead of supervised learning.
With a lot of effort codex 5.5 wrote it in 700 lines. After a couple days of refactors I got it down to 300.
@HotAisle I use agents a lot. But I find they cannot write minimal code yet. Our entire internal codebase is 90k lines. If we let agents do their thing that would inflate quickly.
@jokrvivek@Tesla@maticrobots@comma_ai Absolutely. Paying for robotics data without a product is guaranteed failure. There are obviously useful intermediaries in robotics. Why is there no teleop robotics product?
We just moved all of our metrics from the cloud to our own datacenter. Data goes directly from users to our computers for training, debugging, and metrics.
We've always trained models without the cloud, but now only user-facing services are left in the cloud.
We wrote our own distributed compute library inspired by ray: https://t.co/4wMYOk5Cxc
As @yassineyousfi_ mentioned it handles all of our non-training compute tasks. Machines that aren't training are always automatically available to pick up miniray tasks.
This describes our setup: https://t.co/gaIha9eF7K
@TheGuySwann@BristolHubert@MarioNawfal@comma_ai Maybe! Depends what you want to talk about. I like talking about the importance of open-source, have full control over products you buy, and not relying the cloud.
Not so interested in crypto and politics.
@eriklangille I find them helpful investigating bugs or codebases I am unfamiliar with. But if I rely on them there I learn less, so not sure if that's even a big win.
AI productivity gains are mostly an illusion. Agents trick you into thinking they're helping you.
I use agents a lot, and have them write tons of code. But if I look at what I actually end up merging or shipping, it's almost exclusively stuff I wrote by hand.