Pinned
The world is not zero sum. People can create something out of nothing, and many do. Should be encouraged more.
Harald Schäfer
909 posts
- This take is completely wrong. China's Chinesium era is over. Roborock, DJI, Bambu, Anker, Insta360, ... make better products than their western counterparts. They're now good at ML too, GPUs may be all that's left. We should drop the condescending attitude, and just compete.The only real way for US startups to compete with china is on taste. I ain’t ever see a Chinese product that didn’t feel temu
- Deleting code is what separates us from the LLMs.i love deleting code
- It's not the prettiest datacenter, but it's ours. All openpilot's models are trained here. No cloud, just this room maintained by our technicians and engineers.
- 99% of all life by biomass learns just by mutating DNA, no neurons. Artificial neural nets taught us a lot, but to truly understand intelligence we also need to recreate DNA. I made a repo on artificial DNA. It solves tic-tac-toe reliably in 30min. github.com/haraschax/nogr…
- Interesting to see @ylecun arguing pure ViTs are not practical. We recently moved our compressors over to pure ViTs, no convs. They take longer to train, but work quite well. Left image is compressed to only 224 bytes, right is original.
- Very proud of the hardware team at @comma_ai for making this beautiful thing. Years ago we were shipping great software with almost embarrassing hardware. Over the years it got better, but with the comma 4, the hardware is at least as good as the software.
00:00 - Very excited about our paper. I've been working on this problem for 8 years. When I started end-to-end self-driving was controversial. Now we're the first to show that on-policy training in a world-model simulator can create agents that are useful in a real-world product. ThisWe wrote a paper about training driving agents end-to-end in a worldmodel simulator. Blog post: blog.comma.ai/mlsim
- It's weird seeing people on here rave about living in SF. I hated it. Building technology is not about hype, networking, making deals, or raising money. It's about putting your head down and making a little progress every day. SF unfortunately has a lot more of the former.
- Building things in the real world is the purest engineering pursuit. It's the "touching grass" version of engineering, and we should do it more.The factory is the product
- I've worked at comma for almost 7 years. Our goal has always been to make an autonomous driving agent, by letting an ML model learn to drive by watching humans. We've done that, and shipped it to users with open-source code. It was 3 years ago when we first shipped end-to-endThis post is unavailable.
- openpilot only uses ~5w! I'd say we're doing pretty well at intelligence per watt.Waymo is losing in every metric possible! FSD is >10x more efficient than a Waymo
- We currently take in about ~1000hrs of driving data every day. That means we get plenty of examples of quick reactions in unexpected scenes. Some examples:
00:00 - With the comma 4 release, we've leveled up again in intelligence per cubic centimeter. We don't ship the most capable self-driving system, but @comma_ai is #1 at shipping you useful capability in a tiny package.
