Recent content by eli_

i Be the change you want to see. Plenty of value in discussing things from all different framings and levels.

Not crashing or killing people is such a low bar and not even the difficult part. A driving agent also needs to complete trips without getting stuck, it needs to be actually autonomous (no humans in loop, either in car, the field or call center), it needs to be comfortable, it needs to be...

The hard part of self-driving isn't achieving redundancy, it's not the hardware, it's not getting the permits, it's not the sensors. The hard part is 100% the software. The big improvements to come will be from how the model is trained, going beyond imitation learning. The more of the world...

I suspect the end-to-end model can effectively do some of its own localized navigation, so it might not follow the onscreen navigation exactly. OpenPilot's end-to-end model works this way, they render a map with a desired driving route highlighted and the model uses that to figure out where to...

Yeah there was one point in time where cars with Enhanced Autopilot could buy FSD and literally the only added feature was stopping for traffic lights on TACC or Autosteer. Before that there were zero features for your money so at least it was something.

That's not how that feature works, it doesn't set some secret nav route thats hidden from you. Auto navigation just pre-populates the destination when the driver sits down to either work or home or a calendar event. If there's no route on the screen then it will just drive following the roadway...

That's Arizona, not Los Angeles. I guess the HD maps (truth) said there was a road there.

V12 is mostly good but after several days of use my biggest pet peeve has become its indecision around turn lanes, especially dual turn lanes. Not every intersection causes this behavior but when it does it will oscillate back and forth between lanes sometimes 2-3-4 times, or just kind of...

Solving full autonomy gets you the ability to do lots of cool things, but that's many years away. The current models like V12 FSD don't have any agency, they're purely predictive. To take humans totally out of the loop you need models that can plan/conspire, like MuZero, not simply predict the...

I hope we get some new visualizations that make more sense with V12 since the new e2e model doesn't actually make use of what's visualized on the screen anymore (cars, road edges, etc), that's all legacy. It would be cool to train a model to provide "future sight", i.e. like the backup camera...

I'm predicting that the Tesla Robotaxi will be battery-powered.

I don't know how they input navigation directions into the end-the-end model, there are different ways you can do this. But one day they'll probably need some multi-modal thing that can handle language tokens like "Turn left on 10th street", and that will make it possible to grok the roadway...

I've always wondered if Tesla quietly uses safety score or similar to find the "good driver" cohorts. The other possibility is that good drivers stand out in the training set because good drivers all do very similar things, while bad drivers are bad in a wide variety of ways. The good drivers...

I realized a funny sort of switch happened with the V12 release where previously I'd only use FSD on the highway because the city driving was far too janky and unreliable on V11. But now highway is the crappier experience between the two, and I look forward to seeing "Auto" re-appear in the max...

If Tesla can create a perfect driving agent that works in every possible situation worldwide and it only breaks down because the camera got covered up I am fine declaring it "mission accomplished". Full autonomy is 100% a software problem. FSD (supervised) is a cool intermediary on the way to...