Public safety idea for Tesla:
Use the Tesla fleet to help detect missing persons / Amber Alerts using onboard cameras + compute.
A lightweight model could flag possible matches locally, or Tesla could “mine” the fleet for candidate sightings similar to how edge-case driving scenarios are collected.
With millions of vehicles on the road, this could easily become one of the largest real-time public safety networks ever built.
@Tesla_AI@elonmusk@aelluswamy@Tesla@SawyerMerritt@wholemars@DirtyTesLa
Feels super unlikely. Almost all of the camera data they’ve collected for years is missing a front bumper view. Unless they’ve been able to collect data dramatically more quickly with their new cars, I don’t see how the bumper view advantage would justify the loss of edge case data for training.
@DirtyTesLa@AIDRIVR visiting Austin for the rest of the week. Any advice on attempting to get a test ride in Robotaxi? I assume it’s unlikely, but figured I’d ask!
I completely agree, but in addition to that, I think it’s reasonable to assume (even with simultaneous releases) that Y/3 will see at LEAST slightly better performance for the foreseeable future on the same FSD branch. It’s impossible to discount the model performance advantage that Y/3 have with their data being the overwhelming source for training these models.
Now’s your chance to participate in the Dune: Awakening Closed Beta!
Follow us, like this post, and retweet to be entered in for the drawing. Winners will be chosen and contacted next week. Make sure your DMs are open so we can contact you!
Complete the Beta Player Survey as well as see more information on our forum post here: https://t.co/Z4Hs1Y2pr4
Bless the Maker and good luck.
I catch your point (to a degree), but it seems to me that the 12.5 model must still be running in emulation mode on HW4 (AI4). My reasoning:
A non-emulation mode model implies one fully trained on a HW4 input stream, which would necessarily have a different shape than HW3 could support (just more pixels). They also haven’t powered on the new data center that’s intended to train that model.
The more likely explanation, from my perspective, is the optimizations they’re referencing have always been a part of the deployment process, but require some additional passes in this case because of the larger parameter space.
HW4 works in emulation mode, with less optimization, just because it has so much compute head room. Emulation of an inference task wouldn’t necessarily constrain the hardware.
I read this as still *lootable* on characters 71+, but only *usable* by <=70.
Feels like there must be an implementation problem that they’re bandaiding with this change. Meaning it was either do this or rewrite something that conflicted with TWW loot table (which may have caused a delay or removal).
Assuming that’s the case, this is fine I guess. I’d rather have it than not.
Make his finest asset yours
Follow & Repost with #XboxCheekyControllerSweepstakes for a chance to win a custom Xbox Series X console & 2 Cheeky Controllers designed by Deadpool.
Ages 18+. Ends 8/11/24. Rules: https://t.co/MI0oKWl1Xf
@Marvel Studios' Deadpool and Wolverine, in theaters July 26
Make his finest asset yours
Follow & Repost with #XboxCheekyControllerSweepstakes for a chance to win a custom Xbox Series X console & 2 Cheeky Controllers designed by Deadpool.
Ages 18+. Ends 8/11/24. Rules: https://t.co/MI0oKWl1Xf
@Marvel Studios' Deadpool and Wolverine, in theaters July 26
Wishing Donald Trump good health after an assassination attempt is not a partisan statement, it's a human statement.
I'm getting really tired of people telling me where I stand politically. I'm not right-wing or left-wing. As I've said many times: I'm not partisan. I make up my mind one issue at a time. And I try to approach everyone & every idea with empathy and an open-mind.
I'm sorry if some of you want to categorize people into blue & red bins, probably because it's easier to hate than to understand.
Still, I love you all ♥