4 ms·

>There’s even a solid case to be made that completely autonomous vehicles (AVs) will never take over everyday travel. Have these researchers seen Tesla's lates

>There’s even a solid case to be made that completely autonomous vehicles (AVs) will never take over everyday travel. Have these researchers seen Tesla's latest FSD version 12.3, which uses neural nets instead of hard-coded instructions? For example this video: https://twitter.com/WholeMarsBlog/status/1768714246521827643 https://twitter.com/WholeMarsBlog/status/1768714246521827643 "Tesla Full Self-Driving Beta 12.3 Drives 85 Minutes from Palo Alto to San Francisco during Rush Hour with Zero Interventions". I was also a skeptic until recently.

r3d0c 3y ago
sf is the most heavily mapped & learned area by tech companies in the whole world
- xyzzy4747 3y ago
  They have driving data from all around the country which is collected from anyone driving a Tesla. Waymo is much more limited to specific cities.
- ghaff 3y ago
  I was taking a friend home from the big medical complex in Boston home up 93. I haven't driven much in the city of Boston for ages and mostly not that route. I wish I had video. I had really forgotten just how horrifying driving in Boston can be at a busy time of day. Constant merges that you sort of need to know about. Often aggressive drivers. Apple Maps tried hard but I was glad I had my passenger telling me when to get left, when to get right, follow that car, etc. SF isn't easy but Boston is next level. And it snows there. And can get heavy rain.
recursive 3y ago
For years, the breakthrough has always been in the latest beta software I just hadn't seen yet. Spoiler: It wasn't. One of these times, it will be, but I'm going to bet against it and be right most of the time.
- xyzzy4747 3y ago
  They recently switched it to a neural net model instead of the old approach that had a ton of hard-coded C++ logic. I agree it's hard to time when it's good enough for robotaxis and such, but good progress is being made.
Animats 3y ago
No, not "zero interventions. The human driver puts their hands on the wheel at 08:36.[1] At 12:56, the vehicle drives through a crosswalk with pedestrians present, but misses them.[2] A similar situation appears near the very end. This system drives through crosswalks rather aggressively. There are no hard cases. No heavy traffic. No construction. No emergency vehicles. No double-parked cars. Almost always, either there's plenty of open space ahead or another car to follow. When heavy traffic is encountered, the human driver changes the route. This looks like a cherry-picked try from many trial runs. It's way below Waymo level performance. Tesla has a history of fake self-driving videos [3], so this has to be looked at suspiciously. [1] https://youtu.be/PXy90bAPwn8?t=516 https://youtu.be/PXy90bAPwn8?t=516 [2] https://youtu.be/PXy90bAPwn8?t=776 https://youtu.be/PXy90bAPwn8?t=776 [3] https://www.cbsnews.com/news/tesla-autopilot-staged-engineer-says-company-faked-full-autopilot/ https://www.cbsnews.com/news/tesla-autopilot-staged-engineer...
tippytippytango 3y ago
It's too easy to create an impressive demo right now, yet be very far from solving the problem. Humans set the bar on driving ability incredibly high. Humans achieve 500k miles between collisions on average. The tesla FSD tracker https://www.teslafsdtracker.com https://www.teslafsdtracker.com has been showing that most versions of FSD are well under 200 miles between critical intervention with one version achieving up to 300 but dropping back down in the next release. We don't have data on 12 yet, but it seems exceedingly unlikely they've been able to close a 3 order of magnitude gap in one release to match human ability. I say this as an owner of FSD. It has moments where it does amazing things and moments that make me shake my head. I like it for highways, but driving in a city is too stressful to be more than a toy right now.

r3d0c 3y ago

sf is the most heavily mapped & learned area by tech companies in the whole world

xyzzy4747 3y ago
They have driving data from all around the country which is collected from anyone driving a Tesla. Waymo is much more limited to specific cities.
ghaff 3y ago
I was taking a friend home from the big medical complex in Boston home up 93. I haven't driven much in the city of Boston for ages and mostly not that route. I wish I had video. I had really forgotten just how horrifying driving in Boston can be at a busy time of day. Constant merges that you sort of need to know about. Often aggressive drivers. Apple Maps tried hard but I was glad I had my passenger telling me when to get left, when to get right, follow that car, etc. SF isn't easy but Boston is next level. And it snows there. And can get heavy rain.

recursive 3y ago

For years, the breakthrough has always been in the latest beta software I just hadn't seen yet. Spoiler: It wasn't. One of these times, it will be, but I'm going to bet against it and be right most of the time.

xyzzy4747 3y ago
They recently switched it to a neural net model instead of the old approach that had a ton of hard-coded C++ logic. I agree it's hard to time when it's good enough for robotaxis and such, but good progress is being made.

Animats 3y ago

No, not "zero interventions. The human driver puts their hands on the wheel at 08:36.[1] At 12:56, the vehicle drives through a crosswalk with pedestrians present, but misses them.[2] A similar situation appears near the very end. This system drives through crosswalks rather aggressively. There are no hard cases. No heavy traffic. No construction. No emergency vehicles. No double-parked cars. Almost always, either there's plenty of open space ahead or another car to follow. When heavy traffic is encountered, the human driver changes the route. This looks like a cherry-picked try from many trial runs. It's way below Waymo level performance. Tesla has a history of fake self-driving videos [3], so this has to be looked at suspiciously. [1] https://youtu.be/PXy90bAPwn8?t=516 https://youtu.be/PXy90bAPwn8?t=516 [2] https://youtu.be/PXy90bAPwn8?t=776 https://youtu.be/PXy90bAPwn8?t=776 [3] https://www.cbsnews.com/news/tesla-autopilot-staged-engineer-says-company-faked-full-autopilot/ https://www.cbsnews.com/news/tesla-autopilot-staged-engineer...

tippytippytango 3y ago

It's too easy to create an impressive demo right now, yet be very far from solving the problem. Humans set the bar on driving ability incredibly high. Humans achieve 500k miles between collisions on average. The tesla FSD tracker https://www.teslafsdtracker.com https://www.teslafsdtracker.com has been showing that most versions of FSD are well under 200 miles between critical intervention with one version achieving up to 300 but dropping back down in the next release. We don't have data on 12 yet, but it seems exceedingly unlikely they've been able to close a 3 order of magnitude gap in one release to match human ability. I say this as an owner of FSD. It has moments where it does amazing things and moments that make me shake my head. I like it for highways, but driving in a city is too stressful to be more than a toy right now.