Waymo

[spacecoin]

It's the rate of advancement, the absolutely most critical point, that you left out.

6 months ago, FSDb drove like a 6 yo. Now, arguably a 16 yo after a few lesions. It's the rate of advancement that is crucial, and as the training of the neural nets advances, that will accelerate at a far greater pace than other solutions can keep up with.

Tesla is still learning to follow traffic rules under optimal conditions. The reliability march hasn't even begun for Tesla. Waymo (and many others) drove at current FSDb-levels 6-7 years ago.

Even Musk has said each order of magnitude is exponentially harder to achieve. It took Tesla two years from getting from 30% to 90%. Now they need to get to at least 99.9999%. Do the math, it will take a lot more than 2 years.

 

[spacecoin]

Btw, Waymo killed a dog in SF the other day. Waymo Kills Small Dog on San Francisco Street, DMV Report Says

Safety driver in car, car in autonomous mode.

 

[bkp_duke]

Waymo (and many others) drove at current FSDb-levels 6-7 years ago.

 

[spacecoin]

View attachment 944467

You really have no clue, do you? Elaine Herzberg was over five years ago. Uber ATG drove at least as well as FSDb in 2018 and Waymo was ahead of them.

Edit: Uber launched with safety drivers in Pittsburgh in 2016 and in Phoenix in 2017 IIRC. Here's a fun video for you:

 

[ItsNotAboutTheMoney]

Btw, Waymo killed a dog in SF the other day. Waymo Kills Small Dog on San Francisco Street, DMV Report Says

Safety driver in car, car in autonomous mode.

They'll not be publishing that one, no matter how clearly it shows the car did nothing wrong.

 

[bkp_duke]

You really have no clue, do you? Elaine Herzberg was over five years ago. Uber ATG drove at least as well as FSDb in 2018 and Waymo was ahead of them.

Edit: Uber launched with safety drivers in Pittsburgh in 2016 and in Phoenix in 2017 IIRC. Here's a fun video for you:

Again, I have FSDb and use it daily, and travel enough to have recently used Waymo in SF (fortunately the one I was in didn't hit any animals). You live across the pond where you can't get either, and based your posts upon YT videos, not first-hand experience.

Whatever you say bro, whatever you say.

 

[Bitdepth]

Btw, Waymo killed a dog in SF the other day. Waymo Kills Small Dog on San Francisco Street, DMV Report Says

Safety driver in car, car in autonomous mode.

That's terrible.

 

[spacecoin]

You live across the pond where you can't get either, and based your posts upon YT videos, not first-hand experience.

This is not a matter of opinion. These are facts. Watch the video and do some basic research or stay in your happy bubble.

Oh and btw, I have ridden in a Waymo and FSDb. Last time in FSDb was last month.

Here's another oldie from 7 years ago - but must be another deep fake...

 

[spacecoin]

That's terrible.

Yeah all deaths are. I doubt a human could have avoided an unleashed dog darting out between parked cars... The safety driver didn't see the dog, but the car apparently did, but it didn't have time to break enough.

 

[spacecoin]

Whatever you say bro, whatever you say.

Here's another datapoint for you:

The End of Starsky Robotics

In 2015, I got obsessed with the idea of driverless trucks and started Starsky Robotics. In 2016, we became the first street-legal vehicle…

medium.com

This is from three years ago. Read the text, look at the graphs.

"The S-Curve here is why Comma.ai, with 5–15 engineers, sees performance not wholly different than Tesla’s 100+ person autonomy team."

Your best counter argument so far is "you're from Europe"...

 

[bkp_duke]

Here's another datapoint for you:

The End of Starsky Robotics

In 2015, I got obsessed with the idea of driverless trucks and started Starsky Robotics. In 2016, we became the first street-legal vehicle…

medium.com

This is from three years ago. Read the text, look at the graphs.

"The S-Curve here is why Comma.ai, with 5–15 engineers, sees performance not wholly different than Tesla’s 100+ person autonomy team."
View attachment 944499

Your best counter argument so far is "you're from Europe"...

No, the best counter argument is evidence. The evidence that I drive daily with FSDb (last updated this past weekend - 11.4.2 is pretty darned good), and that I drive frequently to SF and have very recently ridden in a Waymo.

You seem to have some emotional argument you want to hold on to and a need to "win". I don't care which company hits automony soonest. What I observe is that Waymo's version will run into monumental scaling problems (the storage requirements for HD maps alone for large swaths of the USA would likely be a problem - hundreds of TBs to PBs of data). While FSDb is far from perfect, I need marked progress with each release, each month, and I am able to use it everywhere I drive (i.e. the practical "value" of it is exponentially higher to most users).

 

[spacecoin]

No, the best counter argument is evidence. The evidence that I drive daily with FSDb (last updated this past weekend - 11.4.2 is pretty darned good), and that I drive frequently to SF and have very recently ridden in a Waymo.

You seem to have some emotional argument you want to hold on to and a need to "win". I don't care which company hits automony soonest. What I observe is that Waymo's version will run into monumental scaling problems (the storage requirements for HD maps alone for large swaths of the USA would likely be a problem - hundreds of TBs to PBs of data). While FSDb is far from perfect, I need marked progress with each release, each month, and I am able to use it everywhere I drive (i.e. the practical "value" of it is exponentially higher to most users).

I think, as a tech person, that Tesla does some impressive things with FSDb, and I am glad that it's value of to you.

However, this interaction started when you mocked me for pointing out that the reliability that Tesla has now, others had 6-7 years ago. This should not be a controversial statement as it's a fact. I have linked to a few sources.

The development rate is fast in the beginning and gets slower and more costly to get those nines up. It should not be a controversial statement either. This is true for all ML systems.

 

[bkp_duke]

While I think as a tech person that Tesla does some impressive things with FSDb, this interaction started when you mocked me for pointing out that the reliability that Tesla has now, others had 6-7 years ago, This should not be a controversial statement as it's a fact. I have linked to a few sources.

The development rate is fast in the beginning and gets slower and more costly to get those nines. It should not be a controversial statement either. This is true for all ML systems.

I took that as a 100% false statement. 6-7 years ago, even looking at YT videos and not going on my personal experience of riding in Waymo's "way back when" they were NOT reliable. Even back in 2021 there is a relatively famous video of them "piling up" and causing a traffic jam. They were also limited very much in functionality (6-7 years ago). Unprotected left turn (a reality for FSDb today - you set this benchmark), not happening in a Waymo.

 

[spacecoin]

I took that as a 100% false statement. 6-7 years ago, even looking at YT videos and not going on my personal experience of riding in Waymo's "way back when" they were NOT reliable. Even back in 2021 there is a relatively famous video of them "piling up" and causing a traffic jam. They were also limited very much in functionality (6-7 years ago). Unprotected left turn (a reality for FSDb today - you set this benchmark), not happening in a Waymo.

Tesla has indeed some impressive functionality (like UPL), but they also only have about a 95% of the drives without critical disengagements. There is a disengagement on average every 15 miles, and that's not counting speed adjustments and accelerator taps. Source: fsd community tracker.

I'm very confident that Waymo had that level of reliability or better 6-7 years ago, and the DMV stats back that up.

Anecdotes are not really interesting to me, but as a norm people get impressed with FSDb not needing input for an entire single drive, while a system like Waymo's is driverless and makes headlines when it acts up, not when it completes a drive.

 

[bkp_duke]

Tesla has indeed some impressive functionality (like UPL), but they also only have about a 95% of the drives without critical disengagements. There is a disengagement every 15 miles, and that's not counting speed adjustments and accelerator taps. Source: fsd community tracker.

I'm very confident that Waymo had that level of reliability 6-7 years ago, and the DMV stats back that up.

Anecdotes are not really interesting to me, but as a norm people get impressed with FSDb not needing input for an entire single drive, and a system like Waymo's is driverless and makes headlines when it acts up, now when it completes a drive.

Um, no. So, first the DMV stats are "self reported". There is no consistent requirements on how and what is reported there. It's been critiqued to the max in this and other forums (i.e. it's a really REALLY bad dataset to draw conclusions from except on how each individual system progresses from year to year).

Second, Waymo gets that "level of functionality" as you call it by avoiding as many difficult situations as it can. Highway driving? Nope, not a shot in hell. Heavy traffic? It's programmed to be overly cautious there as well. Left turns? It often is reported (on YT) that it will take multiple rights in order to avoid traffic patterns with heavy lefts in it, or even multiple lane changes.

It's an N=1, but I had a 90 mile (2 x 45 mile) round trip from SD to Orange County this weekend. Parts were heavy traffic, parts were on and off ramps, parts were lane changes (both freeway and surface streets). Those were intervention-free drives (although I wish it would have driven more aggressively - but that's the Mario Andretti coming out in me).

Those were drives that Waymo simply could not do. Why? It's not available in my area. The "practical value" of Waymo and Cruise for me are near zero (about the same as their geographic coverage).

 

[spacecoin]

So, first the DMV stats are "self reported". There is no consistent requirements on how and what is reported there. It's been critiqued to the max in this and other forums (i.e. it's a really REALLY bad dataset to draw conclusions from except on how each individual system progresses from year to year).

I agree with this statement re DMV DE-stats. And I am still completely confident that Waymo had a better DE rate than Tesla six years ago. To be crystal clear - the statement is likely correct, and I think it doesn't warrant the response you made tbh.

Second, Waymo gets that "level of functionality" as you call it by avoiding as many difficult situations as it can. Highway driving? Nope, not a shot in hell. Heavy traffic? It's programmed to be overly cautious there as well. Left turns? It often is reported (on YT) that it will take multiple rights in order to avoid traffic patterns with heavy lefts in it, or even multiple lane changes.

You surely know that Waymo is doing UPL:s pretty frequently these days, and highway will likely be added this year in the service areas. They've been doing highways with a safety driver for a long time both with the Gen 5 and with the trucks. I believe they even did highway back in 2014 when Googlers tested the car internally.

It's an N=1, but I had a 90 mile (2 x 45 mile) round trip from SD to Orange County this weekend. Parts were heavy traffic, parts were on and off ramps, parts were lane changes (both freeway and surface streets). Those were intervention-free drives (although I wish it would have driven more aggressively - but that's the Mario Andretti coming out in me).

Those were drives that Waymo simply could not do. Why? It's not available in my area. The "practical value" of Waymo and Cruise for me are near zero (about the same as their geographic coverage).

You're widening the discussion to the ODD. And that's ok, again Tesla does some impressive stuff! But I'm willing to take a bet 2:1 on any amount that Tesla won't take on liability for driving in any ODD by the end of 2024 (on current hardware).

 

[bkp_duke]

I don't expect Tesla to take liability, so I won't take that bet. And I say that knowing one of Tesla's former head attorneys. They will go all the way to L5 functionality while within the L2 legal framework (of requiring hands on) until they get through the march of the 9s. It's pretty much well-scripted out that this is what they are going to do. And I'm perfectly fine with that because the system still brings me VALUE as a user. And I see that value increasing with each version that gets installed.

In a nutshell - I see Waymo, Cruise, and others taking VERY cautious approaches to their self-driving solutions. And that's fine, that's their choice, there is arguable merit to that choice.

But there is also significant merit to Tesla's approach. They are letting their system do harder things and letting it fail, in order to learn from the experience. Legally they are able to get away with this because it's a L2 driver assist, not a fully autonomous system. There are merits to letting the system be allowed to push the edges of functionality and learn from those failures. Arguably, this is why Tesla broke off their relationship with MobileEye (the HW provider for AP1).

 

[spacecoin]

I don't expect Tesla to take liability, so I won't take that bet. And I say that knowing one of Tesla's former head attorneys. They will go all the way to L5 functionality while within the L2 legal framework (of requiring hands on) until they get through the march of the 9s.

Personally I think know too much about ML and CV at this point to have any hopes for autonomy (in a relevant ODD) on current hardware. The idea that you can get to L5 without going autonomous in a smaller ODD is absurd to me. L2 "hey let's try it and see if it works", sure.

It's pretty much well-scripted out that this is what they are going to do. And I'm perfectly fine with that because the system still brings me VALUE as a user. And I see that value increasing with each version that gets installed.

I think there will be some vendors aiming at highway L3 at highways speeds. I'd get that system over Tesla's "L2 everywhere" every day of the week when/if I have that choice. The killer app is eyes-off for most people, I think. An L3 system can free up time, so people can do other things on those long drives (and not have to baby sit the system). I personally do not see any value of using an L2 in the city and in suburbs. I will always drive better and more assertive myself.

 

[Supcom]

What I observe is that Waymo's version will run into monumental scaling problems (the storage requirements for HD maps alone for large swaths of the USA would likely be a problem - hundreds of TBs to PBs of data).

This might be a problem if every Waymo was expected to be driving anywhere at any time. But, the reality is that nobody will be taking robotaxis long distance anytime soon, so they only need the maps for their current service area. And if, on the remote chance, that someone with more money trhan brains wants to do so, I guess that the Waymo may download new maps when it stops at the Waymo service centers to refuel/charge along the way.

Alternatively, if a Waymo were to plan a long drive out of it's service area, it could potentially download HD maps that cover the path it plans to drive.

In any event, no need to download the location of every sewer grate from coast to coast.

 

[spacecoin]

This might be a problem if every Waymo was expected to be driving anywhere at any time. But, the reality is that nobody will be taking robotaxis long distance anytime soon, so they only need the maps for their current service area.

Yeah. Another fact is that over 95% of the taxi revenue world-wide (which is what Waymo is going after) is from drives within large urban centers (including airports). If you need to go somewhere else, get a taxi with a driver? There is no business case going after that 5% from an operations and tech perspective.