Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[Steve m]

Versus your unmarked dirt track comparison, it is intuitively far easier to train a NN to drive in the centre of a nicely defined highway lane with good clear white lines and cats-eyes every 10 metres. But it's not obvious to me why you think it would be impossible to train a NN to recognise where the road edge begins and the grass verge starts. That seems a far easier thing to achieve than tarmacking and line-painting every obscure rural road in the world.

The bit I can see that is trickier, is for example the behaviour response when driving on a very narrow country road, when an oncoming agricultural vehicle cannot pass without you first reversing halfway off the road onto the grass bank. But as I said in my previous message, what percentage of journeys carry a greater than zero probability of this type of event?

Are we in New Zealand? In which case it is a near certainty.

 

[Carl Raymond]

Versus your unmarked dirt track comparison, it is intuitively far easier to train a NN to drive in the centre of a nicely defined highway lane with good clear white lines and cats-eyes every 10 metres. But it's not obvious to me why you think it would be impossible to train a NN to recognise where the road edge begins and the grass verge starts. That seems a far easier thing to achieve than tarmacking and line-painting every obscure rural road in the world.

The bit I can see that is trickier, is for example the behaviour response when driving on a very narrow country road, when an oncoming agricultural vehicle cannot pass without you first reversing halfway off the road onto the grass bank. But as I said in my previous message, what percentage of journeys carry a greater than zero probability of this type of event?

More common is the narrow street with parking both sides, where passing is only possible if one car pulls to the side near a driveway or other gap. I hope they are working on that one.

On an optimistic note, AI driven vehicles will immediately be better at reversing than all human drivers.

 

[Steve m]

More common is the narrow street with parking both sides, where passing is only possible if one car pulls to the side near a driveway or other gap. I hope they are working on that one.

On an optimistic note, AI driven vehicles will immediately be better at reversing than all human drivers.

And this sounds like England. I have had to pull in the side mirrors to squeeze between two rows of cars parked on each side of the road.

And driving on rural single track roads that have pullouts on each end of a blind curve to allow passing.

Any progress on AI driving that involves a large travel trailer and backing it around a sharp narrow bend with tree limbs overhead but uncomfortably low?

 

[SebastianR]

I think it is smart of Tesla not to assign SEA levels to their autopilot product descriptions. It allows them to shift features over with time.

So here is a thought of how they could extract even more revenue of the existing fleet later (much later): wherever the have regulatory approval they could add a new tier of AP “Unsupervised FSD” which will be good enough for picking up kids from school, getting you home drunk from the bar etc.

That level would be required for the Tesla Network to work (you could not verify if everyone using your car has a drivers license...)

On the spin-off idea: I sure don’t hope they go there: why have two securities out there that shorts can mess with? Oil and gas money is big enough - no need to enlarge the Wall St. exposure at all.

 

[Carl Raymond]

On the spin-off idea: I sure don’t hope they go there: why have two securities out there that shorts can mess with? Oil and gas money is big enough - no need to enlarge the Wall St. exposure at all.

If Musk likes spin-offs, Tesla could spin-off their solar division. They could call it ‘SolarCity’ or some such.

 

[Fact Checking]

Sometimes an established industry would fail miserably at predicting and standardizing future technology.
I am not saying the L1-L5 autonomous driving classification had failed, I am only trying to make the point that these classification and standards and regulations are not set in stone, and sometimes would need to be pushed and sometimes re-wrriten.

I agree, but note that while standardization of networking was important due to interoperability requirements between networking products, standardization of autonomy levels is almost 100% pointless, as the products do not and will not have to interoperate in any serious fashion.

FSD vehicles won't network with each other telling their autonomy levels. Any sort of interoperability (such as charging authorization, or future convoying and FSD tunnel access features) are or will be far more specific.

There's no technological need for FSD levels - it's almost entirely just marketing, increasingly used against Tesla, obscuring how much progress "L2 driver assist only" Tesla has made. (The recent "report" that showed Tesla almost dead last is a perfect example of how far detractors will go in downplaying Tesla's lead in autonomy.)

Tesla is correct in ignoring these "autonomy levels".

 

[Singer3000]

Are we in New Zealand? In which case it is a near certainty.

Ha! And yet you can also drive for hours on end in South Island without seeing another living soul, yet alone road user. Maybe it's a simple algorithm: IF (road vehicle identified) THEN (wake up driver)

 

[heltok]

I really like Teslas strategy not worrying about driver allowed to read newspaper on highway or not. Just make a system that gets better and better... At some point it will be good enough to be 2-10x safer than average driver and capable of driving 90%+ of all routes for 90%+ of all customers. At that point they trust that lawmakers will understand their statistics.

Seems like a much more straight forward approach than what I used to be tasked to do working at a competitor where we had pilot projects we demoed to public with specific routes and specific levels that never got to any major scale and really focused on reaching specific ASIL levels for every system to mathematically be able to argue that we would fulfill some specific design requirements.

 

[Antares Nebula]

Taking The Long View: Will Tesla's Q1 Disaster Turn Into Q4 Triumph? - Tesla, Inc. (NASDAQ:TSLA) | Seeking Alpha

 

[Fact Checking]

But what "hard" rules of the road do humans actually know about driving? I'd guess not much more than a few dozen major ones: Speed, lights, yielding, stops, meanings of lines, etc... Some have a few permutations: 2, 3, & 4 way stops, etc..

Along side of that are things that aren't rules, but dictated by physics: I can't occupy the same space as another object, deceleration distances, limits of traction, etc...

And the biggest one that humans are worst at, safety:

• to stay patiently attentive 100% of the time,
• in normal driving always keep a buffer of extra control authority - don't drive just below the edge of the physically possible,
• be aware of the relative speed of the vehicles behind you,
  
• in an emergency, don't panic and don't let panicking passengers affect you!

It's also amazing how many human drivers don't know how to recover a fishtailing vehicle:

More than ~half of all frontal collisions on rural roads comes from one of these factors: driver fell asleep or had eyes off the road, slippery road surfaces, or a fishtailing vehicle drifting sideways into the opposite lane. The fishtailing often starts off as a panicked overreaction, like in the video above.

(And to be fair it's not easy to recover a fishtailing vehicle that has no traction control, as the vehicle will quickly switch between being under-steered and over-steered, and the way to recover it is to be aware of the exact angle of the wheels at a given steering wheel position and resist instinctive corrections, which most drivers have no experience with.)

 

[Antares Nebula]

Tesla Model X steals the show at Jaden Smith’s Coachella performance

 

[KarenRei]

Versus your unmarked dirt track comparison, it is intuitively far easier to train a NN to drive in the centre of a nicely defined highway lane with good clear white lines and cats-eyes every 10 metres. But it's not obvious to me why you think it would be impossible to train a NN to recognise where the road edge begins and the grass verge starts. That seems a far easier thing to achieve than tarmacking and line-painting every obscure rural road in the world.

The bit I can see that is trickier, is for example the behaviour response when driving on a very narrow country road, when an oncoming agricultural vehicle cannot pass without you first reversing halfway off the road onto the grass bank. But as I said in my previous message, what percentage of journeys carry a greater than zero probability of this type of event?

IMHO, understanding the area surrounding the road, and the implications of driving off the road, are absolutely critical to making a self-driving system. There's lots of places around here where if you go off the road, you're going off a cliff to your death. Human drivers drive much more cautiously in such places, and AI drivers need to as well. The system needs to be able to calculate a "maximum width which will safely bear a vehicle" at the edge of the road, the penalties for being wrong, and adjust driving accordingly. The weighting factor for how bad the shoulder is should be multiplied by how bad the road and weather conditions are (aka, how likely you are to unintentionally go into the shoulder to begin with). The car should also be able to assess a negative-width shoulder - e.g. where obstruction or undercutting of the road surface has extended into the road itself.

On bad roads / bad weather around here where there's unsafe shoulders, it's standard for drivers to drive down the middle of the road (even where there's lines), unless there's oncoming traffic approaching. AP needs to learn to do this too.

The day I see, on the AP display, road shoulders accurately modeled, colour-coded by danger, and collision hazards on them (rocks, trees, human structures, etc) flagged by severity, and the current road's driving conditions accurately flagged, is the day that I'll think, "Driverless self-driving is probably ready."

Thankfully, I imagine that Tesla does care about this and is working on this, if only for wanting to have the ability to pull a car off the road if there's an emergency (or at present, if the driver is unresponsive). Shoulder assessment needs to be applied more broadly than that, but it'd be a critical first step.

 

[TheTalkingMule]

Computers have the ability to know where all cliffs reside.

 

[KarenRei]

Computers have the ability to know where all cliffs reside.

You'd think, but not really. Here's a topo map of the road to my land:

Mark the immediately-road-adjacent cliffs and unsafe shoulders by this 20-40 meter deep canyon.

Can't, can you? (the worst spot is where the road climbs up from Grafarmýri to the end of the driveway at Brekka, although the bridge crossing to the southwest and the area to the right immediately north of the bridge would also be a bad spot to go off... there are lots of other cliffs - for example, on the east side of the map, but they're too far off the road to matter)

Very few places in the world have the sort of terrain resolution to be able to accurately identify cliffs**, and I don't think anywhere in the world has the sort of topo resolution to properly measure shoulder widths (let alone assess their stability) above them.

 

[Steve m]

Computers have the ability to know where all cliffs reside.

Oh my having done a fair amount of back country navigating using contour maps one thing I learned was a map with 40 foot intervals will fail to show the 30 foot cliffs every time.

 

[KarenRei]

Oh my having done a fair amount of back country navigating using contour maps one thing I learned was a map with 40 foot intervals will fail to show the 30 foot cliffs every time.

And it's not just cliffs that are the problem. Last time I did a road trip I passed the still-flaming wreckage of an accident (which I later learned had killed a whole family) on perfectly flat land... surrounded by lava flows. If you hit aa lava at full speed... best of luck to you

Substitute "boulders" for people who live in other types of terrain. Or trees, or manmade obstacles, or whatever happens to be in your area.

 

[Pras]

I really like Teslas strategy not worrying about driver allowed to read newspaper on highway or not. Just make a system that gets better and better... At some point it will be good enough to be 2-10x safer than average driver and capable of driving 90%+ of all routes for 90%+ of all customers. At that point they trust that lawmakers will understand their statistics.

Seems like a much more straight forward approach than what I used to be tasked to do working at a competitor where we had pilot projects we demoed to public with specific routes and specific levels that never got to any major scale and really focused on reaching specific ASIL levels for every system to mathematically be able to argue that we would fulfill some specific design requirements.

This is also a good strategy when seeking regulatory approvals, getting them comfortable one by one.

 

[Fact Checking]

The day I see, on the AP display, road shoulders accurately modeled, colour-coded by danger, and collision hazards on them (rocks, trees, human structures, etc) flagged by severity, and the current road's driving conditions accurately flagged, is the day that I'll think, "Driverless self-driving is probably ready."

I expect them to concentrate on the low hanging fruits first, which is the other 95%+ of the road system and miles driven in the western world, which can be geo-marked 'safe' on a case by case basis: most large population centers and the highway systems connecting them.

Generic rural roads won't allow eyes-off-the-road driving for a long time I believe.

I.e. I expect them to treat "safe for eyes off the road FSD mode" differently from "safe with human supervision".

This Russian river crossing probably won't be marked as FSD-safe for a long time:

(I recommend watching this video to the end, there's a small surprise traffic scenario at around 1:30. )

 

[KarenRei]

I expect them to concentrate on the low hanging fruits first, which is the other 95%+ of the road system and miles driven in the western world, which can be geo-marked 'safe' on a case by case basis: most large population centers and the highway systems connecting them.

Generic rural roads won't allow eyes-off-the-road driving for a long time I believe.

I.e. I expect them to treat "safe for eyes off the road FSD mode" differently from "safe with human supervision".

This Russian river crossing probably won't be marked as FSD-safe for a long time:

(Watch it to the end, there's a small surprise traffic scenario in the video. )

Forget "bad Russian bridges in remote areas" - If your system doesn't understand to recognize roadside hazards, it won't even be trustworthy driving from the airport to our biggest tourist attractions.

 

[KarenRei]

This is also a good strategy when seeking regulatory approvals, getting them comfortable one by one.

Here's two problems with that. Let's say you have statistics saying that AP is 10 times safer than a human driver.

1. Except it's not. If you have a human behind the wheel, who can decide when to intervene, or not enable the system at all, then it's not AP's safety, it's the safety of a combined AP+human system.

2. Even if that wasn't the true, you have massive data biases in your system. For example, the largest concentration of Teslas in the US are by far in California, esp. cities like LA and SF. So let's say your data says that AP is 10 times safer than a human driver. It might be 15 times safer than a human in said cities, but 100 times more dangerous than a human in Podunk, Idaho. Is it okay to kill off the population of Podunk? If someone wants to drive to Podunk, does their car warn them that it sucks at driving near Podunk? Does it just geofence off Podunk - "Sorry, you can't go there"? Who decides what gets geofenced off? How do you think Podunkians would respond to reporters and legislators after hearing that this system that keeps killing people near their city just got approved to be driverless? Or contrarily to the news that they've been geofenced off, cutting traffic to their town? How do you think the media would play up the story - would they focus on the "15 times safer than a human in LA and SF" aspect, or would they focus on the "100 times deadlier than a human in Podunk" aspect? Do I even need to ask that? How do you think the fear factor about automated driving among the general public would be affected by a wave of stories with titles like "AP keeps driving people off cliffs in Podunk - is a road near you a deathtrap for AP as well?" How likely do you think parents would be to let their children be self-driven after reading such stories?

AP needs to be better than a human on public roads everywhere. General "X times safer than a human" stats include a number of biases - the biggest being the simple fact that you have a human deciding when they think it's safe to turn on and when they think it's not good enough.