“Crawl Control” for the Cybertruck?

[rr6013]

When talking about FSD, I'm hearing these words more and more :

Magic, Deciding

Any sufficiently advanced technology is indistinguishable from magic.”
― Arthur C. Clarke
FSD is blackbox AI. Similar to LaMDA there is no analyzing the technology and its decisions. Only behavior, after the fact, is very similar, same as magic.

I have a personal trigger with deciding since the human always makes the first choice to use FSD in the frst place. I think that decision may “blanket” responsibility and liability for FSD behaviot thereafter(i.e. deciding NOT to turn off FSD) is human responsibility. So its a funnel for all responsibility for FSD in-use

Sponsored

 

[charliemagpie]

Your FSD / Robot will be one, in whatever guise, one day be with you for life.

 

[slomobile]

Any sufficiently advanced technology is indistinguishable from magic.”
― Arthur C. Clarke
FSD is blackbox AI. Similar to LaMDA there is no analyzing the technology and its decisions. Only behavior, after the fact, is very similar, same as magic.

I have a personal trigger with deciding since the human always makes the first choice to use FSD in the frst place. I think that decision may “blanket” responsibility and liability for FSD behaviot thereafter(i.e. deciding NOT to turn off FSD) is human responsibility. So its a funnel for all responsibility for FSD in-use

https://jalopnik.com/self-driving-trucks-could-be-on-texas-highways-by-2023-1849089082

If I share the road with an autonomous truck that someone else put there, is it still my decision because I knowingly chose to manually drive where they do so autonomously? Do I really have a choice other than giving up my own right to travel, or fighting to ban AI drivers if I do not trust AI?

AI, as advanced as it is, is built on deterministic systems, built from analog components. Analysis of the AI decision making process is possible, despite claims to the contrary, with instrumented AI linked to training data. Unfortunately, analysis is often meaningless because it has no human correlate for us to identify with, or there is no easy fix to implement.

A is weighted higher than B, so the behavior associated with A was chosen.
For that to be meaningful you need to know why A is weighted higher.
To understand that, you need to go back to the training data and note which data records caused a shift in the weightings of A and B.
You try to identify common elements in those records that are absent in records which fail to shift A and B weightings. If those elements are human recognizable, the discovery seems meaningful. Like the AI which identified sport fish in photos by the frequent presence of human thumbs holding the fish.
We can teach the AI by adding several photos of identified sport fish to the training dataset which do not have human fingers. Now, hopefully the AI will adapt by using some other method in addition to the proven thumb method to identify sport fish.

We can duplicate the AI behavior by copying the weighted neuron structure. We can affect the formation of behavior patterns by altering the training. We can directly manipulate neuron structure, but the results are largely unpredictable.

Now, reread the last 2 paragraphs substituting a person for the AI. Parents and life experiences for training data, and a psychoanalyst in place of computer scientist. The roles are similar.

We can often predict that a human will behave a certain way because they were raised a certain way.
AI is a black box only as much as humans are.
But we are unable to clone or directly alter human brains. We can do that with AI, so we have the opportunity for more responsive 'tuning' with a bit more research.

 

[Crissa]

Humans make insane decisions all the time. I have video of a car doing a crazy ivan u-turn across four lanes, cutting me off, on a bridge, just last week.

You can watch any number of videos of Humans making illogical choices of where their vehicle should be, how it should stay in their lane, or stop.

There being an AI in the driver's seat makes no difference in these cases. You just have to trust they don't do something stupid. An AI is going to be less likely than a Human to pull out in front of a motorcycle, period.

-Crissa

 

[rr6013]

If I share the road with an autonomous truck that someone else put there, is it still my decision because I knowingly chose to manually drive where they do so autonomously? Do I really have a choice other than giving up my own right to travel, or fighting to ban AI drivers if I do not trust AI?

@charliemagpie sees Magic deciding. I agreed the state of the art AI magic is humanly imperceptible. Its called technology - blackbox artificial intelligence. And the decisions are inherently imperceptible to humans.

You don’t see a line when AI has decided to cross one, as we humans see in a car on the speedometer cross a speed limit or motor cross the redline. Tesla doesn’t provide any redlines for people to see yet.

A:B analysis presented analytically remains non-realtime non-obvious as a human redline device. This is reflected in the real world. In the case Autopilot was driven on freeway offramp, blowing through a stop signed intersection causing accident fatality, IIRC. Sure stupid human trick AP is stipulated not to be used exiting freeway. BUT we don’t know the accident report. We don’t know if AP decided the offramp was “freeway”. And the blown intersection might be the tell that something went wrong in the process(ing). Its fucked up. It’s fucked up. And its Driver fucked up. Did Tesla?

SO to answer your question”Do I really have a choice?” – YES. This is what is so triggering. Without redlines that humans can use to decide whether all AI operations are “nominal” the only choice you have is to reject AI. Until Tesla AI is sufficiently advanced to be able to arm the car with necessary bells, whistles and redline cockpit warning systems, you must stay grounded, NO AI.

Driving is a privilege, not a right. AI enjoys privileged use of the open roads during its pilot phase of development. You enjoy the same privilege by license to share those same roads.

Philosophically you really have five choices.

1. Unilaterally stop driving open roads that AI fiat-share
2. Drive fiat-shared open roads with AI
3. Lobby OEM’s to provide its vehicles with sufficient bells, whistles and redlines to enable you to determine whether AI, AP, EAP et.al., are operating nominally.
4. Protest against the fiat privilege; against AI driving without a FEDERAL license
5. Protest against all human driver FEDERAL liability for all AI.

First Principles? Who should is the decider? That is determined by who the Law holds responsible. Ask the Tesla driver who killed people when his Auotopilot drove off the freeway.

But for the grace of God there go us all.

 

[Ogre]

we are not quite to the point where FSD is better than human drivers, but it is coming fast. Within 5 years I believe FSD will be vastly safer than the typical human driver.

Right now if an automated vehicle is involved in a crash it is assumed it is at fault. In 10 years if FSD is involved in a crash with a human, people will assume the human is at fault and many will call them irresponsible.

I look forward to safer roads ahead

 

[Sirfun]

we are not quite to the point where FSD is better than human drivers, but it is coming fast. Within 5 years I believe FSD will be vastly safer than the typical human driver.

Right now if an automated vehicle is involved in a crash it is assumed it is at fault. In 10 years if FSD is involved in a crash with a human, people will assume the human is at fault and many will call them irresponsible.

I look forward to safer roads ahead

You give human drivers too much credit. There are lots of human drivers that are awful. From the videos I watch on YouTube, I think the AVERAGE human driver is better than FSD currently.

It would be interesting to see a video of an FSD driven car taking a Driver's License test.

 

[Ogre]

You give human drivers too much credit. There are lots of human drivers that are awful. From the videos I watch on YouTube, I think the AVERAGE human driver is better than FSD currently.

It would be interesting to see a video of an FSD driven car taking a Driver's License test.

Oh not at all.

With you on this. But public perception will lag reality for some time here.

 

[slomobile]

Until Tesla AI is sufficiently advanced to be able to arm the car with necessary bells, whistles and redline cockpit warning systems,

I think Tesla might say the removal of the redlines indicates they are more advanced. Rather than display a warning, the car just takes care of that for you.

AI makes the decisions about what action to do next, then calls a subroutine to perform the action. It would be trivial for the first action of these subroutines to display a prompt asking the user if they would like to proceed with a named action such as "Proceed through yellow light intersection? Y/N".

But can you imagine driving like that? Your attention would be constantly on the screen. How does that make anyone a better driver? I'm sure initial development went something like that, where the check driver had to approve each little action for safety. Then, as the car proved itself reliable for that procedure, it was removed from the check list.

There are a few things remaining that are only 99.999% reliable where the standard is 99.999999% six sigma. ( Not sure I used that right. Engineering school dropout, bad at math. ) Freeway exits are likely one of those things, which is why they advise against taking them on autopilot. Because the system cannot detect that situation, it cannot issue a warning. If it could issue a warning, it wouldn't need to. It would just handle the exit.

I'm curious what is on your list of "necessary bells, whistles and redline cockpit warning systems". What specific criteria would an autonomous vehicle need to meet for you to trust it Full Self Driving?

That is the big question OEMs need answered. What will the public accept? In addition to lots of research and surveys, they sell vehicles with varying levels of functionality. What sells and what doesn't is at least as good as any survey. Maybe they should spend more time here.

 

[ajdelange]

AI, as advanced as it is, is built on deterministic systems, built from analog components.

Yes the logic is ultimately implemented in analogue circuitry but the SNR is so high in these circuits that the probability of bit error is vanishingly small and parity check bits render it effectively 0.

We can duplicate the AI behavior by copying the weighted neuron structure. We can affect the formation of behavior patterns by altering the training. We can directly manipulate neuron structure, but the results are largely unpredictable.

Once the weights have been fixed the transfer function is completely deterministic.

 

[Crissa]

In the case Autopilot was driven on freeway offramp, blowing through a stop signed intersection causing accident fatality, IIRC. Sure stupid human trick AP is stipulated not to be used exiting freeway. BUT we don’t know the accident report. We don’t know if AP decided the offramp was “freeway”. And the blown intersection might be the tell that something went wrong in the process(ing). Its fucked up. It’s fucked up. And its Driver fucked up. Did Tesla?

At no point up to that accident had Tesla activated features which would or could recognize an intersection. The tool the driver was using was plain Autopilot. It had warnings that it could not stop in time from highway speeds for stopped vehicles - in fact, no Active Driver Assist Systems from other manufacturers could. All ADAS had that warning up until last year.

The collision played out the same as if the driver had dumb cruise control - it did not stop, it remained at speed.

Nothing to do with AI. It was a user having no respect for the world around them.

-Crissa

 

[SparkChaser]

I would argue that FDS is already safer than humans. The requirement is that it be multiple time smarter.

 

[charliemagpie]

It just occurred to me that current Beta drivers are above average drivers.

Seems lately I've grown impatient with drivers who linger too long for gaps at stop signs. Or I've just had a bad run. Or it's this new area I've just moved into.

Gooooo...OMG..... What are you waiting for !!!!

On reflection, these slow FSD takeoffs on Youtube, are maybe much closer to average driving than I thought.

 

[Sirfun]

It just occurred to me that current Beta drivers are above average drivers.

Seems lately I've grown impatient with drivers who linger too long for gaps at stop signs. Or I've just had a bad run. Or it's this new area I've just moved into.

Gooooo...OMG..... What are you waiting for !!!!

On reflection, these slow FSD takeoffs on Youtube, are maybe much closer to average driving than I thought.

Out on the golf course, my son and I used to have this as our theme song. With age, I'm mellowing a bit finally.

 

[rr6013]

The collision played out the same as if the driver had dumb cruise control - it did not stop, it remained at speed.

Did it. That’s the narrative released for media reporting.

Or did AP do the steering? That’s a material difference. One that separates the common cruise control narrative people are familiar with, from the AP paradigm. AP I’m not even familiar, much less expert. Cruise control all of us are expert.

Even if the paradigms are equivalent, the unanswered logical is “How can one human be so stupid?” He effectively drove the car off the freeway at full speed; down the off ramp and into an intersection w/o pumping the brakes! That’s what the media expects me to accept.

It simply is unbelievable. But I understand one side of the narrative expertly. Only LEXUS WOT(wide open throttle) accident similarly reflects this accident. That instance researchers proved that the quality of materials in the throttle contol mechanism were defective. Resultant wire strands provided an electrical pathway that short circuited Cruise control, ignition and turning either or both the car and Cruise control off. Tgat killed a CHP officer his wife, daughter and friend.

Sponsored