FSD Beta Videos (and questions for FSD Beta drivers)

[EVNow]

So it's important to look at things for their own merits and be open to considering the fact that tools such as AP introduce new ways of dying that didn't exist before.

Absolutely.

Afterall "automobile" itself is an automation. Definitely people are dying in novel ways compared to riding in horse carriages. Horse carriages themselves introduced novel ways of dying compared to just walking.

ps : People make the mistake of comparing AP to completely manual driving. The correct comparison is to TACC, which is an evolution of cruise control. I think each of those evolutions were indeed upgrades to safety from CC to TACC to AP.

 

[nvx1977]

Properly monitored, automated systems greatly improve safety, no question about it.

But when people aren't properly trained in monitoring automation, things can get fatal really quickly.

I think the unknown here is whether AI can actually significantly surpass human-level decision making in its behavior.

Artificial Narrow Intelligence (ANI) has been shown to significantly exceed human capability in less complex tasks, like playing Chess, Go, Starcraft, etc. It stands to reason that given enough processing power, it can handle driving a car better than humans as well. What is less known is how much processing power is needed, and how long it will take to train such a NN. Dojo seems like an amazing attempt at tackling this problem, along with machine labeling, but these are not operational yet, afaik.

If Musk is right about an exponential learning curve, we could get there sooner than our intuition suggests. But given his failed past predictions, I tend to think we have years (but less than a decade) before we get to L4.

I am also not one to ever fully trust a machine, so I'm perfectly happy if FSD becomes a very proficient L2 system.

 

[EVNow]

If Musk is right about an exponential learning curve, we could get there sooner than our intuition suggests. But given his failed past predictions, I tend to think we have years (but less than a decade) before we get to L4.

Self-driving has always been 5 years away. For over a decade !

 

[Silicon Desert]

Exactly. In addition, the scoring thing is a POS. I’d like access too but I have to make an ongoing effort to qualify for what be for nothing. I have to drive at a crawl, hold up traffic, and then conversely push/run certain yellow lights. Get rid of the busted scoring system and the wait would be less frustrating. Clear messaging from Tesla on the program would also help. I get some decisions are dynamic, but there’s a lot of customer interest and they should really pay a little better attention to what they’re doing here.

Uh, I'm not buying this idea that you have to drive at a crawl. I am now at 3,000 mile with a 99 score. Speed has nothing to do with the score.
And I am not even trying to get a high score. Just driving the way I always do. And on interstate 80, it is fast.

 

[nvx1977]

Self-driving has always been 5 years away. For over a decade !

It's been 1 year away for 6 years, subtle difference. but we have reasonable explanations for that delay (the split from mobileye, the limitations of stateless video frames hitting local maximum, etc). The AI day presentation was sufficiently detailed to suggest that they are finally respecting the difficulty of the problem (Dojo specs are insane, temporal considerations are a must). But you're right. We might hit another local maximum and realize we need to tackle the problem in a different way, causing more delay. For now, I'm hopeful for Dojo/4d vectorspace. It seems like it's good enough for at least a high quality L2 system.

 

[Silicon Desert]

Wow, I just discovered I can engage FSD in my garage!

That's cool. I imagine you can now drive from your tool box to the deep freeze in case you can't remember how to get there.
Yes, I am just being silly, but in reality, I would have a hard enough time trusting this version of FSD on the streets, much less activate it in my garage.

 

[qdeathstar]

We have evolved over a billion years to handle unexpected events that might come up suddenly. Sometimes even once in a lifetime. We should be able to pickup something that is unusual and react.

Distracted driving is a problem - with or without FSD Beta.

ps : I'm more worried about highspeed NOA than low speed FSD Beta.

Most accidents happen on city streets though. Most fatal accidents happen on roads with a low speed limit. City streets are inherently more dangerous than interstates

https://crashstats.nhtsa.dot.gov/api/public/viewpublication/810625

 

[powertoold]

Most accidents happen on city streets though. Most fatal accidents happen on roads with a low speed limit. City streets are inherently more dangerous than interstates

https://crashstats.nhtsa.dot.gov/api/public/viewpublication/810625

Do those statistics differentiate between pedestrian deaths vs in-car deaths? Also, rural doesn't necessarily mean low speed. Many "rural" roads are 40-50mph with high speed cross traffic.

If anything, urban in this context means city streets, lol. I don't think urban means highway and rural means city streets in the document you posted.

 

[EVNow]

Most accidents happen on city streets though. Most fatal accidents happen on roads with a low speed limit. City streets are inherently more dangerous than interstates

https://crashstats.nhtsa.dot.gov/api/public/viewpublication/810625

Not really - this doesn't tell the whole story (for eg. miles travelled, vehicle density etc).

But its undeniable that you have less time to react at higher speed than at lower speed. Also, most of the crashes in urban areas happen because people ignore traffic signals. Automation can really help here.

 

[qdeathstar]

Not really - this doesn't tell the whole story (for eg. miles travelled, vehicle density etc).

But its undeniable that you have less time to react at higher speed than at lower speed. Also, most of the crashes in urban areas happen because people ignore traffic signals. Automation can really help here.

The data shows what it shows. You can have alternative facts of you’d like, but keep in mind that those are wrong.

Well I feel like If both of @powertoold you bothered to click the link you would see that most fatal accidents happen at sub 35mph speeds. It doesn’t really matter that their are more miles driven (which I sort of doubt) or if the accident involved a pedestrian or not. A fatal accident is a fatal accident abs for what ever reason, they occur most often on city streets where fsd operates.

Whether that is because it is more risky, difficult, involves pedestrians, or is just where people drive most of the time, the reason doesn’t really matter.

It’s inherently more dangerous for all of those reasons.

It’s not just a random chance that Tesla chose high speed interstate driving to work on and release first, is it?

 

[EVNow]

The data shows what it shows. You can have alternative facts of you’d like, but keep in mind that those are wrong.

Well I feel like If both of @powertoold you bothered to click the link you would see that most fatal accidents happen at sub 35mph speeds.

Don't assume stuff. I looked at it. I've been looking at these stats for a few years now to evaluate AP/FSD.

Can you tell me how many miles and trips those lower speed drives actually represent as a % ? If 80% of trips, for eg., are those drivers would you expect higher % of accidents during those or not ?

Interpreting stats is as important as linking to them.

ps : You should see SHRP data for for more meaningful stats.

Stuff like this : Crash Factor Analysis in Intersection-Related Crashes Using SHRP 2 Naturalistic Driving Study Data

 

[qdeathstar]

Can you tell me how many miles and trips those lower speed drives actually represent as a % ? If 80% of trips, for eg., are those drivers would you expect higher % of accidents during those or not ?

care to fill me in how you think that is relevant?

 

[rxlawdude]

It's interesting that you mentioned bar coded medicine administration. This eventually helped reduced errors, but introduced entirely new classes of errors where nurses would bypass the system by wearing duplicate barcodes on their person instead of scanning the ones on the patient.

Point being that automation isn't always good and it can and will introduce new failure modes that have to be considered. Sometimes these new failure modes can outweigh the initial problem they were trying to solve.

Those workarounds you mention were in the very early days of BCMA (2004-2006). I actually saw that workaround used at a VA hospital that supposedly was using BCMA. Because the workstations couldn't fit into the rooms and had to stay in the hallway, nurses photocopied the BCs for not only the drugs, but the patients armbands as well! What could go wrong?

I'm happy to say those days are pretty much long gone. The reason is, when those intransigent nurses would use the system as designed and got caught almost making an error, they got into shape really quickly. Or, they were devastated by finding out their workaround cost a patient more time in the hospital or worse.

My point is, when a system saves your bacon (when used properly), you tend to respect what it takes to use that system. On the other hand, if you perceive the system as inconvenient, you work around it (see the "AP Buddy" steering wheel weights, for example). Until your workaround kills or injures someone.

 

[powertoold]

care to fill me in how you think that is relevant?

There are a lot of details missing from the document you posted, so it's difficult to make any conclusions, like:

1) What is the total land area defined as rural vs urban?
2) How are urban and rural defined? Does urban mean both highway and off-highway?
3) Vehicle density
... so on

 

[qdeathstar]

There are a lot of details missing from the document you posted, so it's difficult to make any conclusions, like:

1) What is the total land area defined as rural vs urban?
2) How are urban and rural defined? Does urban mean both highway and off-highway?
3) Vehicle density
... so on

But, so what is your point though. How is an of that important?

 

[EVNow]

care to fill me in how you think that is relevant?

Crash rates depend on a large number of variables. Need to do a multi-variate analysis rather than look at a single variable.

The 25 Top Causes of Car Accidents in the US (2024)

Learn some of the top causes of car accidents on the road as well as helpful tips to help prevent future car accidents while driving.

seriousaccidents.com

ps : This is a major issue which a lot of academics study. The basic objective is to design a road system that reduces crashes.

An example of a multivariate analysis :

A multivariate random-parameters Tobit model for analyzing highway crash rates by injury severity - PubMed

In this study, a multivariate random-parameters Tobit model is proposed for the analysis of crash rates by injury severity. In the model, both correlation across injury severity and unobserved heterogeneity across road-segment observations are accommodated. The proposed model is compared with a...

pubmed.ncbi.nlm.nih.gov

 

[powertoold]

But, so what is your point though. How is an of that important?

I don't really have much of a point, other than to say that the document you posted is lacking in a lot of details. It's difficult to make conclusions based on it.

You're also reading the document incorrectly. You said that most fatal accidents occur at sub 35mph speed limits. But that's not what the graph shows at all. The graph only shows the proportion of fatal accidents (urban vs rural) within a speed limit range. All the bars within a speed limit range add to 100%! Please reassess!

 

[Terminator857]

Most accidents happen on city streets though. Most fatal accidents happen on roads with a low speed limit. City streets are inherently more dangerous than interstates

https://crashstats.nhtsa.dot.gov/api/public/viewpublication/810625

I don't see that conclusion from the url.

> Most accidents happen on city streets though.
Where are you getting that info? Your supplied reference says more fatal accidents happen in rural areas.
Quote:
In 2005, there were 21,924 fatal crashes (56%) in rural areas and 17,265 fatal crashes (44%) in urban areas.

>Most fatal accidents happen on roads with a low speed limit.
Your supplied reference doesn't say that. It does say the following:
Approximately 70 percent of all fatal crashes on roadways with speed limits of 40 mph or less are in urban areas. Slightly less than half (47%) of all fatal crashes occurring on roadways with speed limit between 45 and 50 mph are in rural areas. Over 70 percent of the fatal crashes on roadways with speed limit of 55 mph or higher occur in rural areas.

Lets break that down a bit:
>Approximately 70 percent of all fatal crashes on roadways with speed limits of 40 mph or less are in urban areas.
Of the few fatal crashes in urban areas 70 percent of them occurred with speed limits of 40 mph. Could be that reckless driver was driving at 60 mph though.

 

[S4WRXTTCS]

In

interesting take, and I’m not 100 percent saying your wrong. But really what you need is better driver attention tracking. Then you can deliver level 2 at a high quality level and insure the driver is paying attention.

Driver monitoring is useful at preventing willful driver inattention like texting while driving, doing makeup, etc. If done well it can also help prevent drivers from falling asleep on the highway.

Driver monitoring won't help much to prevent erosion of driver skill, and subconscious inattention like day dreaming.

Being able to drive safe is really about maintaining situational awareness, and I fail to see how a significant percentage of us could manage that over a long time while we're not in control of the task of driving itself. We also live in the time of attention deficit disorder where a pretty substantial part of our population is can't even maintain their undivided attention on a teams call let alone the task of monitoring an L2 vehicle while traveling through a busy city.

There is ample scientific evidence to back this up, and this is why the easiest way to ask if there is a pilot in the room is to talk about humans overseeing automated tasks.

Now just because its dangerous doesn't mean we shouldn't do it, but that we should recognize the danger as it will happen.

The biggest problem I see is liability as L2 manufactures aren't forced to improve their L2 solution if human drivers take all the blame, and its the humans insurance company that has to foot the bill.

Tesla insurance might be the solution to this as it forces Tesla to improve their FSD system to reduce the cost of insurance.

I wouldn't be the least bit surprised if I didn't end up having to switch to Tesla insurance as FSD Beta got to a point where I could get at all comfortable with it. I haven't been comfortable with an L2 system so far at all. They've all had some fault that was clearly evident after a half dozen or so 400 mile trips.

 

[Terminator857]

Fatality Facts 2021: Urban/rural comparison

A summary of fatality statistics about urban/rural comparison compiled by IIHS from 2021 Fatality Analysis Reporting System (FARS) data.

www.iihs.org

Motor vehicle crash deaths by speed limit and land use, 2019
Rural Urban Total* Deaths % Deaths % Deaths % ≤35 mph 1,387 8 5,972 30 7,423 21 40-50 mph 2,827 17 7,037 36 9,881 27 55+ mph 11,666 71 5,592 29 17,270 48 Total* 16,340 100 19,595 100 36,096 100
According to link above:
Crash deaths at <= 35 mph = 21%

Motor vehicle crash deaths by road functional class and land use, 2019
Rural Urban Total* Deaths % Deaths % Deaths % Interstates and freeways 2,360 14 3,773 19 6,133 17 Arterial 3,913 24 11,578 59 15,491 43 Collector 6,809 42 1,773 9 8,582 24 Local 3,218 20 2,443 12 5,661 16 Total* 16,340 100 19,595 100 36,096 100
In 2019 19,595 deaths occurred in urban environment, of those 12% occurred in local urban streets.
Of the total 36K deaths that occurred in 2019 6.77% occurred in local urban streets.
If we add local + collector streets in urban, then (2,443 + 1,773) / 36,096 = 11.68% of fatal accidents.