Karpathy talk today at CVPR 2021

[heltok]

I wonder what Christensen would say about Apple and the cellphone industry. To substitute his thoughts as applied to Apple... "Apple is not disrupting the phone industry, it's just making existing phones better..." Just because he's from HBS doesn't mean he's right. I think most would agree this is an incorrect statement as Apple clearly disrupted the industry. So his statement on Tesla is equally incorrect. Tesla will do the same to legacy automakers and clearly has already disrupted the industry.

IIRC he was employed by Nokia and pretty much said exactly that:

Why Apple and its iPhone confound disruption theorists

Nokia, Motorola, Sony-Ericsson and BlackBerry-maker Research in Motion were all victims of disruption. During the 1990s and 2000s, they shepherded the cell phone during its period of takeoff into ubiquity…

theconversation.com

 

[alsetym]

I am also of the opinion that this is true (using a good ADAS designed for the ODD and combining the strengths of humans and machines is safer than either alone), but I always struggle to conclude from Tesla's data that it is. Which is disappointing, since they probably have the necessary data to show us that it definitely is, but they never publish that data. Arrggh.

I want it to be true, and I think it's true, but I for sure can't quantify how much safer it is, nor can I even see any data which shows that it is safer.

Probably the best, or more reliable, measurement is that Tesla reported accident rate keeps improving, so when it it 10x higher than their first report it is likely that it is at least 10x better than humans alone but not more than 100x. Maybe have to subtract a bit to account for more adas systems being used in general population...but then that is kinda the point right? And you have to give tesla some credit for kicking the ads industry forward faster than they would have. Will be very interesting to see what happens to the stats when the button appears, might crash. I could see them breaking that out separately.

 

[daktari]

Well, seeing as how AP is (arguably) already at 8.6x, is 10x that hard to imagine?

In the 1st quarter, we registered one accident for every 4.19 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.05 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 978 thousand miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 484,000 miles.
source

These data are not worth the keyboard it was written on. We discussed this in another thread. Too much bias, not comparing similar scenarios, not controlling driver demographics etc. Pure marketing, not evidence at all.

 

[mikes_fsd]

These data are not worth the keyboard it was written on. We discussed this in another thread. Too much bias, not comparing similar scenarios, not controlling driver demographics etc. Pure marketing, not evidence at all.

Just because you don't find the data useful doesn't mean that it's not useful!

At the very least, the data is useful comparing like for like! Current version of same fleet against previous version.

With vision only you can easily compare it to the legacy data from Tesla for the past few years.

And that is an apples to apples comparison.
So, if vision only statistics show better performance then the legacy stack - via the crash rate data that they currently publish - they know they are barking up the right tree!

 

[Knightshade]

These data are not worth the keyboard it was written on. We discussed this in another thread. Too much bias, not comparing similar scenarios, not controlling driver demographics etc. Pure marketing, not evidence at all.

This, too, is wrong- but in the other direction.

How the numbers change within their own category for example could provide useful information.

If over years you see the "on AP" numbers steadily decline- that strongly suggests AP is improving safety in situations AP tends to be used.

Ditto for the active safety features numbers.


Now, could tesla make the data more useful by providing a lot more specifics and details? Yes, tremendously.

The data they give now can certainly be, badly, misrepresenting by not considering the potentially huge differences by categories.

But that doesn't mean it tells us nothing at all.

 

[daktari]

Just because you don't find the data useful doesn't mean that it's not useful!

At the very least, the data is useful comparing like for like! Current version of same fleet against previous version.

With vision only you can easily compare it to the legacy data from Tesla for the past few years.

And that is an apples to apples comparison.
So, if vision only statistics show better performance then the legacy stack - via the crash rate data that they currently publish - they know they are barking up the right tree!

No. You need to know if the cars were driven on the same roads, in the same weather and time of day, at the same speed, with similar drivers etc. to compare. There is a difference in accident frequency on highway vs city.

So what they could do is identify a stretch of highway, track all data from that stretch and control for weather, time of day and driver. But driver will be uncertain, they probably only know who the owner is, and not know anything more than age and gender.

Btw: Disagreeing with my post is similar to disagreeing with the scientific method.

 

[daktari]

This, too, is wrong- but in the other direction.

How the numbers change within their own category for example could provide useful information.

If over years you see the "on AP" numbers steadily decline- that strongly suggests AP is improving safety in situations AP tends to be used.

Ditto for the active safety features numbers.


Now, could tesla make the data more useful by providing a lot more specifics and details? Yes, tremendously.

The data they give now can certainly be, badly, misrepresenting by not considering the potentially huge differences by categories.

But that doesn't mean it tells us nothing at all.

The way the data are presented by Tesla, and paraphrased by others here, are certainly misrepresenting. That is my point.

I really don't hope Tesla use these data, at the level of resolution they present to the public, for critical development decisions.

 

[Knightshade]

No it doesn't. You need to know if the cars were driven on the same roads, in the same weather and time of day, at the same speed, with similar drivers etc. to compare. There is a difference in accident frequency on highway vs city.

All cars on AP doesn't need any of that broken out.

Unless you think there's a MASSIVE difference in that stuff between, say, Q1 in year X and Q1 in year Y or something. (pandemic changes aside)


Do you think people suddenly stopped using AP on the same type of roads in one quarter they'd always used it on?


If you do think that- why?


So yes, the data remains useful at least within each category.

If accident rate of Teslas on AP declines, AP is likely getting safer.

Likewise if accident rate of Teslas with active safety features (but not AP) increases, same thing.

(same with decreases the other way of course).

 

[daktari]

All cars on AP doesn't need any of that broken out.

Unless you think there's a MASSIVE difference in that stuff between, say, Q1 in year X and Q1 in year Y or something. (pandemic changes aside)


Do you think people suddenly stopped using AP on the same type of roads in one quarter they'd always used it on?


If you do think that- why?


So yes, the data remains useful at least within each category.

If accident rate of Teslas on AP declines, AP is likely getting safer.

Likewise if accident rate of Teslas with active safety features (but not AP) increases, same thing.

(same with decreases the other way of course).

Well you might examine Tesla vs Tesla or brands ADAS only IF you can control the bias. But there is so much of it, some examples:

The driver:
-gender
-age
-experience
-drugs/alcohol
-after work/nightshift
-emotional status
-disturbing stuff like phone use, crying kids etc
Environment:
-weather: rain/snow/black ice etc
-time of day
-visibility (fog, blinding sun)
-traffic conditions (congestion, busy&fast, scarce etc)
-road surface condition
-Road markings
The car:
-tire condition and type
-technical condition
-Yoke or wheel ;-)

And probably more.

 

[Knightshade]

Well you might examine Tesla vs Tesla or brands ADAS

I might- except AFAIK no other brands provide any data at all on their ADAS vs not safety results.

 

[mikes_fsd]

Disagreeing with my post is similar to disagreeing with the scientific method.

Oh gawd, disagreeing (scepticism) is at the heart of science!

 

[rxlawdude]

Someone doesn't understand the Law of Averages.

 

[daktari]

I might- except AFAIK no other brands provide any data at all on their ADAS vs not safety results.

Correct, and we don't have a controlled baseline either. I asked the association of insurance companies in my country; no complete accident statistic is made at a nationwide level.

 

[stopcrazypp]

Correct, and we don't have a controlled baseline either. I asked the association of insurance companies in my country; no complete accident statistic is made at a nationwide level.

Well insurance companies have different concerns. They are trying to figure out the correct premium for a given individual or type of car, so they obviously need to break down their data more. I don't know about Norway, but here in the USA, the NHTSA certainly does have accident statistics on a nationwide level. What people also are looking at here is simply if the system is improving, and having the fleetwide average is good enough for those purposes, especially when talking about big factors (like multiples, not just a couple percent). Tesla obviously does have more broken down data (like what version is being used, Karpathy touched on that in terms of how they are tracking progress).

When you have a big enough dataset, a lot of the variables get averaged out (this fact is commonly used in statistics). You don't necessarily have to control for every little detail (that's only the case when you have relatively small datasets or you are doing experiments with only a few samples).

 

[nativewolf]

Correct, and we don't have a controlled baseline either. I asked the association of insurance companies in my country; no complete accident statistic is made at a nationwide level.

Right and Norway, great country-returning for a summer is on my bucket list; has a relatively small set of drivers and miles driven. My favorite trip was hitching rides to Nordcap via a diversion to the Lofoten Islands. Great month.

Tesla does far more than most but then that's also self serving because they have a product they believe in and a future product that requires people trust Tesla to improve safety. Tesla has a lot riding on FSD acceptance. Other companies have offerings for AP esque capabilities but they have no future, no FSD capabilities are envisioned.

 

[daktari]

I do believe insurance companies individually have detailed data, but those are kept secret and will have bias also.

We have nationwide statistics on fatal accidents, nothing below.

It is Tesla comparing AP vs non-AP vs "other cars" data that is wrong when we are not presented to the details. Some variables certainly need to be controlled.

But this is off topic for this thread.

 

[daktari]

Right and Norway, great country-returning for a summer is on my bucket list; has a relatively small set of drivers and miles driven. My favorite trip was hitching rides to Nordcap via a diversion to the Lofoten Islands. Great month.

Tesla does far more than most but then that's also self serving because they have a product they believe in and a future product that requires people trust Tesla to improve safety. Tesla has a lot riding on FSD acceptance. Other companies have offerings for AP esque capabilities but they have no future, no FSD capabilities are envisioned.

Welcome back, you will be amazed by the number of EVs and chargers here!

 

[nativewolf]

Welcome back, you will be amazed by the number of EVs and chargers here!

When I was 21 it was just fun to get whatever ride I could. Funniest was getting ride with a french family that was so so disapointed to have picked up a hitchhiking American, the second was a Japanese nature photographer with about 30 words of english, stayed with him for 2 weeks. Good times. Saw a 6 toed puffin hound with him. Saw birds that I have only seen in nature shows, saw the incredible network of power lines and dams that provide so much electricity to Norway. Anyway, wonder if the FSD would ever handle the coastal road?

 

[powertoold]

Tesla software has been stuck on 2021.4 for a while (~5 months or so), so it got me thinking about this Karpathy slide, where he says that within the last year, Tesla overhauled their NN structure (used to be a bunch of heads, but now it's heads -> trunks -> terminals), and the AI teams are mostly refining the "terminal" NNs nowadays, and on rare occasion, they would retrain / overhaul the heads and trunks (backbone). Perhaps .4 is the version of the head and trunk NNs, whereas the .4.X.Y are representative of changes in the terminal NNs. If so, it seems their backbone has been pretty robust.

 

[scottf200]

Doh!

Twitter thread:
v---------------------- ---------------------v

https://twitter.com/x/status/1409298726477352960

^---------------------- ---------------------^