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

[Prunesquallor]

In the short time I have used it, I can say it’s the best drive mode, probably the safest too since it’s default is to come to a complete stop, not 5mph drive into stuff like creep, or the quick gas/brake/gas/brake when you are in “normal” at parking speeds. Where did you get the info that the parking brake is being used? When you put the car in park you can hear it engage, I heard nothing like that using hold.

My bad, I should have said “Vehicle Hold” not “Parking Brake" (the H indicator on the display). I’ll try to edit my post.

Edit: too late to edit

 

[mongo]

Since it’s the weekend...

I don’t know how many have gotten 36.2.1, but "Hold" Stop Mode seems to allow significant reduction in non-regen braking during normal driving. In “Hold”, regen takes out everything down to 0 speed (at which time the parking brake is automatically applied).

I wonder if this significantly reduces brake wear and/or fluid replacement requirements? Since brake fluid replacement is one of the few scheduled maintenance items, increasing the interval could be of benefit to high-mileage use cases (such as, oh, I don’t know, autonomous taxis maybe?).

edit: On the other hand, "Hold" may just be there to eke out a little bit more range.

Hold gets you closer to full one pedal driving which has been a differentiator of other EVs.

As to brake fluid life, the lack of use (no heating of fluid to drive out water) was the reason for the cautionary 2 year fluid replacement. That maintenance item changed from scheduled to as needed previously.

 

[ZeApelido]

Like I said previously, context is always being lost. You are comparing a 2019 chip with a 2015 chip. Ofcourse there would be magnitude speed improvement at lower or similar cost. Any AI inference chip in 2019/2010 destroys the Drive PX2, especially the board that Tesla was using which was a half Drive PX2. I don't understand what the hysteria is all about?


That's not a virtual Lidar, its an occupancy grid map using an RNN out their perception CNN's output of road edges. Again standard deep learning procedure.

There's absolutely no proof of this, confirmed by verygreen/greentheonly.
Infact the percentage of training data being uploaded currently is more like 0.01%


Its Game, Set, Match if you are using standard deep learning 101 architectures and techniques? But the companies like Google's Waymo/DeepMind who are actually inventing new NN techniques and architectures to tackle the problem is actually 5 years behind. Am i right?

Nope. It's about the data being fed into those architectures. You know Waymo doesnt have enough.

 

[Fact Checking]

That's not a virtual Lidar, its an occupancy grid map using an RNN out their perception CNN's output of road edges

Stop being obtuse: I was referring to the top down driveable space visualization, which all major Tesla competitors obtain via Lidar and high-res maps.

The video segment Karpathy has shown demonstrates how Tesla can do this without Lidar and high-res maps - which is why I called it "virtual Lidar".

Tesla could probably even open source all their NNs, and their competitors could still not catch them: Tesla has a fleet of soon to be 1,000,000 cars that are sending them training feedback all around the clock. This is the real NN training computer Tesla has built - which hardware of 1,000,000 cars has to be owned to run Tesla's architecture...

There's absolutely no proof of this, confirmed by verygreen/greentheonly.
Infact the percentage of training data being uploaded currently is more like 0.01%

LOL, you clearly know very little about this topic, you just keep repeating the same debunked talking points.

Firstly, you start with a logical fallacy, verygreen obviously cannot "confirm" a negative:

• Tesla might be detecting rooted cars and doesn't include those cars in training campaigns, to protect their trade secrets.
• As mentioned to you before, verygreen doesn't have access to the LTE packets, only to Wifi packets - which might limit the amount of NN feedback he is able to capture.
• Tesla does targeted training campaigns, and verygreen's car in France might be a low priority test candidate. If Tesla does most of their training in the U.S., to keep LTE costs down and because European Autopilot has regulatory limitations, then verygreen's car would not be normally part of the training feedback fleet.
• We also don't know whether verygreen has managed to find all training related communications channels - reverse engineering without the source code is a very inexact science.

You ignored these counterarguments you got in the past, and you continue to parrot your talking points.

Its Game, Set, Match if you are using standard deep learning 101 architectures and techniques?

No, it's game, set and match due to the second sentence of that paragraph, what you excluded from your quote:

Tesla could probably even open source all their NNs, and their competitors could still not catch them: Tesla has a fleet of soon to be 1,000,000 cars that are sending them training feedback all around the clock. This is the real NN training computer Tesla has built - which hardware of 1,000,000 cars has to be owned to run Tesla's architecture...

I highlighted the operative part.

None of Tesla's FSD competitors has a fleet even approximately this large - the closest one is more than 2 orders of magnitude smaller - so your only trolling tool left is to deny the reality that Tesla's huge neural networking training fleet exists.

 

[Unpilot]

Hmm looks like a argument with a ghost on my end.
Let me guess the ghost knows everything about FSD and AP and sig line brags about how smart they are.

 

[printf42]

Hmm looks like a argument with a ghost on my end.
Let me guess the ghost knows everything about FSD and AP and sig line brags about how smart they are.

Don’t forget roofing and trucking, lots of expertise they have.

 

[ggr]

Mod: posting a link to a porn video is not acceptable under the terms of service. Personally I don't care, but as moderator I/we will respond forcefully. --ggr

 

[Bladerskb]

Stop being obtuse: I was referring to the top down driveable space visualization, which all major Tesla competitors obtain via Lidar and high-res maps.

Occupancy Grid in other companies is using the combination of output of all sensors (Cameras, Lidars and Radars). That doesn't mean they can't use just cameras. That's not being obtuse, there's a clear difference.

The video segment Karpathy has shown demonstrates how Tesla can do this without Lidar and high-res maps - which is why I called it "virtual Lidar".

Occupancy Grid Map isn't virtual Lidar and doesn't need Lidar or HD Maps. The footage Karpathy showed used road edges output.

This is virtual Lidar. (5 mins)

LOL, you clearly know very little about this topic, you just keep repeating the same debunked talking points.

For a guy who knows very little about the topic, i'm definitely glad I'm not calling Occupancy Grid Map virtual Lidar.
OGM has nothing to do with lidar. Occupancy grid mapping - Wikipedia

Tesla might be detecting rooted cars and doesn't include those cars in training campaigns, to protect their trade secrets.

Except that verygreen's findings has been confirmed and collaborated by others in the community, including the exact data that is actually being sent through Wifi. When you actually check your data.

As mentioned to you before, verygreen doesn't have access to the LTE packets, only to Wifi packets - which might limit the amount of NN feedback he is able to capture.

No he sees everything. He is not capturing the WIFI packets. (not that Tesla actually sends stuff through LTE because its too big.) He's looking directly at AP cpu code, instructions and storage area.

Tesla does targeted training campaigns, and verygreen's car in France might be a low priority test candidate. If Tesla does most of their training in the U.S, to keep LTE costs down

verygreen communicates and collaborates with other people within/without the Tesla rooted community all over the US and regularly confirms his findings.

and because European Autopilot has regulatory limitations, then verygreen's car would not be normally part of the training feedback fleet.

You are literally just making stuff up as you go. EU's ADAS regulations which was passed very recently has nothing to with Tesla's data collection which has been the same and verygreen has been tracking it since went live early 2017. In-fact everything that Andrej said in Autonomy Day verygreen had already disclosed 2 years earlier.

We also don't know whether verygreen has managed to find all training related communications channels - reverse engineering without the source code is a very inexact science.

He's a researcher that backs up everything he says with actual evidence and data. You would think that would be something someone like you would want to see but i guess when it doesn't cater to your thesis.

You ignored these counterarguments you got in the past, and you continue to parrot your talking points.

Can you point to the exact post? And you call these counterargument? You literally just made up and tried to use the recent EU ADAS regulation as an excuse.

No, it's game, set and match due to the second sentence of that paragraph, what you excluded from your quote:

And all the evidence and data points to the fact that its not happening the way you think its happening. verygreen actually looks at the code and knows what the triggers are and what kind and data the triggers uploads, what exact data is put into the RAM to be uploaded, when the upload occurs, all the details. but you reject it for some reason.

I guess i will take your words over a large group of people who actually has access to the actual ap cpu logic code and ap internals and are saying the same thing.

But it looks like Facts doesn't matter so continue on because I'm done here.

Twitter

 

[EVNow]

The real numbers are even larger - likely one crash per 150k miles per car in the US after accounting for non reporting of minor accidents (for insurance reasons) which are still detected by Tesla's software. This puts Teslas currently at 18x less crashes per mile. You are saying other luxury cars do better than that? Using what superior technology exactly?

The actual numbers are one accident per 10k miles. This is from the Virginia Tech study.

Yes it does. its called the insurance company.

Which insurance company is releasing such numbers ? A link would be great - otherwise its just a TSLAQ conspiracy theory.

BTW, don't you work for a competitor of Tesla and shouldn't you disclose that since that would be a conflict of interest ?​

 

[TheTalkingMule]

Maybe this FSD conversation could take place in the FSD-appropriate forum?

 

[Silent Ludicrosy]

The Tesla truck unveil happens 4 days after the Ford SUV unveil. Unfortunate timing for Ford.

https://www.ford.com/powertrains/battery-electric-vehicles/

Good news for everyone is they’re listing the advantages of BEVs on their website now.

 

[S3XY]

TLDR...

If you don't have two hours to watch this I have a suggestion. Watch it at at least 1.7X speed. If you only have 5 minutes, skip to the 44:00 mark and watch through 49:00. It's by far the best part of the whole video.

 

[EVNow]

And all the evidence and data points to the fact that its not happening the way you think its happening. verygreen actually looks at the code and knows what triggers uploads, what is put into the RAM to be uploaded, all the details. but you reject it for some reason. But it looks like Facts doesn't matter so continue. I'm done here.

Twitter

Soylent has already said green doesn't know a lot of things that are happening. Would you believe someone who has a team of professional hackers or a single amateur ?

Even if you assume green knows everything, 8 to 10 sec video gives a lot of info.

ps : Ofcourse Tesla will be picking what they need. Not all video from 750k cars is uploaded all the time.

 

[EVNow]

Musk : As an outside investor there are two things you should track
- FSD
- Battery

TMC :​

Maybe this FSD conversation could take place in the FSD-appropriate forum?

 

[MFranc123]

• We also don't know whether verygreen has managed to find all training related communications channels - reverse engineering without the source code is a very inexact science.

When people talk about reverse engineering, it's almost always to refer to the case where you don't have the source code. It's very time consuming, but it can be very exact. You get all the instructions that make up a program, which is usually enough to discover whatever you want about it, even if packed (which Tesla software is not). Certainly reverse engineers don't really care about source code. They're used to working with assembly and pseudo-C from hex-rays (and now Ghidra).

I'll leave the whole rest of the discussion to you both as I know close to 0 about that

 

[EVNow]

Dojo was mentioned once or twice before but without much detail other than that they would use it to train from video (rather than only images - as it takes so much power to do just images, you can imagine how much more performance is needed for video).

From autonomy day, Elon said:
“We do have a major program at Tesla which we don’t have enough time to talk about today called “Dojo”. That’s a super powerful training computer. The goal of Dojo will be to be able to take in vast amounts of data and train at a video level and do unsupervised massive training of vast amounts of video with the Dojo program – or Dojo computer.”

The hope / speculation has been Dojo does end-to-end training. Basically you just input video - and the output tells how to drive. No object recognition or anything like that. Essentially it is unsupervised and may be not even labelled.

That ofcourse would need massive amount of training time and hardware. Karpathy basically talked about the hardware part. He didn't say anything about the end to end training part - though he mentioned fully automated training - "project vacation".

BTW, even with the massive pressure on autonomy team, its good to see them maintain a sense of humor. Shows good morale.

 

[goinfraftw]

The latest Global Risks report from the WEF...named "Out of Control".

"Is the world sleepwalking into a crisis? Global risks are intensifying but the collective will to tackle them appears to be lacking. Instead, divisions are hardening."

http://www3.weforum.org/docs/WEF_Global_Risks_Report_2019.pdf#page=3

Top 5 (in order of likelihood):

1. Extreme weather events
2. Failure of climate-change mitigation and adaptation
3. Natural disasters
4. Data fraud or theft
5. Cyber-attacks

Further, this was an interesting chart in private/public capital.

Also, this was wonderful. /sarcasm

"Researchers last year studied the trajectories of 126,000 tweets and found that those containing fake news consistently outperformed those containing true information, on average reaching 1,500 people six times more quickly."

 

[Tslynk67]

This. While a large chunk of my brain says it was really him, it just seems weirdly not quite his style?

Looks like the Twitter dude's style to me.

 

[JustMe]

If you don't have two hours to watch this I have a suggestion. Watch it at at least 1.7X speed. If you only have 5 minutes, skip to the 44:00 mark and watch through 49:00. It's by far the best part of the whole video.

Who is this guy? He’s adorable! I’d love to see a conversation between him and Gali of Hyperchange. The tortoise and the hare

 

[Doggydogworld]

Why would you claim Tesla accident rates are worse than BMW/Mercedes etc?

It is not even possible for other luxury manufacturers to track their accident rates so the data you are claiming just doesn't exist.

It seems highly improbable though. Taking Tesla's Q3 reported one crash per 2.7 million miles for cars without Autopilot but with Autopilot's free safety features, Tesla cars crash c.10x less regularly than the US average (that is 2.7 million miles divided by NHTSA's reported market average one crash per 498k miles multiplied by NHTSA's reported 1.8 cars on average per crash). The real numbers are even larger - likely one crash per 150k miles per car in the US after accounting for non reporting of minor accidents (for insurance reasons) which are still detected by Tesla's software. This puts Teslas currently at 18x less crashes per mile. You are saying other luxury cars do better than that? Using what superior technology exactly?

IIHS publishes driver death rates. The latest is model year 2014, we should see MY2017 in six months. The average 2014 rate of 30 per million registered vehicle years is roughly 1 driver death per 400 million miles. That's very close to Tesla's 1 death per 320m miles after adjusting for driver/passenger ratio. The IIHS page and linked status report show luxury sedan & SUV death rates are about 1/2 the average car (due in part to more driver assist features). Since Tesla death rate is close to the average car it's ~2x the average luxury car/SUV.

A few have done more detailed analysis and found higher ratios. But they are short sellers with black hearts full of Musk-hate, so we know every word they write is a blatant lie.

How does Tesla claim much lower death rate than average? Because they compare to the entire fleet, not late model luxury vehicles. The fleet contains mostly older cars (average age is 11-12 years) which have less safety equipment. The IIHS page shows model year 2008 death rate was 60% higher and 2004 was almost 3x the 2014 death rate. And those rates were measured when the cars were new. As cars age they fall into disrepair and tend to be driven by less safe drivers (younger and poorer) in less safe areas, so per mile death rate tends to rise over time.