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

[Drezil]

Some interesting AI news relevant for Tesla FSD.

New way of doing object detection in images:

https://twitter.com/x/status/1441052423854956553

Google release new state of the art imagerecognition neural network efficientnetv2

Toward Fast and Accurate Neural Networks for Image Recognition

Posted by Mingxing Tan and Zihang Dai, Research Scientists, Google Research As neural network models and training data size grow, training efficien...

ai.googleblog.com

Note that Tesla are using RegNet according to AI day:

View attachment 713815

So what this mean? The first one is one potential architecture change Tesla can do in the future. They could leave the BEV image output space and just output a list of objects in text format. That would probably be a major rewrite so not likely to see that for a while.

The second one they can probably implement in few days and a few more weeks of validation to improve performance overall. The company they aquired deepscale.ai are experts at this and have done the same thing many times before.

RegNet is only for the first layer before integration to vectorspace.
If you compare regnet to efficientnet they barely differ given same number of params and flops - but I guess it is not as fast on the HW3 chip.

They took regnet BECAUSE it is so simple & fast. As first thing.
That is like the first filter in your eyes, before the information is sent into the visual cortex. Speed is key here for sorting out the useless stuff. Accuracy comes later.

The first thing is also in image space and thus not relevant to Tesla. This thing sounds like "we threw a 2d transformer at the task and it worked"..
Tesla uses a 1d transformer in time in vector space (not image space) with added RNN as neural memory in later layers.

More interesting is another paper kaparthy liked some days ago. It should be the latest one talked about by Yannick kilcher on YouTube. Still on my playlist so I cannot say much about it.

 

[Artful Dodger]

In real life there are seldom consequences for cheating your way out of work. Quite the opposite. You get paid good money to do that.

So you work for VW? (cough)

 

[Drezil]

So you work for VW? (cough)

No.. but I have seen it way too many times in my life to spot the shortcuts

Besides: pay at VW is abysmal.. 60-80k or so for an engineer..
And they want to throw bodies at the problem.. like 3k programmers or so.

If 1 woman can get a baby in 9 months, then 9 woman can get a baby in 1 month!!

I always tell that the Tesla autopilot team is around 60 people for programming. That is enough. Plus the 1000 "unskilled" labelers that annotate the data.

In Germany it is easier to have a 1 hour meeting with your group (~25 people), than to get 1 additional server for ml-training. Business costs would be the same ..

At my old company we were 2 pure ml people + a student. Around 10k salary each month + office space + ...
We had one too small rented server for training our stuff. For 150 bucks/month. There were always talks of upgrades - but in 2 years nothing happened.
So we often sat idle while the server ran...

Management won't get those things.. sadly ..

 

[Buckminster]

https://twitter.com/x/status/1441682601874374663

https://twitter.com/x/status/1441684426400976896

 

[jkirkwood001]

OT, but Elon-related and by definition TSLA-related. Elon & Grimes have separated. We'll see if there is any impact from his personal life on his work life. Not always easy to separate the two.

https://twitter.com/x/status/1441429974326579208

Nobody will believe me, 20/20 hindsight and all that, but when I saw Elon at AI Day, when he seemed subdued on stage at first, I wondered if he had some sad personal news distracting him. Now it seems obvious.

We TMCers obsess about TSLA, Elon's business acumen, and his engineering performance. But we mostly ignore his personal life, like it doesn't happen, as if he lives 24 hours a day only for our obsession. Let's appreciate (despair?) how much he puts his romantic life second to "the mission" for our collective (financial) benefit. Not sure how good a dad he can be given his priorities.

Regardless, as a divorcee like him, I'm very sad to learn he still hasn't found a new lasting love for his emotional health. I read once where he said "it sucks going to bed alone" (paraphrasing). Amen to that (not religious BTW).

Be well, Elon.

 

[PokerFJÆS]

Acceleration is ok!

Safety Score Beta | Tesla Support

The Safety Score Beta is an assessment of your driving behavior based on several metrics called Safety Factors. These are combined to estimate the likelihood that your driving could result in a future collision. We combine your daily Safety Scores (up to 30 days) to calculate the aggregated...

www.tesla.com

 

[Artful Dodger]

No.. but I have seen it way too many times in my life to spot the shortcuts

Besides: pay at VW is abysmal.. 60-80k or so for an engineer..
And they want to throw bodies at the problem.. like 3k programmers or so.

If 1 woman can get a baby in 9 months, then 9 woman can get a baby in 1 month!!

I always tell that the Tesla autopilot team is around 60 people for programming. That is enough. Plus the 1000 "unskilled" labelers that annotate the data.

In Germany it is easier to have a 1 hour meeting with your group (~25 people), than to get 1 additional server for ml-training. Business costs would be the same ..

At my old company we were 2 pure ml people + a student. Around 10k salary each month + office space + ...
We had one too small rented server for training our stuff. For 150 bucks/month. There were always talks of upgrades - but in 2 years nothing happened.
So we often sat idle while the server ran...

Management won't get those things.. sadly ..

Yeah, the "cough" was a jab at the Dieselgate cheating scandal. I guess too subtle.

Cheers!

 

[RobDickinson]

Is it better to ruin your score by braking or just crash?

 

[jw934]

[edit: self-claimed] Tesla engineer Amin A. tweeted info/confirmation
- auto drive industry severely under-estimated computing power needed for vision
- Hw4 will be needed for level 5
- HW4 will be free for all who purchased FSD, in his opinion
- current self park uses only ultrasonic sensors. This will be replaced by one that uses both vision+sensor
- FSD is very hard for the Tesla team. Data to be gathered from 10.1 “button” will speed up the process a bit.

IMO, since HW4 and “driverless FSD” will be made free, full recognition of FSD reserve as income won’t happen until then, to cover at least the cost of the retrofit.

https://twitter.com/x/status/1441621192641712137

 

[Todd Burch]

Do we have any confirmation on this Amin guy’s identity?

My understanding is that Tesla employees are forbidden from participating in any social media interactions involving Tesla-related information, so I’m skeptical he would risk his job to share this info.

At any rate, the safety score metrics look quite fair. My only concern was that if you’re approaching an intersection with a green light and the light turns yellow, it might encourage people to sail thru the light instead of brake appropriately in order to avoid getting dinged for braking events.

 

[heltok]

RegNet is only for the first layer before integration to vectorspace.
If you compare regnet to efficientnet they barely differ given same number of params and flops - but I guess it is not as fast on the HW3 chip.

They took regnet BECAUSE it is so simple & fast. As first thing.
That is like the first filter in your eyes, before the information is sent into the visual cortex. Speed is key here for sorting out the useless stuff. Accuracy comes later.

The first thing is also in image space and thus not relevant to Tesla. This thing sounds like "we threw a 2d transformer at the task and it worked"..
Tesla uses a 1d transformer in time in vector space (not image space) with added RNN as neural memory in later layers.

More interesting is another paper kaparthy liked some days ago. It should be the latest one talked about by Yannick kilcher on YouTube. Still on my playlist so I cannot say much about it.

Not sure why you are downvoting me. And sorry if we derail this thread. I will answer then leave the conversation.

Yes that’s the part they would replace to improve performance. If you compare it at inference time
EfficientNetV2: 24ms 84%
RegNet: 31ms: 82.9%

1% accuracy is a lot in this context and for a speed gain. Not sure what you mean about RegNet being simple, imo what matters is inference time, accuracy and if the code is available or not. The code change is relatively “simple” and the compiler should take care of the rest. If they get similar speed up HW3 as on A100 I don’t know, I would assume it to be similar.

I assume the paper you refer is this one I posted about last week.
Wiki - Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

Let’s be friendly and share papers, not downvote people who share them. And if we do, let’s make sure we actually read the numbers before we criticize them…

 

[ammulder]

At any rate, the safety score metrics look quite fair. My only concern was that if you’re approaching an intersection with a green light and the light turns yellow, it might encourage people to sail thru the light instead of brake appropriately in order to avoid getting dinged for braking events.

The only bummer for me in that list is that I reasonably often get a forward collision warning when a car in front of me is turning off the road, and I can tell it will be out of my way by the time I get there (though prepared to give it some extra space if needed or brake if it suddenly stops) yet the Tesla flags a collision warning anyway. I keep hoping the vision stack will take over from radar and recognize that it’s not a generic obstacle but a car departing the roadway!

I would have though speed relative to limit or running red lights were factors… curious that they are not but I guess that’s what the data analysis told them.

 

[ammulder]

So… do I need a software update to get whatever version will be capable of showing me the Beta Button?

I’m on 2021.24.11 and never seem to get releases the first day (though FWIW I do have ‘advanced’ ticked).

FWIW this morning I got an update to 2021.32.22 and I see .32 is the minimum for the safety scoring, so I guess I did need this update. It hasn’t finished so I don’t know if I actually got the button yet. Here’s hoping!

Edit: got the button; clicked it. 7 more days, sigh!

 

[Drezil]

Yeah, the "cough" was a jab at the Dieselgate cheating scandal. I guess too subtle.

Cheers!

It wasn't too subtle.. but the thing is.. these things go way deeper into Management..
All engineers knew it was wrong and let Management sign it off to cover their asses. But did it anyway to not lose their job.

Edit:
A bit of explanation for my American friends here:
After WW II everyone said they "just followed orders". So this is never a valid excuse in court. Even in Military you have to disobey orders if they go against human rights etc..
As an engineer cheating on emissions you can be held accountable for this personally(!). Not the company. And to go to your boss any tell him "I want it written that I have to do it and you will take the blame" and then get such things signed.. that is a whole other level.
I don't know how those bosses covered their assess.. maybe change companies & destroy records? I don't know. But there surely were legal loopholes..
In Germany you always have to find the one responsible for it. If you can't, then they get away.
A famous lawyer once told a small story: in a big company one in accounting managed the "not so legal" Accounts. And made sure everyone on the floor knew the password. In court everyone testified they COULD have done it and everyone said nothing more. You could not find the individual. Company good a small fine, but no one went to prison.

In the states you got way more money out of VW then we did.. although I think we have more polluters here... -.-

 

[Prunesquallor]

FWIW this morning I got an update to 2021.32.22 and I see .32 is the minimum for the safety scoring, so I guess I did need this update. It hasn’t finished so I don’t know if I actually got the button yet. Here’s hoping!

Edit: got the button; clicked it. 7 more days, sigh!

After you get 32.22, there is also a Tesla App update available (4.1.0).

 

[strago13]

I’m updating to .32.22 as we speak. Will report back about the button.

Edit: yup requested FSD beta. I’m in the queue now.

 

[SN_8]

I’m updating to .32.22 as we speak. Will report back about the button.

Installed and clicked. Unable to find my Safety score in my app though. Perhaps phone app update not rolled out to android yet.

 

[lascavarian]

and I can tell it will be out of my way by the time I get there (though prepared to give it some extra space if needed or brake if it suddenly stops)

Although I agree that there seems to be an overly cautious approach taken, this exact situation resulted in one of my few accidents. It is an OT discussion that could get lengthy so I will just say that there are a lot of variable outcomes that limit the accuracy of any predictive algorithm.

I expect that performance will improve a bit but the best solution requires a turn lane IMO.

 

[ammulder]

I notice that .32 also has the option to show the new FSD visualization (selected by default). Looks available even without enrolling, though I was too eager beaver to confirm? It wouldn’t surprise me if that required AP3 hardware, though.

 

[scubastevo80]

There are more notes posted by the AP Engineer.

1. FSD 10.1 will have reverse function
2. It will not have summon as part of the vision stack
3. Determining one as a good driver includes maintain a good following distance, and overall steady driving. Average speed above limit tolerance will be about 5-7 mph.
4. Expecting HW4 to be a free upgrade for all FSD purchaser(this is his speculation).

Great, so that basically means you need to be the slowest driver on the road on NJ highways or avoid them altogether...