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

[Johnny Ma]

i may have missed it, but did karpathy explain *how* they were able to make requests from the fleet to get more training data? it sounded like they'd identify some edge cases, then ask the fleet to provide more...but how does the fleet *know* what looks similar to the edge cases?

I have minimal experience with neural nets. know the basics, have used a pre-trained net to solve a problem at work. but this request sounds like magic to me. i must be missing something.

 

[SOULPEDL]

Tesla is competing with ICEVs for now, BEVs hitting the market are sparse and that will change.
Tesla is working on FSD and there are others doing the same, well funded, no distractions.
The moment one other company manages to get Level 5 going, it's a price war right there. If Tesla sticks to its own fleet, it can be outnumbered 100:1 for robotaxis on the streets within a few years.
25-30% will not fly anymore. Taking money for FSD software may not fly anymore.
Imagine Tesla sticks to $35K+ cars and then $15K FSD capable econobox BEVs start hitting the market. How much will riders pay extra to be brought to the airport or work by Tesla whom they know make much more money, as the novelty wears off?

Seems like ur ignoring Elons main point. If they are using Lidar, it's a non starter. Are the others using cameras... anyone?

 

[capster]

Yes, it was implied that right now, the NN just does perception and the actual driving is hard coded software. But they are starting to work on a system where the NN would do all the driving. Definitely next level.

Which is where behavioral conditioning models might become relevant including the more esoteric stuff like biologically plausible models of spectral timing (i.e. adaptive timing and temporal discrimination during associative learning).

 

[TheMrX]

So ur assuming everyone missed the first part on the mission and how EVERY Elon prediction has come true? Timing aside while still leading. You are hearing competitors say one thing in public, they are freaking in private... IMO. I know I would.

NO

What I'm saying is that whether you or I do believe in robotaxi next year is irrelevant regarding our (at least mine for sure) financial capacity to buy TSLA stock.

That is why IMO this was the INVESTOR autonomy day. I think it was explicitly meant to diverge institutional money from UBER/LYFT/WAYMO in autonomous driving (we can also add GOOGLE and NVDIA while we are there).

The problem is Elon/Tesla lost a lot of credibility on predictions/projections in the last year. So I do hope that those investor are going to be completely blown away by the FSD demonstration, because headlines IMHO won't be enough to move share price up if it is only an Elon/Tesla statement.

 

[Mars ☰mperor]

Because missed deadlines are used to say Tesla has execution problems and depress the stock price and excitement for future products by saying they can’t deliver

Mainstream media has been saying these forever.
No surprise here either.

 

[Remus]

i may have missed it, but did karpathy explain *how* they were able to make requests from the fleet to get more training data? it sounded like they'd identify some edge cases, then ask the fleet to provide more...but how does the fleet *know* what looks similar to the edge cases?

I have minimal experience with neural nets. know the basics, have used a pre-trained net to solve a problem at work. but this request sounds like magic to me. i must be missing something.

he explained with several examples, 1. ask the fleet to return tunnel pictures. this seems to be straight forward, as recognizing whether you are in a tunnel or not is a solved problem. 2. cutting in: ask the fleet to send in clips with cars in side lanes moving into the center right in front of you.

So basically they specify the rules and deploy the rules to the fleet, fleet upload clips if some of the rules are triggered.

 

[Right_Said_Fred]

Well, contrary to popular belief, Waymo is neither in the "car selling"- nor in the "ride hailing"-business.

Instead Waymo is supposed to become to cars and Robotaxis what Android is to smartphones – a licensable piece of tech. And they very well be able to pull this off, again!

Sure, you can laugh at Waymo's current capabilities and required hardware, yet, they are still miles ahead of what most traditional automakers can do today. If said automakers fall behind even further, they might one day have no other option than to license whatever tech is available on the market.

It's not a matter of Waymo just transferring software to car manufacturers. Those manufacturers have to start producing 'police vehicles' on a massive scale.

 

[Cloxxki]

Panasonic needs to get their production up there, but let's be honest. Whether you produce 7,000/week or 10,000/week robotaxis doesn't change the valuation much compared with the first few thousand you produce.

The thing with robotaxis...
You don't need to build them after the software is ready. Tesla is showing that.
Ideally you'd have a car with sufficiently strong powersteer, but accelerator and brakes can be retrofitted as can of course camera and computer. Let alone if done for one specific make of car that millions are already on the road of. A handful of sensors, another handful of wires and a chip, that's it.
Once the software is made available, getting a fleet organized will not be a problem. If the software works for a certain market but there is no BEV infra, the tech would still steal the drivers' jobs. A robotaxi can very well be run on dino juice.
Imagine a specific type of used VW Golf from years ago isparticularly easy to turn into a robotaxi. $5K for the used car, $1K for the hardware, and you're ready to start competing with Tesla. And it would be like stealing candy from a baby.

 

[SOULPEDL]

I think he was saying that starting tomorrow FSD will be the main focus of marketing a Tesla car. Right now the focus is that electricity is cheaper than gas, but tomorrow they will push FSD as why someone should buy Tesla over BMW etc

Main, yes as long term. Short term, its a bad ass car, sells itself, self funding network. Its further reason to buy a Tesla ON TOP OF existing reasons.

 

[Anstandswauwau]

i may have missed it, but did karpathy explain *how* they were able to make requests from the fleet to get more training data? it sounded like they'd identify some edge cases, then ask the fleet to provide more...but how does the fleet *know* what looks similar to the edge cases?

I have minimal experience with neural nets. know the basics, have used a pre-trained net to solve a problem at work. but this request sounds like magic to me. i must be missing something.

In layman's terms:

1. They bundle known edge cases
2. Using that data, they train the NN to detect those cases.
3. And then ask the fleet for more, matching those criteria.

 

[bdy0627]

We probably aren't going to get a lot of detail about the rides due to NDA.

 

[Cloxxki]

Seems like ur ignoring Elons main point. If they are using Lidar, it's a non starter. Are the others using cameras... anyone?

He could be right, but he may also underestimate the competition. Time will tell.
Even if they achieve Level 5 first and start rolling out the network, they are not safe at all. They are doing it with fancy cars in very limited supply. Google/Apple/etc could double down and make a big effort to catch up. It's not about how long your head start is, it's about whether you can compete when the monopoly ends. When, not if. What's the business plan, build GF4 to make more $50K robotaxis while a big hitter with a more recognized brand start retrofitting the existing fleet of all other car brands?

 

[Johnny Ma]

In layman's terms:

1. They bundle known edge cases
2. Using that data, they train the NN to detect those cases.
3. And then ask the fleet for more, matching those criteria.

Thanks a bunch. that did it for me

 

[SOULPEDL]

Lots of skeptics here. My guess is some supporters have dropped off because options are killing them. Anyway...
So, anyone seen test drives or did I miss something? Nothing on YouTube yet either that I could find. Maybe they are cameras off rides.
What about the videos days ago???

 

[capster]

i may have missed it, but did karpathy explain *how* they were able to make requests from the fleet to get more training data? it sounded like they'd identify some edge cases, then ask the fleet to provide more...but how does the fleet *know* what looks similar to the edge cases?

I have minimal experience with neural nets. know the basics, have used a pre-trained net to solve a problem at work. but this request sounds like magic to me. i must be missing something.

Although I doubt this is what they are doing, you might look at this to begin to build your intuitions about this particular realm of magic: k-nearest neighbors algorithm - Wikipedia

 

[erha]

Any tweets or other information about how the test drives are?

 

[M3Rider]

Presentation on the AI is mind blowing.

I've been seeing the car driving better on each update.

Now I know why.

Fleet Data Collection -> Machine Learning -> Driving Like You Never Thought Possible.

Incredible.

And of course... the market doesn't get it.

Would be nice to have a poll of residents here who think SP would pop tomorrow. My opinion it won't until at least feature complete or the first regulatory approvals with robotaxis going to the streets.

I'm not so sure. For thinking they will have ride sharing next year Elon seemed monumentally unprepared to answer any questions about it. Just a hazy 20-30% cut for Tesla, no certainty as to insurance, no answer to steering wheel, etc.

He was just doing that by the seat of his pants. Someone should've reined him in (can't believe I'm joining those who've been saying that...)

He did say Tesla will assume liability for TN accidents, which means some of their 25-30% cut will pay for those accidents. Which is fair I think,
because the TN revenue is not one time money like FSD sale, but a continuous revenue stream proportional to miles driven.

The big question is what if the passenger takes control of the steering wheel and gets into an accident. I think the passenger is now liable. Because don't touch it. I don't want some drunk to grab the wheel of my car for fun. Probably won't consider lending it to TN if that's an option to a passenger.

 

[SOULPEDL]

Even if they achieve Level 5 first and start rolling out the network, they are not safe at all.

Pls explain this. Citation?

 

[mongo]

Lots of skeptics here. My guess is some supporters have dropped off because options are killing them. Anyway...
So, anyone seen test drives or did I miss something? Nothing on YouTube yet either that I could find. Maybe they are cameras off rides.
What about the videos days ago???

Video was shown before 3rd presenter (but not on the stream). May also be the one on Tesla's Twitter feed.

 

[Johnny Ma]

Although I doubt this is what they are doing, you might look at this to begin to build your intuitions about this particular realm of magic: k-nearest neighbors algorithm - Wikipedia

actually used k-nearest neighbors for an undergrad assignment thanks for the burst of nostalgia!