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

[SOULPEDL]

Wow! Thanks for linking that interview. It was probably the best interview I have seen with Elon. Lots of great details and insights about the self-driving technology. Interviewer: "You're confident with Tesla you can create the world's best autonomous driving system?" Elon: "Yes. To me, right now, this seems game, set, and match. I don't want to sound overconfident, but that's literally how it appears right now."

Watched it too. If an FSD car "appreciates" over time, I think there might be a demand issue!

 

[printf42]

One small detail from the Lex interview puzzles me.
Elon says if you order a Tesla S, X and 3 now you get HW3, if the car is ordered with FSD.
He didn’t say all cars rolling out of the gate has HW3, maybe this means they still have some HW2 boards stock not depleted yet, or maybe he was thinking about existing fleet in the middle of the sentence.

But anyway, this officially confirmed HW3 is in production and to me that cleared the biggest risk in Tesla FSD roadmap. If the hardware is certified to run current networks, there is no doubt it would be able to run bigger networks, and bear in mind what current network is already capable of.

My opinion is now it might be the last chance for anyone who understands the implications of this to front run wall street bean counters.
Soon people would wake up and realized they need to apply Waymo valuation to Tesla, and probably multiple times over and over.

Not an advice, because before that happens who knows what kind of tricks Wall Street would pull out.

 

[Krugerrand]

We have rules and regulations, whether that causes unreasonable delays in permits or construction I don't know, but you can't just start building a factory where and when you want.

For sure in some cases you see projects under construction 24/7, but it's rare, mostly because workers in the EU are compensated appropriately for unsocial hours - 2x, 3x pay, with time-off in lieu on top of that. This isn't red-tape, feet-dragging or political mud, it's normal in western society.

I'd also hasten to add that several EU countries have made a pitch to Tesla with proposed sites, no doubt with some sweetenersunder the table, but Tesla didn't, to the best of our knowledge, bite yet.

I don’t think normal in Western society is going to cut it anymore than needing a Chinese partner in China cut it. Normal Western society is killing us.

Nevada won GF1 because they were willing to do things fast and cut through regulation and political red tape.

So who can you potentially see in Europe getting off their high normal Western society horse and get down to business? Or can you think of another place not in Europe that could fill the void?

 

[Rarity]

They don't need HW4. HW3 with double redundancy is only using 10% of compute capacity at full resolution, full frame rate.

We'll see if there are limits that present themselves on the "march of 9s."

 

[MarcusMaximus]

Well, that pushed the opposite direction from what I expected at close. Still closed roughly flat, but looks like market makers ceded the day.

 

[EVDRVN]

I'd like to highlight 1 point Elon made around 8:30 in the video (which is not new information per say, but the way he puts it is eye-opening): Tesla having close to 500K cars with sensor suit collecting data on the road vs all autonomus driving competitors all-together having in the order of 5K cars or less means, that Tesla collects about 99% of the data in the field. That is a huge advantage! Some would even say YUUGE

If we all opt out of the data and disconnect from the Tesla servers for a few months I'm sure we can leverage $500 HW3 upgrades for everyone But of course most could not live without their apps for a day.

 

[bdy0627]

Just reread the Autonomy Day blog post. No mention of an FSD test drive or demo. Literally just says:

“Investors will be able to take test-drives to experience our Autopilot software first-hand, including features and functionality that are under active development”

Willing to bet that Elon will show advanced summon/NoA and that’s it. Won’t be anything special. Just another pitch to investors. No one believes him anymore. Last three major announcements that we thought were bullish af were complete letdowns. $35k Model 3/store closing, 4 minute Model Y reveal with no S/X refresh, Q1 deliveries.

Hopefully this isn’t the case. Don’t want another letdown. I want Tesla to control the narrative again.

I'm hoping to see some progress, but I don't personally care too much what he shows there. If we climb before the event, I will definitely be prepared for a big sell the news. Check out the podcast interview of Elon by the MIT researcher. It's getting pretty clear where the neural net is at now. I think it's getting close. Lots of rave reviews on the latest navigate on autopilot. We really aren't that far from highway FSD. Elon confidently said he thinks we may be hands off in 6 months (I know, does it always have to be 6 months?!), definitely next year.

 

[MarcusMaximus]

They don't need HW4. HW3 with double redundancy is only using 10% of compute capacity at full resolution, full frame rate.

I think this isn't very well understood by a lot of people: The compute requirements are affected by resolution and frame rate, but are not determined by them. I could design an algorithm that "processes" full 8k video at thousands of frames per second even on relatively meager hardware. That processing just wouldn't amount to very much.

Taking those video frames and turning them into consistently safe and correct car control actions, on the other hand, will require quite a bit of compute power(nobody really knows for sure how much, yet). Running networks capable of that on low resolution or low frame rate may be far more compute hungry than running their current networks on full res/full frame rate.

They seem to have high confidence in HW3 being able to handle it, and Karpathy, in particular, is someone I'd trust to best make that judgement, but just taking their current performance at face value doesn't necessarily guarantee that.

 

[K3jSe7Llp7]

'teslas are appreciating assets'! EM in the podcast.
If they took that seriously leasing on their terms is very profitable. They could loan them out for free and take them back years later at a profit.

 

[carterm2]

If HW3 is so incredible... why not sell it... as a graphics card, or AI co processor. OEM it... see if your competitors can outpace your software development. Lower cost per part.

I will take 2.

 

[SOULPEDL]

I’m a big fan of Lex but disappointed in his questions.

Apparently he had a bunch of them tee’d up about the hand-off between FSD and the human, but I would have thought he could have pivoted better when Elon basically told him that’s all moot, because FSD will surpass humans in 6-18 months.

I’m no AI expert, but it seems like one central issue is that each 9 in the “March of 9’s” becomes more and more difficult because of edge case explosion. I wanted more detail on how Tesla is handling this edge case explosion without it involving increasingly intensive amounts of human labor in labeling...

To take one example of Road hazard detection:

It would seem that the car has to classify every object in the road along a spectrum of how safe it is to hit, vs. how imperative it is to avoid.

To get to 99%, it may be enough to identify pedestrians, wheels, plastic bags, roadkill and potholes.
To get to 99.99%, there may be hundreds of different road hazards, going up to 100’s of thousands as you march to 99.999%.

Am I missing something. Does someone know how Tesla can handle edge case explosion?

Not an expert if algorithms could perhaps better filter the edge cases for the truely novel experiences only...
Bigger picture is that the value of FSD should self fund to the pointwhere the car has zero profit with pure software growth. I keep coming up with options on how I want it to drive. Example, if Tesla were to charge an extra $200 for mad max, it would clearly sell to this human. This is an untapped business model ready to explode!

 

[MarcusMaximus]

If HW3 is so incredible... why not sell it... as a graphics card, or AI co processor. OEM it... see if your competitors can outpace your software development. Lower cost per part.

I will take 2.

From working in this sector, it's highly unlikely the chip would work as a graphics card(well, kinda... I have heard of neural networks that can take in sets of vertices and colors and reproduce the graphics pipeline, but I don't think that's what you're getting at). It's unclear if it'd work particularly well for any network architecture radically different from what Tesla is using/expecting to use.

 

[Anstandswauwau]

If HW3 is so incredible... why not sell it... as a graphics card, or AI co processor. OEM it... see if your competitors can outpace your software development. Lower cost per part.

I will take 2.

It's a single-purpose board, that's why.

 

[defc0n]

If HW3 is so incredible... why not sell it... as a graphics card, or AI co processor. OEM it... see if your competitors can outpace your software development. Lower cost per part.

I will take 2.

Tesla believes they can make more revenue with it over time via AP than selling it outright to others and potentially giving competitors an edge in that space.

 

[EVDRVN]

I think this isn't very well understood by a lot of people: The compute requirements are affected by resolution and frame rate, but are not determined by them. I could design an algorithm that "processes" full 8k video at thousands of frames per second even on relatively meager hardware. That processing just wouldn't amount to very much.

The reason is Tesla uses 8K HDR and Dolby Vision hence the new HW3 requirement. All processing images are sent the screen for entertainment while you are on your ride. The HDR is the secret "safe sauce"

 

[j6Lpi429@3j]

yup, its an ASIC. it makes graphcis chips look general purpose (and they really aren't). Tesla havent invented new alien tech, the huge processing capability is partly due to the very fact that its hardware designed to just do a very specific thing in a very specific way.

 

[Sparky]

I’m a big fan of Lex but disappointed in his questions.

Apparently he had a bunch of them tee’d up about the hand-off between FSD and the human, but I would have thought he could have pivoted better when Elon basically told him that’s all moot, because FSD will surpass humans in 6-18 months.

I’m no AI expert, but it seems like one central issue is that each 9 in the “March of 9’s” becomes more and more difficult because of edge case explosion. I wanted more detail on how Tesla is handling this edge case explosion without it involving increasingly intensive amounts of human labor in labeling...

To take one example of Road hazard detection:

It would seem that the car has to classify every object in the road along a spectrum of how safe it is to hit, vs. how imperative it is to avoid.

To get to 99%, it may be enough to identify pedestrians, wheels, plastic bags, roadkill and potholes.
To get to 99.99%, there may be hundreds of different road hazards, going up to 100’s of thousands as you march to 99.999%.

Am I missing something. Does someone know how Tesla can handle edge case explosion?

Maybe, but I wouldn't have expect such from Fridman. I gather he's more of human-computer AI researcher. From MIT page: “Human-Centered Artificial Intelligence,” which is defined by two goals: (1) the AI system must continually improve by learning from humans while (2) creating an effective and fulfilling human-robot interaction experience. "

The nuts and bolts of autonomous driving equations may be outside of his wheelhouse.

 

[bdy0627]

You are picking and choosing what you want to take at face value then. That same article also specifies a "360,000 to 400,000 vehicle" forecast.

If you assume all of those are model 3's them that's upwards of 28GWh. If 25% of those are 100KW pack S/X vehicles then that's 32.5GWh

This ignores cells needed for Tesla Energy, and to ramp production for Semi, Y, etc...

It also doesn't mean that just because the lines are yet running at max capacity. It's also understood that Tesla/Elon aslo speaks of rates often.

The sales guidance is for all vehicles. GF batteries are for model 3 only.

 

[Remus]

I'm puzzled by this as well. TSLA wanting to lease well optioned vehicles vs selling SR+ makes a lot of sense. If SR+ demand isn't out the door.. well we have a much much larger problem. Also, forcing SR+ buyers into paying for AP (even at a discount) can't be too bad.

I guess they want more demand to create bigger batches with the same configuration to avoid production slowdowns.

 

[SOULPEDL]

yup, its an ASIC. it makes graphcis chips look general purpose (and they really aren't). Tesla havent invented new alien tech, the huge processing capability is partly due to the very fact that its hardware designed to just do a very specific thing in a very specific way.

Not far off from the expert human in that sense. So there could also be some ai doctor hw or ai social expert if we project that thought a bit.