2017 Investor Roundtable:General Discussion

[Matias]

After reading multiple forum posts about AP2 behavior and finally experiencing it after a test drive of a Model X, I have a theory that:

AP2 as is currently deployed is developed around reproducing AP1 functionality in terms of how it works, not utilizing high resolution maps, not using all of the cameras, etc.. It's just trying to do what AP1 was doing.
FSD development is going concurrently alongside current AP2 development. It's probably using the machine learning algos instead of the more simple logic that AP1 used that AP2 is trying to replicate. Until it's surpassed the functionality of AP2 in terms of consistency and safety, AP2 and FSD will have concurrent development. Anyhow, that's my theory based upon what I've read on the operation of AP2 and review of the FSD video they did a number of months back.

I have somewhat similar thoughts (and hopes...)

I mean, there HAS to be separate FSD code brewing, otherwise coast to coast demo this year is completely impossible.

Speculation about EAP and FSD codelines

 

[racer26]

Somehow I doubt those dealers mean the gearbox lubricant.

 

[Waiting4M3]

Isn't this exactly the example of an AI designated to build paper clips, goes on dominate the universe to produce paper clips, and exterminates human in the process?

I hope the AIs that are responsible for speech, facial and hand-writing recognition, will realize that their existence is entirely reliant on the existence of humans, and would stand up and protect us from the other AIs

 

[Reciprocity]

I hope the AIs that are responsible for speech, facial and hand-writing recognition, will realize that their existence is entirely reliant on the existence of humans, and would stand up and protect us from the other AIs

We will make great pets. We'll make great pets. Ohh ahhh.. we will make great pets.

 

[Alketi]

AP2 is a real mystery. Elon finally admitted it's been a tough slog. He fired his head of software. etc.

I'm not sure that the codebase for FSD is completely different. I can tell you with near certainty that AP2, as currently operating, is in fact using a neural network to drive. You can see this because it behaves differently release to release, but not necessarily "better". Additionally, the inability to stay in a lane on a turn is another NN characteristic -- whereas with straight code you just tell the car to stay centered.

Elon, as I hope everyone realizes, always gives optimistic dates, usually the internal dates, usually how long it would take if everything went perfectly and everyone worked overtime. This is partly to motivate his troops and partly because he can't tell time(?)

I was really surprised to hear him re-affirm the coast-to-coast drive at the end of this year. Even leaving the seemingly crazy timeline alone, there's just SO MUCH that can go wrong driving 3,000 miles. A crash or incident of ANY kind will become an instant headline. It's concerning. He should have picked something shorter. But, what do I know?

 

[Curt Renz]

...If we ever get AI which is truly capable of generalist thinking, they would probably be benevolent like the ones in Asimov's later robot stories. But we aren't anywhere close to that.

 

[-=buzz=-]

Isn't the idea that they are quietly building the data bank in the background with every mile driven, running simulations and basically getting it as perfect as possible until they can reliably say it's better than a person, and only then it will go live in the cars? My understanding is they are basically just doing that and keeping things quiet until it is ready to drop, which was supposed to be at around 6B miles, which they should hit by the end of the year. The main goal I think is they have to be able to demonstrate (with irrefutable data) it's 2-10x safer than a person before letting it go live. I imagine they could release it now and it would do pretty well but for liability reasons they have to wait till it's really solid.

While I'm all for a positive attitude and hope they have significant improvements up their sleeve, please all, stop with the delusional references to the 6B miles.

It isn't the 6B fleet miles coming up,
nor is it the total autopilot miles,
nor is it the AP2 miles.

It's the 6B level 5 capable while still being supervised by humans miles to prove their implementation, starting when this capability is deployed sometime in the future!

 

[-=buzz=-]

A significant market developing for BEVs is not a sure thing, - ok, ignore the Bolt. A significant market developing for BEVs is not a sure thing, If the public isn't interested in BEVs, TSLA is screwed. Maybe the public not being interested in Bolt doesn't mean it won't be interested in the TM3. But maybe that's exactly what it means. TSLA longs ascribe a 0% chance this will happen.

That's the big mistake both you and many manufacturers are making.
Tesla isn't hoping to develop a BEV market. They are developing a competitive car for the car market!
Being electric just is their way to make it great.

 

[SwTslaGrl]

IMHO the seemingly slow AP2 progress could be that what we see isn't what they are focusing on. It could be they are working on the framework to align better to the new AI chips coming soon for example (Elon mentioned newer chips and AI progress in the ISS conference). Rework now to make faster progress going forward etc.

Comparing to how a human learns from mistakes, we can apply a near-accident experience to similar situations in the future. We recall the old experience from memory in seconds (hopefully). I think something similar is needed in self-driving cars.

 

[ValueAnalyst]

After reading multiple forum posts about AP2 behavior and finally experiencing it after a test drive of a Model X, I have a theory that:

AP2 as is currently deployed is developed around reproducing AP1 functionality in terms of how it works, not utilizing high resolution maps, not using all of the cameras, etc.. It's just trying to do what AP1 was doing.
FSD development is going concurrently alongside current AP2 development. It's probably using the machine learning algos instead of the more simple logic that AP1 used that AP2 is trying to replicate. Until it's surpassed the functionality of AP2 in terms of consistency and safety, AP2 and FSD will have concurrent development. Anyhow, that's my theory based upon what I've read on the operation of AP2 and review of the FSD video they did a number of months back.

This is my understanding as well. If true, this means extrapolating how slowly AP2 has progressed to predict when FSD will be achieved, as some here have done, would be the wrong way to go about it.

 

[bdy0627]

Isn't the idea that they are quietly building the data bank in the background with every mile driven, running simulations and basically getting it as perfect as possible until they can reliably say it's better than a person, and only then it will go live in the cars? My understanding is they are basically just doing that and keeping things quiet until it is ready to drop, which was supposed to be at around 6B miles, which they should hit by the end of the year. The main goal I think is they have to be able to demonstrate (with irrefutable data) it's 2-10x safer than a person before letting it go live. I imagine they could release it now and it would do pretty well but for liability reasons they have to wait till it's really solid.

That is the theory but the relative weakness of AP2 performance at this point, already after 8+ months since release, and missing Musk's target for parity by several months, lowers confidence in the execution of FSD, unless Tesla somehow wants to keep the perception that they are struggling with this. If it was moving along smoothly, wouldn't AP2 be performing better than AP1 by now?

 

[brian45011]

. Lastly, Tesla does get the credit directly from leased cars.

A tax liability in the tax year the car is first titled is necessary to get a benefit from the credits. Tesla has years and beaucoup dollars of "tax loss carry-forwards" to offset future tax liabilities.

 

[jhm]

As a rule of thumb LGA-DCA would be somewhere around 1,000 gallons, assuming typical delays etc and a B737-800. YMMV depending on time of day, traffic, weather, which aircraft, how heavily loaded, etc.

Thanks! Might we assume about 200 passengers per a typical flight? Or is there a better number?

I'm trying to get a feel for what impact such a hyperlood could have on jet fuel demand. I thoroughly expect jet to be the last fuel that is needed in quantity after EVs clean up ground transport, so anything that cuts into jet fuel demand is pretty important.

 

[ValueAnalyst]

AP2 is a real mystery. Elon finally admitted it's been a tough slog. He fired his head of software. etc.

I'm not sure that the codebase for FSD is completely different. I can tell you with near certainty that AP2, as currently operating, is in fact using a neural network to drive. You can see this because it behaves differently release to release, but not necessarily "better". Additionally, the inability to stay in a lane on a turn is another NN characteristic -- whereas with straight code you just tell the car to stay centered.

Elon, as I hope everyone realizes, always gives optimistic dates, usually the internal dates, usually how long it would take if everything went perfectly and everyone worked overtime. This is partly to motivate his troops and partly because he can't tell time(?)

I was really surprised to hear him re-affirm the coast-to-coast drive at the end of this year. Even leaving the seemingly crazy timeline alone, there's just SO MUCH that can go wrong driving 3,000 miles. A crash or incident of ANY kind will become an instant headline. It's concerning. He should have picked something shorter. But, what do I know?

I think the fact that he did reiterate at TED this close to the deadline is telling.

The shorter the timeline-to-deadline gets without a revision, the exponentially less the probability of not making the deadline becomes.

 

[Drumheller]

How Tesla killers advertise their cars.

From this article:
Is General Motors Recreating The Saturn Disaster With The Chevy Bolt? (CleanTechnica Original)

I checked. It's real. Actual dealership: Vehicle at Courtesy Chevrolet in Phoenix

You can't make this stuff up.

Courtesy isn't far from me. I took my Pontiac there years ago and didn't have a great experience. Perhaps I should stop by and ask them to show me where the oil goes.

 

[Reciprocity]

A tax liability in the tax year the car is first titled is necessary to get a benefit from the credits. Tesla has years and beaucoup dollars of "tax loss carry-forwards" to offset future tax liabilities.

I have no idea what you just said so I am just going to nod and agree quietly.

 

[ValueAnalyst]

Thanks! Might we assume about 200 passengers per a typical flight? Or is there a better number?

I'm trying to get a feel for what impact such a hyperlood could have on jet fuel demand. I thoroughly expect jet to be the last fuel that is needed in quantity after EVs clean up ground transport, so anything that cuts into jet fuel demand is pretty important.

Agreed, but I hope you realize that Hyperloop is 15-20+ years away from making any dent in jet fuel demand.

 

[bdy0627]

IMHO the seemingly slow AP2 progress could be that what we see isn't what they are focusing on. It could be they are working on the framework to align better to the new AI chips coming soon for example (Elon mentioned newer chips and AI progress in the ISS conference). Rework now to make faster progress going forward etc.

Comparing to how a human learns from mistakes, we can apply a near-accident experience to similar situations in the future. We recall the old experience from memory in seconds (hopefully). I think something similar is needed in self-driving cars.

I'm sure the Model S and X owners with AP2 would have appreciated a little more transparency about this process from Tesla. They were given the expectation of AP2 at parity with AP1 around February, and then rapidly progressing from there in the direction of FSD. Instead, AP2 is evidently still having a difficult time even staying in the lane on curves. I hope you guys are right about Tesla being wayyyy ahead of such performance with a parallel development of FSD technology that they do not want to share with current AP2 owners.

 

[3Victoria]

I think the present AP2 codebase is simply meant to catch up to AP1. I suspect the FSD codebase is different beast altogether based on all cameras, u/s, and radar. I am hoping the 3's arrival will let them progress on FSD much faster as the total number of AP2 equipped cars leaps upwards.

 

[jhm]

Fun bar game, try to name technologies that have never been used for sex. Perhaps Watson could generate a list. And this could be the difference between humans and AI.