Elon Musk $41B offer for Twitter

[EddieSteadyGo]

soupdragon1 said:

I agree that there is enormous room for improvement via technology, AI etc. The only way is upwards as far as that's concerned.

Elon is talking out of both sides of his mouth though. Designing a new robotaxi? Why? He said last night that there are millions of robotaxi ready cars already in the Tesla customer fleet, and they can jump in and out of their ride hailing app, like a homeowner can jump in and out of Air BNB rentals. So why build a new generation robotaxi when, to paraphrase Elon, FSD software is the differentiator between Tesla and other car makers.

If FSD is the golden goose, why build the Goose a brand new nest, when its already got millions to choose from? If the new platform is going to be dedicated to the Robotaxi, but its too much of a capital risk for the M2, why risk that capital in the 1st place, just for Robotaxi, when all the existing Teslas are robotaxi ready anyway?

When you try and join the strategy dots, they don't make any sense. All I can smell is BS to be perfectly honest.

Reuters leaked the news of the shelving of the M2, share price dips hard, then immediately, we get the Robotaxi 8/8 announcement. Then investors still aren't happy about the vacuum of information, why no M2, so Elon fills the gaps in last night and says there is an M2, but not as we know it.

And that's before we even talk about hardware, and vision only. The minute the weather is unfavourable, there is no robotaxi, even with perfect software. Its a fundamental flaw, vision only. I guess that's why the 8/8 robotaxi makes sense - better hardware.....but....then he says there will be millions of robotaxis from the existing fleet on the roads today. So maybe Robotaxi isn't getting new hardware?? Its nonsensical BS. The dots don't join up in a logical manner. Its feels like seat of the pants whimsical BS.

I'm not trying to connect every dot together, because I realise there are big contradictions.

But I don't think you either have to believe everything Musk says or discount all of it.

For example, I think it is possible they 'solve' FSD on existing cars in the next 2 years, but only in certain weather conditions, not all. So maybe in heavy rain, or fog, or snow, it doesn't work. Maybe for the 'Cybercab' they add some kind of self-cleaning to the cameras and possibly some kind of heating element which fixes the all-weather capability.

I'm wondering do they now have the right architecture for FSD? And do they have sufficient computing power to process a gigantic quantity of native video clips? And do they have a sufficient fleet size to automatically feed them hundreds of millions of miles of training data? And if they do all of this, how easy is it for a competitor to match them? And what could be impact in terms of future profitability if they hold the solution for FSD?

That I think is the basis for the 'bull case'.

[Byker28i]

Not so subtle dig at Musk saying it's not the role of social media to control misinformation.

https://www.washingtonpost.com/world/2024/04/18/mu...

From

https://theshovel.com.au/

[mko9]

Byker28i said:

Not so subtle dig at Musk saying it's not the role of social media to control misinformation.

https://www.washingtonpost.com/world/2024/04/18/mu...

From

https://theshovel.com.au/

[Img]https://i.imgur.com/ly4sxRd.jpeg[/thumb]

The Shovel brilliantly pointing that Elon Musk is correct.

Is that misinformation, or is that satire? Who gets to decide?

[EddieSteadyGo]

soupdragon1 said:

...
I agree that there is enormous room for improvement via technology, AI etc. The only way is upwards as far as that's concerned.
.....

You might be interested in this comment from Fred Lambert. Fred has been a big critic of FSD and has expressed similar views to yourself. He now says, if he sees some further improvements in v12, he could be prepared to become a 'true believer', which is an interesting volte-face.

https://twitter.com/FredericLambert/status/1783251...

[tangerine_sedge]

EddieSteadyGo said:

soupdragon1 said:

...
I agree that there is enormous room for improvement via technology, AI etc. The only way is upwards as far as that's concerned.
.....

You might be interested in this comment from Fred Lambert. Fred has been a big critic of FSD and has expressed similar views to yourself. He now says, if he sees some further improvements in v12, he could be prepared to become a 'true believer', which is an interesting volte-face.

https://twitter.com/FredericLambert/status/1783251...

I think FSD will follow the 80/20 rule. Tesla will achieve 80% of the functionality in 20% of the time and effort. The remaining 20% will be the difficult bit which takes the remaining 80% of the time and budget. FSD will have to meet a higher bar than the relatively low standard of human driving, so although it's impressively come a long way, it's still got a long way to go.

[dimots]

tangerine_sedge said:

I think FSD will follow the 80/20 rule. Tesla will achieve 80% of the functionality in 20% of the time and effort. The remaining 20% will be the difficult bit which takes the remaining 80% of the time and budget. FSD will have to meet a higher bar than the relatively low standard of human driving, so although it's impressively come a long way, it's still got a long way to go.

But how do you know how far along the curve they are? Perhaps they were 80% of the way there two years ago?

[EddieSteadyGo]

dimots said:

But how do you know how far along the curve they are? Perhaps they were 80% of the way there two years ago?

I have checked fairly regularly on the progress of "FSD" over the last couple of years. Despite promises to the contrary from Musk, apart from improved visualisations and a couple of basic features, it barely improved, and it always had critical errors. Every new edge case needed lots of code to write exception handling. Seems like they were creating an increasingly complex spaghetti of overlapping logic rules, which just became impossible to maintain. I think that is why rate of improvement was so slow.

Then they switched to this new design (v12), removed nearly all the code, replaced with a neural net, trained on short video clips from the cars. That version has only been training for a few months and it has already far surpassed what the old version could achieve.

And I'm thinking, how many miles does it take an average human to get good at driving i.e. to see enough edge-cases so they know what to do? Let's assume something like 50,000 miles. Tesla are collecting 10 million miles of training data every day. And they now have the computing power to process those video training into the neural net. Of course, humans use far more of their general cognitive ability when driving, so there isn't a direct comparison in terms of 'miles driven' but when you are collecting millions of miles per day, you are quickly going to collect a lot of edge cases for training.

So I'm thinking they are going to have 'solved' FSD on the current cars with the current cameras, assuming good weather, within 2 years. I think the cybercab will need an upgrade in hardware to be fully automatous so that it works in poor weather etc.

In terms of what this means, I don't think this just means they become a competitor to Uber or Lyft. I think within a decade it will start to be changing everything we know about personal transportation, including whether most people even need to bother owning a car.

[mko9]

Personally, I don't think FSD will ever work unless/until every car on the road is doing it. There are an infinite number of combinations of vehicles, road conditions, weather conditions, and most importantly human reactions to all different those factors. There is no way you can program an AI to compensate for all of them to respond promptly and safely every time.

[tangerine_sedge]

dimots said:

tangerine_sedge said:

I think FSD will follow the 80/20 rule. Tesla will achieve 80% of the functionality in 20% of the time and effort. The remaining 20% will be the difficult bit which takes the remaining 80% of the time and budget. FSD will have to meet a higher bar than the relatively low standard of human driving, so although it's impressively come a long way, it's still got a long way to go.

But how do you know how far along the curve they are? Perhaps they were 80% of the way there two years ago?

Let's say that they were at 80% 2 years ago.
That means that they've taken ~8 years to get 80% of the way there (they started the easy stuff way back in 2014), but the last 20% is significantly harder (and takes 80% of the time).
If 8 years is 20% of the time, then the remaining 80% of the time is 32 years. I look forward to FSD being ready in 2056...

The 80/20 rule is a general rule of thumb which seems (in my experience) to be a reliable guestimate in the software world. Teslas super-duper AI might speed up some of that learning, but it's still software relying on things it's seen before to predict how it should react.

At some point, Tesla will declare it good enough, they might even convice some authorities to accept it, but the moment that a FSD autonomous vehicle kills someone there will be a metric st-storm of legal wrangles.

Edited to add : I'm impressed by the videos of Tesla running v12, but it's a million miles away from being road-ready - if anything it looks more stressful trying to predict what stupid things it might do, than just driving the damn car yourself.

Edited by tangerine_sedge on Friday 26th April 10:44

[EddieSteadyGo]

tangerine_sedge said:

....
At some point, Tesla will declare it good enough, they might even convice some authorities to accept it, but the moment that a FSD autonomous vehicle kills someone there will be a metric st-storm of legal wrangles.

I think Musk might lean into some of the politics of this when they announce the 'Cybercab' in August. It's not beyond the realm of possibility that they get someone like DeSantis on stage, and announce some kind of state-level partnership for the launch. I'm not saying that will happen, but I think Musk could seek to use political rivalries to exert leverage onto the regulator.

As you say, when FSD is directly involved in a death, it is going to get very political, very quickly.

[shakotan]

EddieSteadyGo said:

I think Musk might lean into some of the politics of this when they announce the 'Cybercab' in August.

Is this the same CyberCab (albeit named RoboTaxi) that Musk announced in 2016 and promised buyers of Teslas they could use their cars as RoboTaxis and make the purchase price back on their investment through cab earnings in as little as a year?

Then after 6 years of it 'coming soon' with lots of eager Tesla owners waiting patiently for their investment to start earning them money, the idea was completely scrapped and a brand new vehicle was going to be developed instead.

That CyberCab?

https://www.bloomberg.com/news/articles/2022-12-07...

https://www.bloomberg.com/news/newsletters/2024-04...

https://www.forbes.com/sites/bradtempleton/2019/04...

https://electrek.co/2022/05/17/tesla-1-million-rob...

[dimots]

tangerine_sedge said:

Let's say that they were at 80% 2 years ago.
That means that they've taken ~8 years to get 80% of the way there (they started the easy stuff way back in 2014), but the last 20% is significantly harder (and takes 80% of the time).
If 8 years is 20% of the time, then the remaining 80% of the time is 32 years. I look forward to FSD being ready in 2056...

Well that doesn't make any sense at all, I suspect you are applying this incorrectly

[EddieSteadyGo]

shakotan said:

Is this the same CyberCab (albeit named RoboTaxi) that Musk announced in 2016 and promised buyers of Teslas they could use their cars as RoboTaxis and make the purchase price back on their investment through cab earnings in as little as a year?

Then after 6 years of it 'coming soon' with lots of eager Tesla owners waiting patiently for their investment to start earning them money, the idea was completely scrapped and a brand new vehicle was going to be developed instead.

That CyberCab?

https://www.bloomberg.com/news/articles/2022-12-07...

https://www.bloomberg.com/news/newsletters/2024-04...

https://www.forbes.com/sites/bradtempleton/2019/04...

https://electrek.co/2022/05/17/tesla-1-million-rob...

Yep, that's the one. Although, at that time, I would argue any objective analysis could see it was years away from being delivered. I think, objectively, this time it is different.

[soupdragon1]

tangerine_sedge said:

FSD will have to meet a higher bar than the relatively low standard of human driving, so although it's impressively come a long way, it's still got a long way to go.

The chat from Tesla bulls is all about FSD needs to be safer than a human, looking at road deaths per mile travelled and FSD needs to be better than that. I don't think they understand how low a bar that is.

To make it anywhere near a comparison of human vs computer, there is a load of human accidents that need removed from the calculation. FSD won't stop all accidents. Eg, a tractor or a lorry spills its load on the motorway - complete havoc. Neither a human or FSD has much of a chance. Drunk drivers - they already decide they don't want to pay for a taxi, so drunk drivers will remain on the roads. People on mobile phones, people hitting a patch of dirt on a farm exit road, people wilfully driving dangerously, people putting their makeup on while driving.....the list goes on and on.

So you need to remove ALL those accidents, and you're left with a more 'true' comparison. FSD vs a good human driver. Unless they ban cars and the roads are robots only, that's the actual bar that needs achieved. Releasing FSD into the wild won't stop people drink driving, it won't stop farmers strapping a load incorrectly, it won't stop the lady in her 2 ton range rover applying her make up, so all those accidents will remain - FSD doesn't fix them.

Nobody seems to grasp this concept of safety improvement. Its not about being better than the average. Anyone with a grasp of statistics will tell you that FSD being a little bit better than the average accident rate will result in a net increase in overall accidents. And despite that, you're still only getting into the ballpark of a true comparison. Accident statistics aren't a case of yes/no. Its not binary. Its a combination of incidents and severity.

Say you had 2 employers, and employer A had 5 safety incidents in a year, and employer B had 20. People automatically assume that the employer A is the safest, as it only had 5 safety related incidents. But what if employer A had 1 x fatality, 3 x broken limbs and 1 x minor injury, vs employer B who had 15 paper cuts, 3 x fingers trapped in the photocopier and 2 x falls on a slippy floor, all minor injuries. Suddenly employer B is safer.

Same with road incidents. Frequency and severity combined is how you measure this stuff. The FSD fans, and even Elon Musk himself, fail to grasp this concept of how to properly measure safety. Perhaps Tesla truly do understand it, I would expect a large organisation to be on top of this stuff, however, the way they present their safety data they either (a) don't understand it or (b) are being disingenuous and are trying to wipe peoples eye with misleading data.

I work in Finance as a business analyst these days but I've been IOSH and NEBOSH trained in years gone by. This is basic, mid level management safety stuff I'm talking about here, not like advisor to the board type stuff. Its basic analysis, yet its completely absent in the discussions about future safety on our roads. And from what I've seen of FSD, its literally miles away from the benchmark of a 'good driver', not the average driver, which is just a rediculous target to shoot for. Like I mention, if FSD becomes slightly better than the average driver, its a statistical certainty that it would increase accidents on the roads.

[tangerine_sedge]

dimots said:

tangerine_sedge said:

Let's say that they were at 80% 2 years ago.
That means that they've taken ~8 years to get 80% of the way there (they started the easy stuff way back in 2014), but the last 20% is significantly harder (and takes 80% of the time).
If 8 years is 20% of the time, then the remaining 80% of the time is 32 years. I look forward to FSD being ready in 2056...

Well that doesn't make any sense at all, I suspect you are applying this incorrectly

Using the power of paintshop to hopefully explain the 80/20 rule. Functional completeness is not linear, some use cases/requirements are significantly more difficult to achieve than other use cases/requirements :



The easy functionality can be done really quickly, so in this case you get a car that can drive along well lit roads, tackle junctions, roundabouts and mixed traffic. It looks like it's almost working and only needs a little more work to fix - hence the constant changing delivery dates and broken promises from Elno. It's not as close to completion as people think (that pesky last difficult 20%).

The hard functionality, driving in poor weather/lighting conditions, dealing with unexpected random stuff, and not injuring people is the stuff that takes considerably longer. Normally, there's a trade off when functionality is 'good enough', but in autonomous vehicles that trade off will be very high because peoples lives are potentially at stake.

[EddieSteadyGo]

soupdragon1 said:

....
The chat from Tesla bulls is all about FSD needs to be safer than a human, looking at road deaths per mile travelled and FSD needs to be better than that. I don't think they understand how low a bar that is.
...

I take your main point that Tesla are being somewhat evasive/non-transparent with their stats. They certainly prefer the ambiguity. But on the point I've highlighted, I think they are saying FSD needs to be not just better than the average driver but at least twice as good.

[soupdragon1]

EddieSteadyGo said:

soupdragon1 said:

....
The chat from Tesla bulls is all about FSD needs to be safer than a human, looking at road deaths per mile travelled and FSD needs to be better than that. I don't think they understand how low a bar that is.
...

I take your main point that Tesla are being somewhat evasive/non-transparent with their stats. They certainly prefer the ambiguity. But on the point I've highlighted, I think they are saying FSD needs to be not just better than the average driver but at least twice as good.

Do you think twice as good is good enough though?

[EddieSteadyGo]

soupdragon1 said:

Do you think twice as good is good enough though?

Hmm... I suspect the accurate answer would be 'it depends'. I do also think it is going to be impossible in the short term to measure it properly.

As you have highlighted, current FSD requires hyper-vigilance from the driver. So they won't be able to measure accident rates with FSD enabled (plus a driver watching attentively), and compare that to an 'average driver' and conclude what FSD would have achieved on its own.

But I think, if it could be measured accurately, and some metric like "deaths/km" for each type of road showed it was actually x2 better than average, I think that would probably be good enough, even taking into account the statistical issues you mentioned.

There is another point above trajectory of improvement - let's say it achieved a threshold of being x2 safer on urban roads in the next 24 months. I think, if it gets that far, one could assume that given another 24 months, it is going to get better again.

[soupdragon1]

EddieSteadyGo said:

soupdragon1 said:

Do you think twice as good is good enough though?

Hmm... I suspect the accurate answer would be 'it depends'. I do also think it is going to be impossible in the short term to measure it properly.

As you have highlighted, current FSD requires hyper-vigilance from the driver. So they won't be able to measure accident rates with FSD enabled (plus a driver watching attentively), and compare that to an 'average driver' and conclude what FSD would have achieved on its own.

But I think, if it could be measured accurately, and some metric like "deaths/km" for each type of road showed it was actually x2 better than average, I think that would probably be good enough, even taking into account the statistical issues you mentioned.

There is another point above trajectory of improvement - let's say it achieved a threshold of being x2 safer on urban roads in the next 24 months. I think, if it gets that far, one could assume that given another 24 months, it is going to get better again.

I put a quick excel sheet together to help explain my point.

I've used round numbers, 1m cars on the road, 1,000 deaths per 1m cars, 40% of deaths caused by good drivers, 60% caused by people being complete a-holes on the road. In reality, its probably more like 80 or 90% is due idiots but I've had to pick a number, to illustrate the point. Its worth noting that NHSTA in america estimates that 40% of road traffic incidents are caused by drivers with alcohol in their system. My underlying point is that no matter how good FSD becomes, it will have zero influence on a semi-truck driver getting into his truck after drinking alcohol. If you're in an FSD car and that truck swerves into you, the accident will happen regardless. This is a very important point. The lady doing in her makeup in her Range Rover - if she drives into you head on, FSD won't bring you back to life. That statistic will still happen and go on record.

So lets look at the stats. If FSD was twice as good as the average driver:

Total deaths with no FSD cars on the road: 1,000
Total deaths with 10% of cars on the road having FSD: 1,010

Conclusion:

FSD creates additional deaths, despite being twice as good as the 'average' driver. Unfortunately, that's how statistics work. There is a huge amount more nuance of course, but the below screenshot explains how the data is arrived at in this basic example.

Edited by soupdragon1 on Friday 26th April 15:41

[EddieSteadyGo]

soupdragon1 said:

I put a quick excel sheet together to help explain my point.
...

I think there is an error in your spreadsheet.

You have the "deaths avoidable" in "regular cars" being at a lower rate than the robotaxi car rate. Hence why it gives more deaths with robocars when it should be less.