Musk endeavours

[odysseus2000]

The structure of Elon Musk's incentive package means that he will shortly receive shares via performance options and will at that point need some cash to pay the tax on these. With this background it was almost inevitable that he would sell some shares to generate the cash and his Twitter poll now looks like a brilliant piece of stock marketing. More details in the first approx 1/4 of this video:

https://youtu.be/LBlhwzfY52E

Regards,

[BobbyD]

The structure of Elon Musk's incentive package means that he will shortly receive shares via performance options and will at that point need some cash to pay the tax on these. With this background it was almost inevitable that he would sell some shares to generate the cash and his Twitter poll now looks like a brilliant piece of stock marketing.

That background was contained in the original post on the subject. Musk is generally quite happy to borrow against his holdings to get hold of cash, would this mark a change in his attitude to the long term returns of borrowing to continue holding, or a change in the attitude of those offering a secured loan against the holding?

[odysseus2000]

The structure of Elon Musk's incentive package means that he will shortly receive shares via performance options and will at that point need some cash to pay the tax on these. With this background it was almost inevitable that he would sell some shares to generate the cash and his Twitter poll now looks like a brilliant piece of stock marketing.

That background was contained in the original post on the subject. Musk is generally quite happy to borrow against his holdings to get hold of cash, would this mark a change in his attitude to the long term returns of borrowing to continue holding, or a change in the attitude of those offering a secured loan against the holding?

One the options vest he will have a tax bill and likely is not keen to borrow more to cover this when he can pre-sell some shares now, generate the cash, put off the tax on these sales to April and then have enough cash to cover the tax from the vesting of his options.

It is not clear which 10% he is talking about. Is it 10% of current stock or 10% after the latest options vest?

Regards,

[BobbyD]

One the options vest he will have a tax bill and likely is not keen to borrow more to cover this...

So why is it different this time?

It is not clear which 10% he is talking about. Is it 10% of current stock or 10% after the latest options vest?

It's crystal clear.

Much is made lately of unrealized gains being a means of tax avoidance, so I propose selling 10% of my Tesla stock.

- https://www.theguardian.com/technology/ ... sla-shares

Uncaveated '10% of my Tesla stock'.

[odysseus2000]

BobbyD
Uncaveated '10% of my Tesla stock'.

Yes, but he does not say when and there is the ambiguity.

Regards,

[BobbyD]

Tesla (TSLA) stock plummets, is Elon Musk starting to sell? His brother got $100 million out

- https://electrek.co/2021/11/09/tesla-ts ... llion-out/

[odysseus2000]

Tesla (TSLA) stock plummets, is Elon Musk starting to sell? His brother got $100 million out

- https://electrek.co/2021/11/09/tesla-ts ... llion-out/

Interesting day for Tesla equity with volume of 59 millions shares compared to the average of 22 million.

Does Musk have something up his sleeve to arrest the stock decline or will we see further profit taking?

I did nothing although I sold out of several other positions to raise some cash. Clearly I was again too influenced by future potential and too concerned about not being able to get my position back if the shares rallied, to sell Tesla and clearly this was a mistake, a repeat of what I did in January.

Whether this is a short term blip or the beginning of another hard winter for Tesla stock remains to be seen.

Volatility is ones friend if used wisely and today I was foolish, but so it goes.

Regards,

[odysseus2000]

Diess & VW internals on Tesla:

https://www.bloomberg.com/news/articles ... eaten-jobs

Regards,

[odysseus2000]

Ford CEO praises Tesla:

https://electrek.co/2021/11/08/ford-ceo ... ev-better/

Regards,

[onthemove]

Tesla FSD beta 10.4 ...

https://youtu.be/EF1fWFyG238?t=596

"Oh my gosh! ... this is terrible ... This is probably one of the worst drives I've had in a long time"

"It was sooo good before, this exact same thing, it was so good...and now you get this and it's messing up all over the place..."

At one point, the Tesla is mistaking the car in front's rear lights for a traffic cone (shown in the visualisation) and trying to go around it.

Too many issues to list them all.

The guy even comments that he hopes these mess ups are deliberate, that (he hopes) Tesla is deliberately tweaking the parameters to cause more interventions to get more data.

My view : that's wishful thinking.

What I think we are seeing is how fragile the technology stack is. This isn't traditional programming. Each new 'enhancement' is typically not going to the addition of a new neural network, but further training of the existing one. And the functionality in a neural network is effectively spread across the network - so any further training runs the risk of undoing previous functionality.

And I think a few of the youtube Tesla owners (not evidenced in the above video, I'm referring to others here) don't seem to realise the above. I've seen a number of the tesla videos where the owners have commented that they have reported the same issues over and over again and that they just don't seem to be getting addressed. One or two clearly there was a hint of disillusionment in their commentary.

But I don't think they realise that when they report something, it's not like a regular software bug report where a software engineer investigates the code and tries to find the line(s) of code causing the problem!

The same 'code' in a Tesla has to handle an almost infinite continuous spectrum of possibilities. And if you try to tweak it to work better in one situation, you really do run the risk of degrading it's behaviour in another situation that might have been working. And I make that sound like 'tweaking' is fairly trivial - it's not... as I say above, this isn't 'code'. There isn't a software engineer editing a line of code.

It's actually a fairly impenetrable network of 'weights' that have been established by an algorithm repeatedly presenting the training data, looking at the output, and feeding back the 'errors' back through the network to 'tweak' the weights... the idea being, that by applying a lot (I mean a LOT) of training data, and repeating the algorithm many MANY times over, the weights eventually, iteratively settle to a reasonably good 'function' to transform the input data into the desired outputs.

But it's a very fuzzy process. You can't say to the algorithm 'you got this wrong... in this situation you need to do X instead of Y'... it can't distinguish this situation... all you can do is run the training algorithm with new data and hope it biases the weights to account for the situation that it didn't get right, while still retaining the capability to still do the things it was previously getting right. And it will require much more than just a single intervention from a beta tester to get enough data to tweak it.

You can also get the problem of 'local minima' in ML training algorithms. In these situations it can settle into a not-so-ideal situation but where it will take a substantial change to shift the collection of weights to achieve a better solution. Think of it like bad habits in people... once you've got into a bad habit, it can take a lot of effort to break your behaviour out of that bad habit.

While people can reason about their behaviour and be selective about which habit to change, it's not quite that simple for machine learning. To break an existing network out of a local minima / bad habit, it could take quite a push of training data to cover the situations that it's not handling very well. But then when you hit it with such a forced amount of training data, you really do up the risk of breaking existing good behaviour.

As an aside, this Waymo video (that I've linked to before) shows the lengths Waymo are going in order to train the system to handle the sorts of levels level of variation... https://youtu.be/oJ96bgmSaW0?list=PLcvM ... dJU&t=2177 ... it's showing how they've setup their simulation environment to recreate the camera scenes in different conditions and different times of day, etc.

That illustrates the lengths that Waymo are going to get the training data, and more importantly how they are going about making it robust.

I suspect this is why the Waymo videos that JJ Ricks used to post, showed a very stable and settled behaviour for the Waymo vehicles - the self driving system has been trained on a vast array of variations of each situation - exactly the same scene (road users, etc) but in different lighting, fog, rain, sunshine, and different shadows indicating different times of day... but all with exactly the same layout of vehicles, to really give the networks a HUGE amount of breadth in their training. Really making sure that the network gets to realise that it's the layout of cars that matter, and the lighting around them is not relevant, etc.

When you consider what Waymo is doing there, and then look at the Tesla videos where the guys press their "Report" button and say "Hey I'll report that so that they can fix it"...

Yeh guys, get real! Tesla aren't going to be treating your intervention report as a bug report in the traditional sense and tweaking their algorithm to suit.

No, in reality they're just going to be using you guys, the FSD beta testers, as a "test set", to evaluate how it's performing in a numerical sense -- interventions per mile driven, etc.

They might review them to see the *kinds* of situations where it's particularly struggling to figure out where they need to beef up their training data. And they might integrate some of the data directly into their unit tests. But they certainly won't be making individual fixes based substantially on each individual intervention reported to them by the FSD beta testers.

But, yeh, I think the penny is starting to drop with a few of the Tesla channels, that a full release of FSD is not just around the corner. It isn't anywhere near as close to ready as they were hoping, and they're slowly starting to realise that.

In fact, one of the Tesla channels even ran a poll asking the question along the lines of ... 'if Tesla offered you a refund on FSD given where it is at the moment, would you take it?'. I think more and more owners are starting to realise that they're not likely to see that hoped for - and paid for - software update for quite a while yet - will they even still own their Tesla by that time!

[odysseus2000]

Bloomberg say Musk has sold his shares:

Elon Musk Sells Tesla Stock After Twitter Vote To Meet Taxes - Bloomberg
https://apple.news/A2gG88Y3aR6-kVMKvAiXomg

Regards,

[odysseus2000]

onthemove
I suspect this is why the Waymo videos that JJ Ricks used to post, showed a very stable and settled behaviour for the Waymo vehicles - the self driving system has been trained on a vast array of variations of each situation - exactly the same scene (road users, etc) but in different lighting, fog, rain, sunshine, and different shadows indicating different times of day... but all with exactly the same layout of vehicles, to really give the networks a HUGE amount of breadth in their training. Really making sure that the network gets to realise that it's the layout of cars that matter, and the lighting around them is not relevant, etc.

I can see how this will make the system more robust for various lighting and weather conditions, but it seems too narrow to be of much use other than in geofenced locations where the road, (save for road works, accidents etc) does not change, only the users of that road change.

One of the recent developments by Tesla is to create a new programming language where variable length is configured to suit the situation. This is covered e.g. in this youtube: https://youtu.be/UWqTylXZ4AE The rational for these configurable variables is to cover a much larger range of values via the exponent but with a less accurate mantissa. I am assuming that this is being done to have the weights that you refer to more conveniently represented with less use of memory and faster processing.

Simplifying the process, as I understand it, the software is creating a weighted mean of all data (Camera plus car plus weather etc) from which it determines what to do. The biggest weight going to "don't crash" with the weights of other less important things scaled by the exponent with a less accurate mantissa.

Assuming I have this right then I imagine that the majority of the tweaks are first to the exponent to get a rough overall mapping and then to the mantissa to move things around. With the configurable variables memory usage is reduced to only what is needed giving potentially faster response.

I may have this wrong, but as I now see things, the tweaks become mostly mantissa ones to increase or decrease the weights once the crude exponents are settled by the cars not doing dangerous things.

This increased efficiency does not prevent the latest tweak making good behaviour go bad and that will still have to be addressed by tweaks, simulation and then real driving to see what happens in practice.

It is easy to get pessimistic here and the stack may not be good enough, but I am not sure we are at a point where one can say with very high confidence that the Tesla system will not work or will work.

Interesting times!

Regards,

[odysseus2000]

I do not know how accurate this is regarding Elon's selling but it seems plausible:

https://twitter.com/garyblack00/status/ ... 32869?s=20

Regards,

[BobbyD]

* ID only, VW sold 22 fully electric models in China last year.
**https://autonews.gasgoo.com/china_news/70019039.html
*** https://www.reuters.com/business/autos- ... 021-10-12/
****https://electrek.co/2021/11/08/tesla-maintains-extremely-strong-electric-vehicle-output-china/

Despite this high level of export from China, originally described as being for Chinese domestic market only, wait times for 3 and Y remain unchanged.

Together with non-ID. models, the total Volkswagen New Energy Vehicle sales amounted to 16,138.

For reference, Tesla sold in October about 13,725 units (wholesale shipments)...

- https://insideevs.com/news/546810/china ... tober2021/

[BobbyD]

also:

#VW China #NEV Oct retail sales: 16,138, 72% of Sino-foreign JV sales.
(CPCA)

- https://twitter.com/DKurac/status/1457639567767445504

So It's basically the Chinese, VW & Tesla...

[BobbyD]

Tesla MIC Oct retail sales
Model Y: 13,303
Model 3: 422
(CPCA)

- https://twitter.com/DKurac/status/1458005420107128834

[TUK020]

Tesla FSD beta 10.4 ...

But, yeh, I think the penny is starting to drop with a few of the Tesla channels, that a full release of FSD is not just around the corner. It isn't anywhere near as close to ready as they were hoping, and they're slowly starting to realise that.

In fact, one of the Tesla channels even ran a poll asking the question along the lines of ... 'if Tesla offered you a refund on FSD given where it is at the moment, would you take it?'. I think more and more owners are starting to realise that they're not likely to see that hoped for - and paid for - software update for quite a while yet - will they even still own their Tesla by that time!

Industry sources say that Tesla are planning to offer customers who have paid for FSD an opportunity to upgrade to the Ultra Premium FCD, which is a repeat subscription service model. This would neatly side step issues seen with FSD, and also lead to a re-rating of Tesla stock, to account for the ongoing margin potential for Tesla from this service.
Industry sources also say that preparation for this new ultra premium subscription service is behind the shortage of Romanian truck drivers.
Those close to the plans indicate that Full Chauffeur Drive offers innovative premium featuress such as holding the umbrella while opening the door.

[onthemove]

onthemove
I suspect this is why the Waymo videos that JJ Ricks used to post, showed a very stable and settled behaviour for the Waymo vehicles - the self driving system has been trained on a vast array of variations of each situation - exactly the same scene (road users, etc) but in different lighting, fog, rain, sunshine, and different shadows indicating different times of day... but all with exactly the same layout of vehicles, to really give the networks a HUGE amount of breadth in their training. Really making sure that the network gets to realise that it's the layout of cars that matter, and the lighting around them is not relevant, etc.

I can see how this will make the system more robust for various lighting and weather conditions, but it seems too narrow to be of much use other than in geofenced locations where the road, (save for road works, accidents etc) does not change, only the users of that road change.

Heck no! Definitely not. NO!!

The difficulty with neural networks is getting them to assimilate the aspects that are important.

At university we were presented with the example of a team who wanted to train a neural network to detect tanks (on the battle field). They went out and dotted tanks around everywhere and took photos. Then got rid of the tanks and took photos. Then went back and trained the neural network to categorise photos with tanks and without tanks, in order to act as a tank detection network.

They thought they'd done brilliantly. They'd split the data they'd gathered into training and test sets, and when their resultant network trained on the training set was then tested on the 'unseen' test set (standard practice), it did brilliantly. They thought they'd got it working (generalising) well.

So, pleased with their results, they went out, dotted some tanks around again and took more photos to do more testing.

The network performed absolutely dreadfully.

What went wrong? Well, it turns out that between then taking the photos with the tanks, and taking the photos without tanks, the weather had changed.

The neural network had simply learned to recognise what the weather was doing. It hadn't assimilated any capability at all to do with recognising tanks.

So to bring it back to Waymo... what they describe in the video about testing the same scene in multiple conditions (fog, rain, sun, lighting at different times of day, etc) is ABSOLUTELY NOTHING to do with perfecting it on a geofenced area.

It is about making sure that the network(s) are able to extract - and generalise from - the pertinent training information related to driving.

And to be clear, if you watch the video, Waymo are doing this with completely artificial scenes as well. Even totally made up cityscapes.

They've even used machine learning to automatically create a large number of completely artificial cityscapes and roads, which they then completely artificially add vehicles to, and give those vehicles (and other road users) behaviours based upon neural networks they've already developed to learn the behaviour for different entities using the roads.

They are then applying those same techniques that I mention above - about putting the scene in different lighting, etc - and generating almost photo realistic images (as shown in the video that a human wouldn't notice at first that they aren't real) and using all of those to help train the network.

Their training data goes MASSIVELY beyond any geofenced area, beyond even real world roads, and includes completely artificial layouts, all to help ensure that the networks are good at GENERALISING across a very wide range of scenarios.

I mean, really, if you think Waymo are using training data from their geofenced area as a substantial input into their ML training in order to get good in a specific area, then you've got it very wrong. Very, very wrong.

They are certainly not creating a self driving system that drives well in Phoenix Arizona, but only in Phoenix Arizona!

I mean, their choice of Phoenix Arizona to test driverless taxis, was taken a long way down the road of their self driving development. It was more of a "hey guys, I think this is ready for the real world, where do you think we can safely test if first?". ... rather than a "lets develop a self driving car for Phoenix Arizona first, and let's worry about the rest later!"

That is just where they've chosen to real world test driving without a safety driver. Waymo have hundreds of other cars in other places being driven extensively WITH a safety driver - just like Teslas!

Really, seriously, if people think that Waymo's geofencing in Phoenix Arizona is somehow because they've only developed self driving to work there, seriously, that is a HUGE misunderstanding of what they are doing.

Though if that's what you think Waymo are doing, I can see why you'd think Tesla are ahead.

But sadly, no, that isn't what Waymo are doing. By a long shot. It would be incredibly naïve and stupid if Waymo were training their neural networks on a Phoenix Arizona biased data set. Absolutely ridiculously naïve.

Let's get real here, Google are one of the (probably THE) world leader in artificial intelligence and machine learning. They sure as hell aren't going to develop self driving machine learning algorithms on training sets that are biased towards one city, in one state, in one country!

Similarly, I would hope and expect (as eluded to in my previous message) that Tesla aren't going to be biasing their training data to the feedback they're getting from their small number of beta testers who tend to be driving the same roads over and over again. This is why the Tesla videos a reporting that they keep reporting things to Tesla but they aren't getting fixed. Well, yeh, dur, Tesla aren't making their self driving specific to a particular set of roads being driven by one or two people either! If a Tesla FSD beta tester reports a problem on (e.g.) a specific junction, Tesla aren't likely to dump that data into the training set 'as is' job done, otherwise Tesla would end up being good in the areas where the beta testers are, and then perform poorly elsewhere!

(BTW - thanks for the link ... I'm supposed to be working at the moment, so I'll watch it later, but definitely interested to see the technical details for Tesla as with any other FSD efforts ... I'm sceptical of what it's talking about when you say 'exponent and mantissa' ... sounds like it may be talking about low level details... some neural nets use a lower 'resolution' (less bits) floating point number representation for performance reasons, and that may be all it's referring to... floating point values are composed of an exponent and mantissa ... , but I'll watch the video before commenting further... )

[odysseus2000]

onthemove
Heck no! Definitely not. NO!!

Thank you for the correction.

The difficulty I have is why are these systems constrained to geofenced areas. If the training is fully generalised, surely they could test them anywhere? Or is this geofencing just about what they consider safety, not about what their software/hardware can do?

The use of simulated data is also done by Tesla.

Perhaps I am too focused on practicality and not on the underlying performance although this being commercial is not shared by the companies.

Regards,
`

[BobbyD]

Bloomberg say Musk has sold his shares:

Elon Musk Sells Tesla Stock After Twitter Vote To Meet Taxes - Bloomberg
https://apple.news/A2gG88Y3aR6-kVMKvAiXomg

Regards,

A fraction of the shares he committed ot sell:

Tesla (TSLA) confirms Elon Musk has sold roughly $5 billion worth of stocks so far, likely to keep going

- https://electrek.co/2021/11/11/tesla-ts ... eep-going/