1x Tsla Discussion Threads

[careful] 15/4/2024 19:02

Tesla are now just another car company as many predicted.
It will be a grind, ferocious competition.

[cfb2] 15/4/2024 18:52

Johnwise: The article you've posted is just garbage, it has confused Cybertruck with overall vehicles and misunderstood a production problem. There are mistakes in pretty much every paragraph, for example:

In Q1 2024, Tesla stated that it had only sold 386,000 Cybertruck units, despite producing over 430,000.

Cybertruck is just ramping up production, and Tesla are hoping to be producing 250k Cybertrucks by 2025. They have sufficient back orders (~2m) for the whole of 2024 and beyond.

There is always a difference between manufactured and delivery numbers as it takes time to deliver vehicles around the world.

Tesla overall car delivery numbers undershot analyst estimates of 425k with 386k. Sales in China have dropped off a cliff for all manufacturers (43% for BYD!) but only 20% for Tesla.

Production has stalled and slowed due to Tesla being unable to get parts due to problems transporting through Red Sea. Production at Gigafactory Berlin was down for almost a week due to sabotage of power to the plant.

[cfb2] 15/4/2024 18:27

I'm not sure a 10% lay-off should be seen as a bad thing by Tesla investors. This was announced a few months ago as a result of Tesla being between roll out waves and should be seen as an attempt to reduce costs and inefficiencies. Apparently a dead wood cull is a regular event there where teams need to fight to keep their people.

Andrej Karpathy talks about this at around 16:50 in this video:

[johnwise] 15/4/2024 18:19

Tesla lays off 10% of staff after falling Cybertruck sales

Tesla's Cybertruck sales fell short of expectations in the first quarter of 2024.

[juliemara] 15/4/2024 15:53

Amazing how many just can't see the wood for the trees!

[hpcg] 15/4/2024 13:43

They don't, which is why the plurality of analysts have Tesla as a hold, which of course means sell. Stockopedia data, which is from Refinitiv, has 4 strong sell, 7 sell, 22 hold, 11 buy and 5 strong buy. Their combined target comes out at $186, which is barely volatility above the price today. The consensus is gradually getting more negative.

[cfb2] 15/4/2024 12:08

Tesla are shortly (my guess is sometime in June/July 2024) going to start the approval process with NHTSA to get FSD software to level 4. This requires vast amount of data showing FSD safely driving over billions of miles. The more cars running the latest software the faster Tesla capture the evidence.

$99/month seems to be a sweet spot for people continuing with the FSD trial with around 25% of people prepared to pay this. I would guess the price reduction is to maximise the number of cars running the software rather than revenue.

Many analysts don't even include risk weighted revenue in their research for FSD software. Once Tesla have got to level 4 they will have to start modelling it as well as Robotaxis, unless they are incompetent.

[hpcg] 15/4/2024 09:59

Cost of FSD in the US has been halved. This presumably is a reaction to dropping demand, or possibly more likely, slow or negligible take up from new car buyers. I can't see the market liking that, though early doors there might be some buying of the spin. Analysts will have to get some red ink out, especially if they had factored in it only ever going up in price, per company announcements to this point.

[cfb2] 11/4/2024 17:18

Dominic is correct about the billions of captured miles from Tesla vehicles, which will be useful for getting approval from NHTSA and the various US states. Tesla seems to think 6 billion miles of safety data will be required to get approval.

For training their cars about 1 million multi-second clips of video data is required for basic operation but Tesla believe about 5 million clips are required for "god like" driving. Tesla carefully select and manicure their data before using it for training to enhance the required areas the previous version of software was lacking.

[cfb2] 11/4/2024 16:57

The compute hardware in a Tesla is not your standard microprocessor. Whilst it does have CPU cores it also has a GPU for running multiple processes simultaneously and a neural network that effectively computes dot products. Roughly speaking, the GPU runs an order of magnitude faster than the CPU and the neural network an order of magnitude faster than the GPU.

There are two complete units running in each car. This allows Tesla to run new versions of their software in "shadow mode" before they release it. Shadow mode runs on the second processor and allows them to see how their software performs compared to the current active version.

[cfb2] 11/4/2024 16:47

Pierre: The roll out model for Tesla self driving cars is different from Waymo. Waymo have launched self driving cars that earn money operating a taxi service but do not require a driver. Tesla's FSD software costs $12k or $214/month and currently requires a driver to take responsibility of the car, sitting in the driving seat, and intervene if necessary. This means that Tesla generates a revenue from their software, Waymo operate at a loss.

Waymo is restricted to 45mph, does not perform unprotected left hand turns and is geo-fenced (only drives in areas that Waymo have mapped to high definition for their Lidar sensors). A single traffic cone in the wrong place is enough to disable a Waymo taxi. They have a monitor centre which can intervene when a Waymo gets confused; recent figures show this occurs on average every 2.5 miles.

[dominiccummings] 11/4/2024 13:28

In any case, it is not necessary to be perfect, like any other product, it just has to be significantly safer than human drivers.

[pierre oreilly] 11/4/2024 12:33

Even 1 billion hours takes 115,000 years to run through. At least that illustrates the magnitude of the problem they face testing software. I'm not sure there are enough resources on the planet to save that amount of data. It looks like Tesla processes data from their fleet (whatever that means, possible several evs from workers in California?) in real time with advanced ai techniques to improve their autonomous driving software and fix problems as they happen. Phew, quite a claim! There may be an element of truth in that, but really it is 99.9% PR imv.

[dominiccummings] 11/4/2024 11:17

Sorry Pierre, but Tesla captures billions of hours of real world driving to run through its software in a simulation mode. On the road experiences of actual driving, mistakes and all.

[pierre oreilly] 11/4/2024 10:51

Jeeze, what's happened to advfn? I seem to be having an intelligent and respectful conversation - that doesn't happen often! Thanks!

Surely Tesla will have the same restrictions as Waymo, until and if they can be lifted? Do Teslas go 70 on normal roads without someone in the drivers' seat? (I have no idea but I wouldn't get in one if they did).

For neural networks, (like ai) read software. Again, I know you know it's software, but to many neural networks sounds like magic with perfect execution. Irealise the software has many inputs in an attempt to drive a car, but to me that just means more software in real time for the chips to handle. Add on talking to each other, and soon it becomes untestable in order to guarantee it works.

Of course there's regression testing to some degree. Not sure how in a multi cpu, multi degrees of freedom that is dome in general, only specific test scenarios can be set up (and automated). If you edit software, then of course you can check it still works for all those cases where it worked before (by regression testing for a specific scenario), but it doesn't test for cases yet to be discovered in real life. (Incidentally, aircraft systems are far easier and less complex than car system, nd there are still plenty of bugs making themselves known there, often covered up and blaming things on the pilots).

I think the best Musk can do is to claim, probably correctly, that self drive is safer than humans (in decent weather conditions and slower speeds) - no problem with that. The question is will that be good enough and will the law be changed to give Tesla a dispensation from being sued when people are killed? (in every country where it wants to operate its autonomous cars)

[cfb2] 10/4/2024 16:23

Pierre: I am discussing the possibility of Tesla autonomous cars, rather than Waymo. Tesla cars operate at 70mph and above. If they get approval for level 4 then this will be direct comparison to the equivalent humans - hence like for like.

Whilst most listed companies know their P/E ratio will go up if they waffle about how their software uses AI, for Tesla this is an accurate description. Their FSD software has a neural network with parameters fed from the cameras and outputs that drive the steering, pedals, display etc. of the car. The neural network is an array of interconnected weighted numerical values, tuned by a goal orientated training process. Even by older definitions of AI, this is mimicking the operation of the human brain by perceiving the environment around it and taking actions to achieve a goal.

I agree that complex software is difficult to test (I remember how difficult it was to produce mathematically provable computer software over 30 years ago and the language restrictions to do this, even for the simplest of programs). AI software has made testing orders of magnitude more complex, as you are now feeding "values" into a black box that is performing maths using the weighted numerical values and rippling them through a network to eventually drop out to control or display something. You can't possibly check all the values so you do the best you can by simulating the scenarios and confirming it works correctly. When you find a scenario it fails in you add that to the test sequence. Every time a new release of software is produced you perform regression testing by running through all the previously accumulated test scenarios.

[cfb2] 09/4/2024 14:45

Pierre: We already have statistical data for car fatalities in a year, as well as the number of miles travelled per accident. This can be compared to cars running FSD software. If FSD performs significantly better than a human then they should be allowed.

The "control group" to compare the AI software would be those who continue to drive their cars.

As you say, there is no guarantee the AI software will continue to improve and that is why you capture data at each accident, analyse it and react. Exactly the way aircraft accidents are analysed.

Unit testing to recreating accidents in simulators can prove that a fix has been correctly applied and no regression has taken place as new software is rolled out. If you've watched some of Tesla's technical presentations you'll know they already do this.

[pierre oreilly] 09/4/2024 12:40

Your Waymo statement seemed to me to not represent the actual situation (while being true). From a quick google into Waymos documentation, it looks like many of the first million miles on which the safety report was based, were with a driver in the car ready to take over (although some unspecified amount were RO, i.e. no driver.

Also, the cars are in 45mph limits (or at least never exceed 45mph) - another factor which may not or maybe, no idea) appreciated by the average joe (not you). Another factor - they only drive in decent conditions, and no driving 'in heavy rain' or fog (not great for San Francisco!) or adverse weather conditions.

So a long way to go (which will never be reached in the forseeable future in my view) to replace taxis (hich do go 45mph+, and do go in rain and fog and bad weather. There must be many more restrictions (perfectly correctly) which I haven't found yet.

It's software complexity. Many will think that software will improve over time so there are no accidents and they'll go in all conditions. But software (called ai these days) isn't like that. At a certain level of complexity, changes to fix or improve things always, imv, result in other problems caused by the fix.

[pierre oreilly] 09/4/2024 12:20

cfb2, Is your argument known as selective data (or rather selective imagination)?

Say there were 1 human accident and 8,000,000 robotaxi malfunction deaths. If you allowed rootaxis you'd be killing 7,999,999 people.

(Not an argument I use, just illustrating exactly your selective argument).

[hosede] 09/4/2024 11:30

Pierre
Agreed - The Robotaxi has to have FAR less accidents than a driver - but One thing that might work thru AI is a robotaxi with a remote human backup that could see everything the car could see and step in if it got stuck. The benefit being that the backup could maybe act for several taxis at once.
But there will be massive resistance from taxi drivers ( Remember " the man in the white suit") and it will take the public quite a while to come round to them

[cfb2] 09/4/2024 11:29

hpcg: AI is a difficult one to assess. Everyone is playing their cards close to their chest, especially "openAI" (misnomer if ever I heard one!). I think Telsa's AI is behind with respect to AGI but has advantages with respect to training data (large amounts of video from Tesla cars and twitter text that competitors don't have access to) and compute power (they have stated their FSD software is no longer compute restrained).

Boston Dynamic's Atlas has hydraulics rather than electric motors making it expensive ($1m+ for latest model, compared to Tesla's Gen2 with an estimated $80k) and power hungry. Battery life in Atlas is less than 1 hour compared to over 8 hours for Tesla Gen2.

The software in BD's robot is heuristics based, those demonstrations with back flips were carefully scripted movements, rather than neural networks.

I would describe BD's Atlas as a research project, not as something to be sold. Evidence for this is the number of times their company has changed hands; it is bought and then the purchaser doesn't know what to do with it.

There are serious competitors to Tesla's Optimus Gen2, such as OpenAI's Figure robot which will be used in BMW's factories.

At last count there are over 30 robotic companies looking for a niche. To Tesla's advantage it has huge compute power for training, leading the field with actuators and motors, has proven manufacturing skills and possibly cross speciation with their car AI. It is likely to be an extremely competitive field to be working in.

[dominiccummings] 09/4/2024 10:34

No. I was replying to hosede. Main issue being 'first mover' advantage and value.
However, your argument can be rebutted for other products, guns, cigarettes, noxious chemicals, pollution, and so on. Government steps in with Law, the Law determines the degree to which thew producer of user can be liable.