Tesla Talk

liamog

https://www.boards.ie/discussion/comment/121049920#Comment_121049920

If I said that UEFA were going to block camera based ADAS systems would you think the claim was ridiculous?

listermint

https://www.boards.ie/discussion/comment/121049925#Comment_121049925

If I said EU doesn't work with member states and their bodies to ensure there is a outcomes that suit the EU, would you be nieve enough to believe that doesn't occur?

Similarly if say the UK are putting forward arguments as part of a grouping that a pending trade deal with say China doesn't come into those conversations. There are various political and economic forces that come to play in these forms of framework talks. And I've no doubt you know this already. Tbh

liamog

https://www.boards.ie/discussion/comment/121049969#Comment_121049969

I assume by your avoidance of an answer that you understand why UEFA are not involved in the certification of camera based ADAS systems, i.e. they are not a supranational body with competency in the area. Just like the EU is not a supernational body that has a competency in the area.

I don't quite understand why you want to insist that their is an EU competency in the area when their isn't. There is a technical standards body which is working to regulate the operation of automated driving systems, you just have the wrong one.

How we got down to this particular piece of educating @listermint on where light vehicle regulations are set was your starting position that the EU would not allow camera based ADAS systems. What we can say is that the relevant regulatory body (WP.29 of the UNECE) is setting standards based on the requirements for systems that are used to deliver automated driving features. Whether Tesla's FSD system will meet these requirements is a whole other question but we can be certain that they will be required to meet a UNECE standard and not an EU standard.

As to your point re groupings within the working committee's, from the documents and minutes published, there is as much back and forth between regulatory bodies from signatory states that are also EU members as there are between those from regulatory bodies that are signatories but are not a member of the EU. I don't see any evidence to support your accusations of a shadowy EU influence on the proceedings.

cnocbui

https://www.boards.ie/discussion/comment/121047850#Comment_121047850

10K years of driving experience and they still regularly slam into the backs of police cars, fire tenders ambulances...

unkel

https://www.boards.ie/discussion/comment/121050232#Comment_121050232

Sigh. Restraining myself not to be tempted to feed the troll. V11 is beta software, coded. V12 is alfa software, no coding, machine learning. If you don't understand what the alfa stage of software development is, please google it.

listermint

https://www.boards.ie/discussion/comment/121050314#Comment_121050314

It's Alpha, and it's still quite poor. And not up to task for full self driving on public roads.

And if you think machine learning requires no coding. This is an absolute bizarre perception.

Bovakinn

Alpha comes before beta and is the very first stage of testing, so it makes sense that it's not perfect, whether or not it will ever be perfect is yet to be seen, but IMO the progress looks promising.

Saying that V12 has no coding is a bit of a misnomer, but is an easy way to talk about it. V11 uses ML outputs as input to human-coded algorithms which control the vehicle. V12 is designed so that the output of the ML network controls the vehicle directly. The workflow basically goes Sensor Input -> AI inference -> Vehicle Control with no coded steps in between the AI Inference and Vehicle Control. I don't think anyone is saying that ML requires no coding.

listermint

https://www.boards.ie/discussion/comment/121050824#Comment_121050824

Well he did say no coding. So I think I was correct to clarify.

Additionally clarification there is code involved within the AI portion of the analyse whether its tweaking the algorithms consumption of firmware modifications in Hardware output or tweaking the Machine learning barriers parameters or filters. There is always refinements made based on outputs. There are a large portion of folks that assume slapping the term AI on something means its sentient. Its not, the library's that it utilises to base assumptions on are frequently adjusted.

I firmly state that teslas model is not the correct one, and I personally think they jettisoned it because of cost and difficulty. They should adjust and reintroduce non video analysis as part of their platform sensor inputs. It's the most sensible approach. And yes it costs more.

Anyway it's all good discussion. I'm all for an FSD. Just the best possible one we can all get as drivers and pedestrians.

unkel

@listermint - "Well he did say no coding."

Ah come on, you know what I meant. In my earlier post I already said there was coding to hold the rest together. I stand by my point that the number of lines of coding in V12 is a fraction of what it was in V11 as the new version is a paradigm shift. Imagine the billions of dollars Tesla have spent on versions before V12 😮 and to throw all that out, just like that.

listermint

https://www.boards.ie/discussion/comment/121051449#Comment_121051449

No code v12 is incorrect. We shouldn't make such statements. Simple.

There's a white washing of teslas FSD that is continued to be propagated rather than serious discussion of its limitations. Great tech yes, limited also yes.

unkel

https://www.boards.ie/discussion/comment/121051541#Comment_121051541

I agree it is limited tech. I always have done.

Try look at it from a philosophical point of view. And not from a technological point of view. Once one gets their head around that there is no reason to get it perfect (also quite impossible to get it perfect) and that all we need is it to be many times safer than a human, isn't that all we need to strive for?

Better release it (once above can be proven) to start saving a few thousand lives per year almost immediately, rather than wait (possibly for decades) before it can save hundreds of thousands of lives per year, no?

The car working just as a human but with senses, reactions and experience far, far greater than a human, should do the trick nicely. FSD will cause accidents and it will kill people. There is no doubt about it. This is acceptable imho as long as it does so in far smaller numbers than humans do...

listermint

https://www.boards.ie/discussion/comment/121051628#Comment_121051628

There's no proof its safer than humans though.

And just like flight. We don't allow planes to not be next to perfection It's why the standards are incredibly high and have multiple redundancies are required.

We should not by any means drop the high bar we have for flight where autonomous cars are concerned.

The mantra of 'a sure its kinda better than what we have' isn't good enough and is finance driven not ethically driven. You have kids yourself you should want any autonomous vehicles to be the absolute best they can possibly be before they're allowed share our roads in any volume. Low bar acceptance is a purely money making view point and a race to the bottom in terms of wages suppression and conveying people from a to b via subscription services. $$$ we should view it in the same cynical view wallstreet is viewed.

Bovakinn

https://www.boards.ie/discussion/comment/121051405#Comment_121051405

There are always refinements and tweaks to be made, but they're rarely done by people in the ML space.

With deep learning techniques, you set the model architecture (shape and size, types of layers such as fully connected, convolutional, pooling, etc., and neuron activation function), learning rate, and the number of epochs. Once the training is complete you shouldn't manually change any of the parameters. If you're not happy with how a model performs you change the training parameters and train again, generally from scratch to prevent overfitting.

Here's a paper from Yan LeCun that outlines the benefits of Deep Learning, as opposed to hand-coding feature extractors, although it is behind a paywall.

https://www.nature.com/articles/nature14539

As for whether Tesla's approach is the right one or not, I think it remains to be seen. There's progress being made on several fronts, with a myriad of different sensor systems. Tesla's approach is definitely the cheapest, and quite possibly the hardest to develop for, but I think it's still too early to dismiss it just yet when their system is the only one, to my knowledge, that isn't geofenced to relatively small and controlled areas. I think people would be more forgiving of Tesla's systems if Elon didn't run his mouth and say that FSD would be here by the end of the year for the last 5 years or so.

KCross

if Elon didn't run his mouth and say that FSD would be here by the end of the year for the last 5 years or so.

+1. That's the real problem.

Let Tesla bang away at it as long as they want and test it with the proviso that the driver is always responsible (which in fairness they do say) and let their technological approach take them as far as it can, best of luck to Tesla.

But the incessant "its just around the corner" from Musk is ridiculous, but it has pumped the share price no end... so there's that! 😉

unkel

https://www.boards.ie/discussion/comment/121051992#Comment_121051992

It's naive to think Musk promising anything about FSD will bump the share price. Like you said, he has been promising this for years. Nobody pays attention to this really. But pretty much all his previous promises eventually became true. I do feel bad for people who have forked out for FSD in the past based on these promises and who have long since sold their cars, without being able to take their expensive option with them. They deserve compensation. I'm amazed this hasn't already led to class action lawsuits in the USA

@listermint - "You have kids yourself you should want any autonomous vehicles to be the absolute best they can possibly be "

No, I want them on the road asap as they will take over from humans who are inferior drivers (obviously once the software is ready for release and this can be proven with statistics). As I said, they will cause accidents and deaths, but to a far lesser extent than humans. The roads will get safer with every car that has FSD enabled. We will never reach the absolute best they can possibly be with a guaranteed zero accident rate. Or it will take a very long time. Anyone who wants that is indirectly responsible for the deaths that could have been prevented in the mean time

cnocbui

https://www.boards.ie/discussion/comment/121050314#Comment_121050314

Someone who thinks public safety endangering software should be literally out on the streets for testing and development ...

Alfa software - that's incredibly appropriate: Looks nice and flashy on the outside to those easily dazzled by surface appearances, but is prone to interminable faults the minute you drive it out of the showroom.

My son has a computer science degree and lives one bedroom away, I don't need google to tell me the problems with AI and neural network derived algorithms, he is quite sufficient.

'Hey boss, our BSD system keeps running over policemen, fireman and ambo's...'

'Well just fix the lines of code that are causing the problem...'

'But it's a black box...'

'Well get a cat and put it in the box, and maybe it will fix it, and maybe it won't.'

Aug 29 (Reuters) - The U.S. auto safety regulator investigating Tesla's Autopilot driver assistance system is demanding an explanation for a software change that allows drivers to keep their hands off the wheel for longer, risking collisions, documents released on Tuesday showed.

The National Highway Traffic Safety Administration (NHTSA) ordered Tesla to answer questions about the Autopilot change and produce documents in a so-called special order dated July 26. The NHTSA order did not compel Tesla to recall vehicles.

"The resulting relaxation of controls...could lead to greater driver inattention and failure of the driver to properly supervise Autopilot," NHTSA said in its letter to Tesla.

The NHTSA order underscores the challenges regulators face keeping up with safety-critical vehicle technology that can be changed overnight with a software upgrade delivered over the air...

In April, Chief Executive Elon Musk tweeted that Tesla is gradually reducing the steering wheel nag - alerts aimed at making sure drivers using Tesla's "Full Self-Driving" (FSD) system keep their hands on the wheel. He also this week said "Yeah" in response to an X message that "No steering wheel nag will be a *game changer* for FSD user satisfaction."

https://www.reuters.com/business/autos-transportation/nhtsa-issues-special-order-regarding-teslas-driver-monitoring-system-autopilot-2023-08-29/

listermint

https://www.boards.ie/discussion/comment/121051862#Comment_121051862

All of the changes you mentioned are hand coding feature extracts. Note your entire First paragraph is about changing configurations and 'shouldn't' adjust during modeling.

This space is fast developing but also myriad of smoke and mirrors. Chat GPT is a superb example of language library's and the capability of humans to modify the output based on trickery with of the queries being entered.

Papers are fantastic but the reality is they aren't doing anything from scratch they are adjusting the modeling daily based on thrown up scenarios. As with alot of 'machine learning' humans are controlling the inputs.

Teslas is blatantly the cheapest worst approach, why ? You are reliant on a single technology input when there are many available today to enhance and increase its accuracy. But they choose not to.

KCross

https://www.boards.ie/discussion/comment/121052060#Comment_121052060

It's naive to think Musk promising anything about FSD will bump the share price.

I think you might be the naive one there.

FSD goes to the heart of the share price because its insanely high valuation is based on that future potential FSD income. Its not entirely based on Tesla being the number 1 car producer in the world. If the share price was only based on the future car sales it wouldn't be as high as it is.

tesla_newbie

Took a drive in a 2019 model 3 today, does anyone know if the 2023 model 3 has the same chrome styling around wing mirrors?

it’s hideous

unkel

https://www.boards.ie/discussion/comment/121052316#Comment_121052316

Don't lecture me on what the TSLA share price is based on, I've been following it almost daily for many years, invested and divested in it many times, made a good few bob along the way too 😂

You show me the last time Musk mentioned FSD would be ready imminently and the TSLA shareprice the next NASDAQ working day to prove your point (and disprove mine)

Am I Evil?

Tesla have today heavily discounted inventory Model 3s in the US and Canada, rumors of highland announcement tomorrow in China seems to be gaining a lot more traction. I think RHD markets will severely lag behind with deliveries but interesting to see if we follow suite here and in the UK

Bovakinn

https://www.boards.ie/discussion/comment/121052332#Comment_121052332

They got rid of the chrome about 1/2 way through 2021 if memory serves. The black trim looks much nicer imo.

Bovakinn

https://www.boards.ie/discussion/comment/121052206#Comment_121052206

Im pretty sure hand coding feature extractors would be determining the best kernal for the convolutional layers manually, which is a terrible idea. Feature extractors are "learned" as part of the training process. I believe you're also suggesting changing the weights of the fully connected layers too, which doesn't make sense, although I might have picked you up wrong on that.

The space is moving fast, which is more the reason to keep an eye on it and not discount a system because it's not perfect at the moment. In ML achieving 80% accuracy is quite easy, it's the last 20% that's super tough.

Papers are fantastic, and they do lead to real-world use cases. Humans control the inputs in lab controlled environments, but as soon as the systems are tested in the real world, there's so many variables that a human can't modify. Even things like sensor noise will have an impact on a system and a lot of times these are reflected in the results.

Sensor fusion is tough, especially when you're dealing with different data types such as RGB images and LiDAR Point clouds, ToF sensors etc, Using a single sensor type simplifies this, although it does make the drawbacks of a certain sensor more problematic. For example a camera not being able to see through fog, whereas a Radar is completely unaffected. Only relying on cameras makes fog more problematic. The upside is if you can get the accuracy high enough and the reaction time short enough, the drawbacks can be minimised.

Like I said before, I don't think you necessarily need a redundant system in a personal self-driving vehicle, as long as the human can take over when a sensor fails or the system has low confidence for whatever reason.

KCross

https://www.boards.ie/discussion/comment/121052372#Comment_121052372

the TSLA shareprice the next NASDAQ working day to prove your point

That wouldn't prove the point so that's a straw man argument.

Musk repeatedly (and you too by the way!) say that when FSD is proven to work that the company will be worth many multiples of what it is today because the thought process is that Tesla will be able to charge insane money for FSD (its insane already to be honest). And that money will be almost pure profit at that point as its just enabling a feature.

Im not sure why you are even arguing the point. Clearly Tesla is not worth its insane valuation based on simply selling the cars. Its based on a future promise.

wassie

Was earlier than that. First batch of the M3 2021 refresh model with chrome delete were actually delivered in Q4 2020.

tesla_newbie

https://www.boards.ie/discussion/comment/121052484#Comment_121052484

I’ve a Y but might need to sell and get a 3 , the 2019 I drove today was no more comfortable than the 231 Y I have

bar physical size , is the tech any more advanced in the Y ?

unkel

https://www.boards.ie/discussion/comment/121052594#Comment_121052594

Agree with all that. Obviously!

But your point was:

"its just around the corner" from Musk is ridiculous, but it has pumped the share price no end" and that Musk said

"when FSD is proven to work that the company will be worth many multiples of what it is today"

So not even Musk himself believes that whatever he has said or done about FSD has or will up the TSLA share value until it is proven to work.

KCross

https://www.boards.ie/discussion/comment/121052636#Comment_121052636

So not even Musk himself believes that whatever he has said or done about FSD has or will up the TSLA share value until it is proven to work.

You are naive then.

Share price is based on future promise. If Musk says that, then investors invest on that promise/gamble.

Its pumping the share... not sure why you disagree with that. There are lots of examples where Musk said something on twitter and the share price bumped.... sure hasn't he even been successfully sued for it!

Put another way.... if he came out in the morning and said that he was giving up on FSD do you think the share price wouldn't move?!!!! We both know the answer.

listermint

https://www.boards.ie/discussion/comment/121052558#Comment_121052558

I'm 100 percent confident you are throwing your responses into chat gpt. I'd push the odds higher than that.

It adds little weight to the conversation by the way. You are attempting to make a human interaction inaccessible to the reader.

Try better.

unkel

@KCross - "There are lots of examples where Musk said something on twitter and the share price bumped"

You keep saying that, I don't believe you. That's why I asked you to come up with an example in the last year or two when Musk promised that FSD was imminent (he's promised that many times as you and I both agree) with the TSLA share price jumping up the next working day. You can pick the occasion that's most suitable for your argument, I'm not fussy.