1x Tsla Discussion Threads

[simon gordon] 09/5/2024 09:25

cfb2,

Cheers, I will look up Tony Seba.

-----

This is an interesting discussion, one quote that sticks out is "modelling and data are not reality":

Firms Consulting - 27/1/24

FLUKE: Chance, Chaos, and Why Everything We Do Matters — with Brian Klaas

Welcome to Strategy Skills episode 417, an interview with the author of Fluke: Chance, Chaos, and Why Everything We Do Matters, Brian Klaas. In Fluke, myth-shattering social scientist Brian Klaas deep-dives into the phenomenon of randomness, unpicking our neat and tidy storybook version of events to reveal a reality far wilder and more fascinating than we have dared to consider. The bewildering truth is that but for a few incidental changes, our lives - and our societies - would be radically different. Offering an entirely new perspective, Fluke explores how our world really works, driven by strange interactions and random events.

[johnwise] 09/5/2024 08:33

US delays ban on Chinese graphite in batteries while ex-China suppliers scrabble to source critical minerals

Chinese dominance on graphite supply

[cfb2] 08/5/2024 21:40

If you want someone who has an amazing track record about technological progress look up "Tony Seba".

[cfb2] 08/5/2024 21:35

I'm not sure if even the people working for Tesla, Google, OpenAI etc. know whether their competitors are BSing so we stand no hope!

Musk is super optimistic because it's in his nature, he doesn't understand the complexity of the problem and to motivate the people who work for Tesla. I personally believe he has some vile personal traits but he has inspired people to achieve what might once have seemed impossible.

If you're trying to do something that nobody has done before it is difficult to estimate when it will be done. Most of Musk's companies are working at the forefront of technology so I tend to cut him some slack about his predictions.

[cfb2] 08/5/2024 21:26

I have heard that either v12.4 or v12.5 of FSD will contain more than just LLMs for resolving mid-term route planning issues so we should find out in a month or so whether these new version are really leaps in technology, as Musk claim they'll be.

Multimodal is the new thing for AI, where problems are passed to specialised modules, which should cut down on the "hallucinations".

I haven't come across Gary Marcus before but I think his predictions are underwhelming.

I have found Dr. Alan Thompson provides good background on the state of developments in the AI world. He has posted some interesting technical comparisons of the different models on youtube without going too deep into the weeds.

Personally I hoping for another breakthrough in reinforcement training. Currently it seems like an awful lot of computing power is being thrown at the problem, which is why plenty of research is being done in this area. The current compute makes it difficult for anyone that hasn't got a few billion dollars to spare. The exclusive club restricts people like me and even Universities from being able to contribute to the field in a meaningful way.

[cfb2] 08/5/2024 20:32

simon: I expect people in the Tesla AI team smiled when they saw that. Both they and Wayve know they weren't the first but it give them some publicity, which they desperately need as they've just raised $1bn and may need multiple billions more in the near future if they are to compete.

[simon gordon] 08/5/2024 20:24

cfb2,

Do you think LLMs will take AI far enough for FSD?

I read the thoughts of Gary Marcus and he's been an LLM sceptic on the hype and is fairly sceptical that FSD is imminent.

His latest thoughts on OpenAI:

Gary Marcus - 8/5/24

Five predictions about OpenAI 2024-2025 -- from the person who told you first (even before it was released) that GPT-4 would continue to be unreliable and to hallucinate, that it would not have the imagined gigantic economic impact, and that pure LLMs would reach a point of diminishing returns:

• Unless GPT-5 is spectacular, OpenAI will eventually be viewed as the WeWork of AI, greatly overvalued relative to what it could deliver. They won't go bankrupt, but people will wonder why investors ever thought of them as a $100B+ business.

• Competitors will continue to narrow gap; OpenAI will never again be as dominant as it was in 2023.

• Nothing this year is likely to be worth the name of GPT-5. Instead, much of this year will be spent hyping a speculative future product that does not yet exist, while releasing only 4.X versions. (Sora will continue to generate a lot of press, but be costly to run, and too unreliable for many purposes.)

• OpenAI will not release a reliable agent architecture in 2024 or 2025.

• Both investors and the public will lose faith in OpenAI if Sam doesn't deliver something big within a year. (Microsoft is already hedging its bets.)

-----

From what I can see Marcus has been calling out Musk's hype on robots and FSD for a number of years. To date, he's been right. Are LLMs going to be enough to make FSDs a commercial reality?

[simon gordon] 08/5/2024 19:45

cfb2,

Found a quote from yesterday by the Wayve CEO about how Tesla has pivoted to their methodology:

"We’ve been all in on an end-to-end AI approach since we started in 2017, and were unique in doing so. It was fun to see Tesla last year pivot to this approach."

[dominiccummings] 08/5/2024 19:18

It is perfectly clear that Space Rockets CANNOT land vertically on their tails and therefore cheap space transport on reusable rockets will NEVER happen. Eh?

[hosede] 08/5/2024 17:36

Simon
Sounds like an area FULL of pitfalls.
Cfb
I think the mileage would in general be based on last years figures, and an adjustment would then be made via the MOT. Income tax works much the same. You pay an installment in advance based on expectations

[blusteradjuster] 08/5/2024 17:09

tbh Simon, even if autonomous vehicles were already 'perfect', we'd initially see public responses like this.

Humans being a suspicious/superstituous bunch.

The passage of time would, in such a situation, win them over.




"Nearly 70% of drivers say they’re afraid of fully self-driving cars, per a survey from the American Automobile Association. About a quarter say they’re unsure about them, while just 9% say they trust them."

[simon gordon] 08/5/2024 16:58

cfb2,

Thanks, I stand corrected.

Seemingly, Wayve will be selling their solution to OEM's. So many companies working on this. What does it mean for Telsa if they don't get their first or there is no moat? Apple just stopped their FSD programme, they probably spent a few billion.

"A remarkable $100bn has been spent on self-driving cars cumulatively (according to McKinsey), but little progress has been made. Some automakers have scaled back ambitions, while Ford and Volkswagen have pulled the plug, for now, on their self-driving car efforts completely. This has primarily been as a function of too many unpredictable ‘edge cases’ for autonomous cars to figure out. Nearly 70% of drivers say they’re afraid of fully self-driving cars, per a survey from the American Automobile Association. About a quarter say they’re unsure about them, while just 9% say they trust them."

[cfb2] 08/5/2024 16:44

hosede: People self reporting mileage along with a government department managing vast amounts of data sound like a recipe for disaster.

I agree about private cars, if you didn't need a car why would you want one with all the hassle and cost that comes with it. I know a number of people living in London who use the underground and walk rather than having a car. I don't have a transport infrastructure where I live but if I had access to robotaxis so I could beckon one within 10 minutes I'd happily give up my car.

[cfb2] 08/5/2024 16:34

simon: Andrej Karpathy, head of Tesla AI, around a decade ago was saying that neural networks (what he calls software 2.0) will eventually engulf the entirety of the C++ code. A quick search has popped up numerous youtube videos over 5 years old with him talking about it.

To get to end to end neural network much of the work that Tesla have been doing was needed as a stepping stone towards that goal. Obviously it's not possible to just wave your hands at a huge amount of video data and out pops FSD! There is still a lot of traditional software in FSD that can't be done in NNs, for example route planning or the "mind of the car" on the display.

[cfb2] 08/5/2024 16:11

mm: The cost is shown in their capex. They predict it will be $8-$10bn for each of the next two years.

I agree it's bad for the FSD to hit a curb. I expect training data has the car positioning itself badly on some turns, which Tesla will need to weed out. Given there were over 1m cars driving the FSD s/w last month you'd expect something like this to be amplified - if it happens people jump onto social media to complain. You don't get the equivalent from the millions of corners the car correctly navigates.

[simon gordon] 08/5/2024 13:10

cfb2,

From what I read Tesla has copied Wayve's methodology as it was failing with the previous one.

-----

Interesting piece in the FT about how brutal it is in China for some foreign car makers as the big dealerships drop them for local brands due to price.

[hosede] 08/5/2024 13:05

cfb
"Cars don't get MOTs until they're almost 3 years old."
That could easily be changed - by regular mileage checks - and people could be asked for advance estimates.
Anyway - as I've said - I don't expect there to be many private cars in the future

[cfb2] 08/5/2024 10:29

Cars don't get MOTs until they're almost 3 years old. Besides the government would prefer a steady revenue rather than a bill every year in case the driver is unable to pay it. One of the reasons why HMRC like a 6 month payment on account for next year tax. If they had their way you'd be paying every month, like PAYE/NI.

[johnwise] 08/5/2024 07:28

.
Stupidity has no limits