|
infernal machines posted:tesla had a massive first mover advantage thanks to their ability to use commodity batteries, they put a lot of work into their charging and power control systems, but those are still based on the premise of using a few thousand individual cells, rather than specifically designed prismatic batteries like everyone else is now doing This. The idea that Tesla has some magic tech is silly. Their infotainment is good, that's about it.
|
#
?
May 19, 2019 19:32
|
|
|
|
| # ? Jun 2, 2024 18:25 |
|
|
ford sold volvo's car division to china almost ten years ago now. (the rest of volvo group is still independent, publicly traded and has no single majority owner.)
|
|
#
?
May 19, 2019 19:37
|
|
|
MunchE posted:This. The idea that Tesla has some magic tech is silly. Their infotainment is good, that's about it. The whoopee cushion icon that farts?
|
|
#
?
May 19, 2019 19:49
|
|
|
The faith in FSD being just around the corner (need just a little bit more fleet data!) is especially annoying. Neural networks are cool as hell, but at the end of they are just statistics and have noticeable error margins. Either people don't understand the diminishing returns when Musk talks about "only" having to squeeze 99.99 to 99.999 percent (either number assuredly a loving bold faced lie anyway) or they actually believe in comic book geniuses that build magic crystal robot brains in their secret labs.
|
|
#
?
May 19, 2019 19:51
|
|
|
StabbinHobo posted:this is just false, BYD and SAIC both sell more EVs per month than tesla, the #3 ev seller. Yeah, but I was clearly talking about EV sales outside of China. Anyways, it seems weird that China would risk so much money on a company that's teetering on collapse. Maybe you're right, I shouldn't perceive the Chinese government as inscrutable, but rather a bunch of Bazingas that are passionate about Musk's vision! Hopefully Tesla goes bankrupt soon, and the fate of Gigafactory 3 will become apparent then.
|
|
#
?
May 19, 2019 19:52
|
|
|
Most of tesla's brand value is just Elon's cult of personality anyway
|
|
#
?
May 19, 2019 19:53
|
|
|
infernal machines posted:via yospos: 413.8 is a Homestuck as gently caress number an extinction event scale meteor is going to hit Gigafactory this year
|
|
#
?
May 19, 2019 19:53
|
|
|
Kindest Forums User posted:Yeah, but I was clearly talking about EV sales outside of China. what the gently caress would they care about ev sales outside of china for do you have *any* idea what you're talking about or are you just hitting a bong and shitposting?
|
|
#
?
May 19, 2019 20:01
|
|
|
BMan posted:Most of tesla's brand value is just Elon's cult of personality anyway Is it though? His cult of personality WAS the brand at one point. But now that EVs are mainstream, almost everyone has heard of Tesla. But it's not like they obsess over Musk like we or Bazingas do. They just know about the fast electric car everyone is talking about.
|
|
#
?
May 19, 2019 20:04
|
|
|
StabbinHobo posted:what the gently caress would they care about ev sales outside of china for https://www.youtube.com/watch?v=sZHCVyllnck
|
|
#
?
May 19, 2019 20:07
|
|
|
ok so your brain is just stuck in 90's kid lib brain economics mode
|
|
#
?
May 19, 2019 20:19
|
|
|
Kindest Forums User posted:Is it though? His cult of personality WAS the brand at one point. But now that EVs are mainstream, almost everyone has heard of Tesla. But it's not like they obsess over Musk like we or Bazingas do. They just know about the fast electric car everyone is talking about. they might have heard of them but theyre definitely not buying them
|
|
#
?
May 19, 2019 20:22
|
|
|
Shifty Pony posted:wtf is up with softbank anyway? they always seem to turn up when the dumpster fire hits maximum inferno softbank's play on sprint makes sense if you view it as softbank always intended to sell it for parts, because the parts were worth far more than the sum when comcast and others passed, softbank certainly lost money because now the fcc is the only one that will truly win in the end (tmo is basically buying sprint for the 2.5 spectrum and my gut says fcc will ask to re-auction the majority of what is leftover that sprint holds) hallebarrysoetoro has issued a correction as of 20:30 on May 19, 2019 |
|
#
?
May 19, 2019 20:28
|
|
|
StabbinHobo posted:ok so your brain is just stuck in 90's kid lib brain economics mode okay, I'm lost. Why doesn't China want to export goods? Because it sure seems like they're doing a lot of it if it's a bad thing. But maybe there's a specific reason they wouldn't want to export EVs. I honestly don't know.....
|
|
#
?
May 19, 2019 20:31
|
|
|
for the last few years China companies have no issue throwing the gauntlet in the face of the likes of Apple and gaining market share while simultaneously making money it really has been a good while since Chinese tech companies NEEDED to relabel themselves as a western brand to sell anything
|
|
#
?
May 19, 2019 20:31
|
|
|
I was just at an event in Barcelona where I had to walk through the big car expo for access. Tesla was there but I didn't see anyone test drive one (also didn't see any fires, explosions, or deaths).
|
|
#
?
May 19, 2019 20:47
|
|
|
lmao in what world do us regulators let the chinese government get literally anything of value from a tesla liquidation. god i hope it does happen because it would be a tremendous loving poo poo show
|
|
#
?
May 19, 2019 21:00
|
|
|
BioMe posted:The faith in FSD being just around the corner (need just a little bit more fleet data!) is especially annoying. the other half is assuming image classification is the problem. it's a problem for sure, and not nearly a solved one, but it's frankly a lesser problem than the elephant in the room. they're trying to drive a car by reacting to objects it identifies. there is no contextual awareness of the environment, and even your hypothetical circus dog has that. that's why they'll stop trying to charge through a closed patio door after bonking off the glass three or four times. even if they can't see it, they're aware that something is there that will prevent them from moving through that space. autopilot doesn't have that. autopilot doesn't have even the most basic conceptual framework to have that idea, and until it does it's going to drive under flatbeds and into highway gores following rules that don't apply in situations it can't differentiate
|
|
#
?
May 19, 2019 21:14
|
|
|
orange sky posted:In other news, Elon Musk is now QA absolutely cannot wait for this to start loving poo poo up (assuming he's not utterly full of poo poo lmao) other than i'll have to be vigilant in looking out for driverless teslas trying to murder me in the grocery store parking lot e: lol owned by the date but just assume i quoted the new one
|
|
#
?
May 19, 2019 21:29
|
|
|
Well in that specific tweet he's definitely full of poo poo because thats from november 2018
|
|
#
?
May 19, 2019 21:31
|
|
|
orange sky posted:Well in that specific tweet he's definitely full of poo poo because thats from november 2018 It is ready they just don't have time to implement cause there are so many car orders to fill.
|
|
#
?
May 19, 2019 22:02
|
|
|
Elon will personally merge every 10th line of code.
|
|
#
?
May 19, 2019 22:45
|
|
|
He's definitely gonna look at a bunch of invoices for expenses that have already happened and refuse to pay some suppliers or some poo poo like that because he decided they aren't essential.
|
|
#
?
May 19, 2019 22:56
|
|
|
BioMe posted:Either people don't understand the diminishing returns when Musk talks about "only" having to squeeze 99.99 to 99.999 percent (either number assuredly a loving bold faced lie anyway) No? Look they've basically trained the neural network on every visual scenario the car could ever encounter, just gotta finish up and add "truck blocking the lane" and it's done actually a Q for any neural net experts in here - you're basically training the network to provide a certain set of responses to a certain set of inputs, right? But every time you expand the complete range of input data (say by adding pictures of a cow because the car freaked out when it saw a cow on the road) and retrain the network to handle it, you risk losing "correctness" you had before for the original inputs, right? You're basically rolling some dice until you get a model that can hit certain targets, then you define even more targets and roll again until you're hitting all of those. And the more targets you add, the harder it is to hit them all with the same accuracy - and the input space for "everything you could see while driving anywhere in the world in any conditions" is insanely huge. Even if they maintain a massive test data set and keep retraining with all the inputs they handled before, isn't it basically a given that they'll struggle to hit a model that happens to handle all that? And correctness will start to regress as they have to settle for "good enough" results on a massive range of handled inputs, instead of only needing to handle a much more limited scope?
|
|
#
?
May 19, 2019 23:00
|
|
|
gschmidl posted:I was just at an event in Barcelona where I had to walk through the big car expo for access. Tesla was there but I didn't see anyone test drive one (also didn't see any fires, explosions, or deaths). Perhaps they only shipped Teslas without batteries to the show 
|
|
#
?
May 19, 2019 23:05
|
|
|
Kindest Forums User posted:okay, I'm lost. Why doesn't China want to export goods? Because it sure seems like they're doing a lot of it if it's a bad thing. But maybe there's a specific reason they wouldn't want to export EVs. I honestly don't know..... *then* when they get into the export game, its going to piggy back on the belt-and-road plan where they will provide half the ev's in all of asia and africa. at no point, in the entire future of the human race, does china need to give a poo poo about ev brand marketing in the US. if it ever comes to us buying imported chinese EVs it will be just like the last time we started importing cars heavily, because we're so desperate from an energy crisis we'll take whatever we can get. china's "investment in" (loan to) tesla is a completely pissant nothing-number in their overall energy and trade policy. they are not pulling some tricky move to save a few hundred million on a factory or steal some ip, they're just doing exactly what they say they're doing in their state policies. StabbinHobo has issued a correction as of 23:12 on May 19, 2019 |
|
#
?
May 19, 2019 23:08
|
|
|
Bofast posted:Perhaps they only shipped Teslas without batteries to the show Surely they drove themselves!!
|
|
#
?
May 19, 2019 23:15
|
|
|
Kindest Forums User posted:I don't disagree with any of this. China has a very competent and productive manufacturing sector, and I could see them mocking Tesla's approach. They may not be interested in any of the processes, tooling, or technology. But they would be interested in the markets. As far as I know, China has almost zero market penetration for their vehicles in the West. Having access to one of the biggest brands could change that. didn't chinese manufacturing just figure out how to make ballpoint pens in 2017
|
|
#
?
May 19, 2019 23:16
|
|
|
Yah, not so sure china believes in climate change.
|
|
#
?
May 19, 2019 23:24
|
|
|
gschmidl posted:Surely they drove themselves!! Brought via advanced summon to the show floor, only to have their batteries removed and placed in a climate controlled room in a fire and explosion proof basement.
|
|
#
?
May 19, 2019 23:25
|
|
|
infernal machines posted:the other half is assuming image classification is the problem. it's a problem for sure, and not nearly a solved one, but it's frankly a lesser problem than the elephant in the room. they're trying to drive a car by reacting to objects it identifies. there is no contextual awareness of the environment, and even your hypothetical circus dog has that. that's why they'll stop trying to charge through a closed patio door after bonking off the glass three or four times. even if they can't see it, they're aware that something is there that will prevent them from moving through that space. autopilot doesn't have that. autopilot doesn't have even the most basic conceptual framework to have that idea, and until it does it's going to drive under flatbeds and into highway gores following rules that don't apply in situations it can't differentiate This is exactly the case. Even if they somehow had a 100.0% accurate image segmentation tool that could identify everything in the video perfectly, that still doesn't mean it knows how to navigate these obstacles in a 3-d world. That's not even taking into account the future behaviour of those objects and following traffic laws. Recognising every item in a video doesn't mean you can drive!
|
|
#
?
May 19, 2019 23:28
|
|
|
StabbinHobo posted:the chinese communist party doesn't have a single republican in it. they believe in climate change and have arranged state policy around investing in, among lots of other things, battery and EV production. both in the form of literally state owned companies and in the form of massive subsidies/credits/taxes around favoring EVs over ICEs. they have no need, or intention, to export a single EV until domestic demand is entirely satiated. Everything here makes sense and I agree with you. But I'd like to think that China is creating the circumstances to majorly own the United Snakes of America and it's techlord Elon Musk. I just want to see America get owned, is that too much to ask?
|
|
#
?
May 19, 2019 23:32
|
|
|
Agronox posted:Yeah, plenty of funds eat poo poo on bad stocks even with (possibly because of) their access to insiders. A good one that comes to mind is Bill Ackman's Pershing Square fund and Valeant. Valeant was also a hilarious scam which probably deserves it's own Theranos like thread. For years big money investors believed Valeant had invented a new magical way to run a big pharma company and also acquires hundreds of companies.
|
|
#
?
May 19, 2019 23:39
|
|
|
baka kaba posted:actually a Q for any neural net experts in here - you're basically training the network to provide a certain set of responses to a certain set of inputs, right? But every time you expand the complete range of input data (say by adding pictures of a cow because the car freaked out when it saw a cow on the road) and retrain the network to handle it, you risk losing "correctness" you had before for the original inputs, right? so that out of the way, neural networks are not really acting like massive lookup tables, at least not the way youre implying and not as far we can tell. networks definitely pull some structure out of the data they get, even if there is not a very good answer about why they do it. the adversarial attacks where fifteen specific pixels can be changed to make the neural network think a cat is a dog actually demonstrates just how good nns are at picking structures out of data; a deep enough net is going to find structures in your data, and it is going to use them regardless of whether or not its a meaningful one or an artifact of bias in the dataset. networks also do not have to be tuned to average performance; the loss function can be anything so long as its continuous and sort of differentiable. the real question youre asking is "is it possible to make a dataset that 1: contains enough relevant information for the neural network to find a generalizable structure 2: free from biases that distort that structure 3: still small enough for the network to feasibly train on?". i would say probably yes? based on my own experiences with neural networks this is really a data engineering problem, not a fundamental design problem. of course, everything musk has said strongly implies that tesla's plan is to bootstrap their way into their dataset, and that gives me zero faith in their ability to actually deliver. i basically have zero faith in any of these companies except maybe google to actually have an even halfway rigorous approach to this problem.
|
|
#
?
May 19, 2019 23:55
|
|
|
think of the whole thing as america owning itself. musk wouldn't bother flying to china for meetings let alone build GF3 there if the u.s. did such basic things as "not cancel the battery credit" and "tax gasoline" and "not rescind the obama cafe standards" and "ban ICEs in various cities and counties".
StabbinHobo has issued a correction as of 23:58 on May 19, 2019 |
|
#
?
May 19, 2019 23:56
|
|
|
But in owning themselves, they owned the rest of the world through climate change. sad.
|
|
#
?
May 20, 2019 00:00
|
|
|
some sort of fish posted:im going to guess that tesla is not trying to go fully end to end and actually have the network go directly from camera feeds to driving commands. the only company i know of thats been trying that approach is nvidia, and from what ive heard they basically given up. based on what the videos of autopilot output, tesla is probably just trying to get object detection and depth approximation from their network. they definitely claim they're getting more than just object identification and depth from their camera data. The two examples from Fully Self Driving Day Honest I remember are predicting when a car is going to cut into your lane (by feeding it video where that happens and annotating the vehicle as "about to cut in") and also learning how to calculate paths through an environment by just feeding it user driving video and getting it to predict a similar route I know neural networks aren't a lookup table, they're a system of weighted connections that (ideally) map "similar" inputs to "similar" outputs, so it represents a machine that perceives pertinent structure in the input and produces the correct, relevant type of response. What I'm getting at is it's a lot easier to create a machine that can, say, classify handwritten numbers than it is to reliably handle every possible image a car camera could see at any time anywhere, and correctly "perceive" what's going on enough to produce the correct output. It seems like an, uh, ambitious project, no? Like there was that project where the military was trying to automatically detect tanks visually, right? And they trained their network on all these tank images, different designs all from different sides - and when they went out in the field it didn't work, because the lighting conditions weren't the same as in their training data pics. And yeah they can correct that, but it just goes to show how one tiny difference in the environment can completely throw a system designed solely to look for one particular visual element in a certain kind of image. Musk believes he can train a system that can handle any environment, any conditions, accurately identifying every relevant element in an image and doing it in real time while in control of a vehicle that relies on the network always being right just seems to me that any neural network that could possibly handle such a massive and complex set of potential inputs would need to have a hell load of internal connections, just to make it fine-grained enough that it could be tuned to correctly handle everything. And that tuning process would become massively more time-consuming as the set of training data grows to encompass a big enough sample of the world. Every generation risks regressing more and more, as a tiny change to the complex machine ripples out into differences in the way it responds to the original set of input data, you know? You find an equilibrium that gives good results for everything, then you throw something new into the mix and have to find a variation that handles all of it, and it'll get harder and harder to maintain a good score for everything sorry for the gigapost!!
|
|
#
?
May 20, 2019 00:40
|
|
|
The tank story is an urban myth, but the point it was making was that AI takes easy shortcuts if you give it the chance (like only detecting sunny skies instead of tanks if all your photos with tanks had sunny skies and the ones without tanks were cloudy). In general you can get a neural network to do a lot of things at once and there's some generalization of function - the one that got used to make https://www.ganbreeder.app is trained for a shitload of different imagetypes, instead of single-category networks that only work on faces ( like https://www.thispersondoesnotexist.com ) In principle, you could also just train the network by having it watch real drivers and try to predict what they do. This is what teslalistas belive is happening with their cars, even though it isn't. The strategy works for mario kart at least https://youtu.be/Ipi40cb_RsI Tunicate has issued a correction as of 01:07 on May 20, 2019 |
|
#
?
May 20, 2019 01:03
|
|
|
Yeah but that's still "parse some structure from these images and produce a combination that follows some rules" rather than "here's an image from a car travelling at speed! what's happening in it - think fast". Plus how many of those generated images look completely messed up rather than the nice results they put in the gallery? A car's visual state-classifying system can also produce messed up images, but those are the after photos didn't realise the tanks thing was a lie though 
|
|
#
?
May 20, 2019 01:10
|
|
|
|
| # ? Jun 2, 2024 18:25 |
|
|
Tunicate posted:In principle, you could also just train the network by having it watch real drivers and try to predict what they do. This is what teslalistas belive is happening with their cars, even though it isn't. The strategy works for mario kart at least fwiw tesla claims they're doing exactly this https://www.youtube.com/watch?v=Ucp0TTmvqOE&t=8014s
|
|
#
?
May 20, 2019 01:16
|
|