-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
- Status
It was called the "Signature" series and required a $40,000 deposit
I know what you are saying, but I'm gonna add this anyway:
Just a better version, but SpaceX made their own flow/ combustion simulation SW
Just a better version, but SpaceX made their own land a rocket software
Just a better version, but SpaceX came up with a new material for their engines (i think specifically the turbopumps).
At least (unlike Edison) they didn't need to find a hundred ways not to land a rocket.
I will state that I still read myusername due to his occasional short-term technical analysis (he seems to have some talent at it) and I have put *very few* people on ignore, not even some of the visitors who are admittedly Seeking Alpha bear FUDsters. MMD is the first person I ever put on ignore. He never *ever* posts anything of any value, and never even accepts correction when he's caught out making a clearly false statement (most people accept correction at least *sometimes*).
I really would recommend his permanent removal.
Yesterday TREFIS, a stock analysis company put out a release titled "Why Battery Improvements Will Determine The Success of The Tesla Semi".
The gist is that success of the Semi when it begins to seriously ramp in 2020, is possible but not certain. I.e. significant further reduction in pack cost must occur by then for Semi margins to be what Tesla is claiming they will be (given pricing).
The analysis seems sound but only if the author's pack price per Kwh is correct.
There are a few members who have tracked and commented on where pack cost currently stands. Perhaps one or more might share what they think of the authors math. He/she states that Tesla claimed in 2016 that pack cost was around $190 per KWh. Then this year stated or implied that for one or more reasons (like GF scale up reducing costs) there had been a further 30 - 35% reduction. Bringing current
pack cost to $130. From that author makes a conservative guess that by 2020, IF continued cost reductions bring pack cost down to $100, a 1 MWh Semi pack would cost Tesla $100,000, and this would be low enough for good gross margin.
Does $130 pack cost per KWh seem like a good current day figure? I think I recall some who have followed this closely believe the
pack cost may already be lower that $130. Apologies if I haven't recalled that correctly.
From TREFIS:
Battery Prices Will Determine Tesla’s Margins
Tesla doesnt regularly disclose the cost of its batteries, but the company claimed that its battery packs cost under $190 per kWh in early 2016 and earlier this year, Tesla claimed reductions of about 35%, implying that its battery costs could stand at levels of under $130 per KWh, considering the companys large production volumes at its Gigafactory and its move to the new 2170 battery cell on newer vehicles like the Model 3. Costs could decline further over the long run, as Tesla scales up production. For instance, studies have shown that electric vehicle battery costs have declined by about 80% in six years. If Tesla is able to bring down per kWh costs to levels of around $100 by 2020 (assuming rates of declines significantly below historical rates), as the Semi begins to scale up production, this could put prices of the battery on the 500 mile version at $100,000, which could still leave room for reasonable gross margins.
The gist is that success of the Semi when it begins to seriously ramp in 2020, is possible but not certain. I.e. significant further reduction in pack cost must occur by then for Semi margins to be what Tesla is claiming they will be (given pricing).
The analysis seems sound but only if the author's pack price per Kwh is correct.
There are a few members who have tracked and commented on where pack cost currently stands. Perhaps one or more might share what they think of the authors math. He/she states that Tesla claimed in 2016 that pack cost was around $190 per KWh. Then this year stated or implied that for one or more reasons (like GF scale up reducing costs) there had been a further 30 - 35% reduction. Bringing current
pack cost to $130. From that author makes a conservative guess that by 2020, IF continued cost reductions bring pack cost down to $100, a 1 MWh Semi pack would cost Tesla $100,000, and this would be low enough for good gross margin.
Does $130 pack cost per KWh seem like a good current day figure? I think I recall some who have followed this closely believe the
pack cost may already be lower that $130. Apologies if I haven't recalled that correctly.
From TREFIS:
Battery Prices Will Determine Tesla’s Margins
Tesla doesnt regularly disclose the cost of its batteries, but the company claimed that its battery packs cost under $190 per kWh in early 2016 and earlier this year, Tesla claimed reductions of about 35%, implying that its battery costs could stand at levels of under $130 per KWh, considering the companys large production volumes at its Gigafactory and its move to the new 2170 battery cell on newer vehicles like the Model 3. Costs could decline further over the long run, as Tesla scales up production. For instance, studies have shown that electric vehicle battery costs have declined by about 80% in six years. If Tesla is able to bring down per kWh costs to levels of around $100 by 2020 (assuming rates of declines significantly below historical rates), as the Semi begins to scale up production, this could put prices of the battery on the 500 mile version at $100,000, which could still leave room for reasonable gross margins.
To elaborate this point: MMD is essentially spamming.
I don't really know what the lead time is on the new robots and factory equipment. If it's a rush job, however, "spare no expense", I really would expect it to take about three months. Some things like gigantic stamping presses take longer, but this is smaller stuff -- nothing should have a super long lead time.
Especially when it's tying everything else up. But if there is a third party part needed...
Anstandswauwau
Member
But that's exactly the point why now is the exact time FSD will be solved: The current, true exponential increase of knowledge and efficiency concerning deep learning methods! I'm not sure if you're aware of it, but the amount of research and progress in this area is just bonkers. It's like witnessing a century worth of computer science compressed into half a decade or so.
To get back to the quote above: By leveraging, let's say, convoluted neural networks (CNN), there's no need to understand what exact problem you're trying to solve and how it can potentially be done. The only thing you must be able to do, is to verify/quantify the output of said CNNs. And that's quite easy to do in the case of FSD by using simulations and shadow driving.
To give you another example, we have no (intuitive) understanding of how state of the art visual classifiers work. The only thing we know is: They do work – in some cases even better than what we're capable of. Figuring out how they exactly work will likely take 10 times longer than, you know, getting them to work in the first place.
I'll disagree that SpaceX didn't invent anything. They've invented many new rocketry and related technology hardware and improvements.
I think what you're saying is SpaceX didn't discover any new fundamental physics.
The same applies to Tesla, their innovations/inventions were guided by close attention to the fundamental physics saying what is theoretically possible. Tesla's long list of patent applications are for these specific inventions. The company has stated they will allow other companies to use IP in granted patents.
That's fair; it would probably double the capital cost.
But with the lifetime of the racking system being essentially infinite (how long do steel-frame buildings last?) I'm not sure how to amortize the cost over the power production. Since a canopied parking lot has its own value for other reasons (snow avoidance, sun avoidance) it really should probably be thought of as a real property improvement and not figured into the cost of the power. Different companies will probably view the value of this differently.
No, you've missed the point. You're simply wrong here. We absolutely do need to understand the problem in order to verify or quantify the output. There are areas where it is actually disputed what the driver should do in a given situation.
We need a specification of success. Even if our method for solving the problem is "try random crap until we find something which works", we still need a specification of success.
(Did the US win the Iraq War? Well, what do you mean by "win"? When a US major in Vietnam said "'It became necessary to destroy the town to save it", he was using a different meaning of "save" than the entire civilized world.)
Nope. It's not even possible to tell after the fact whether it did a good job or not, because we don't have consensus on what a good job is, and this is the problem. To give an exaggerated example, a Mumbai driver and a Swiss driver will *disagree* on whether they were looking at competent driving.
*I* have disagreed with Papafox on whether I'm looking at competent driving.
Unfortunately, some people have pretty scary definitions of competent driving which allow for a lot of mayhem and pedestrian death. I don't want an autonomous car designed by those people!
This is (usually) applied to problems with a solid and testable definition of whether the classifier is correct or not. In the medical context, this is not always the case, and in those contexts, they quite honestly don't actually work; they lead to a bunch of false positives or false negatives.
This is essentially a version of the GIGO problem. You can get something which performs wonderfully at doing what you told it to (or in the case of so-called "deep learning", what you *trained* it to do), but what happens when that isn't actually what you wanted it to do?
Last edited:
Okay so I'm far from an expert, but to common sense it here I gotta disagree considering that google has been doing self-driving for quite a long time, self driving cars are on roads today in some capacities, and most tellingly Tesla is selling FSD right now. Now Tesla can get away with a lot, but I don't think they can get away with selling a product that is 5+ years away, and it's not worth it to them to try and fudge things.
On the more technical side there are a couple youtube videos from the guy who used to head up Autopilot, and some from the comma ai guy, and from what I can gather from those, if you crunch the numbers, is that Tesla likely can do FSD right now, they are just gathering enough data to be able to prove beyond a doubt that is as safe or safer than the average person, Musk pretty much said that on the last conf call. It seems like the regulatory and liability issues are much bigger hurdles at this point than the technological part, the "problem" is pretty much solved they just have to keep logging miles.
Last edited:
Anstandswauwau
Member
As I'm working in this field, I highly doubt it
There's little point in discussing this with you as you seem to have a very limited understanding of this topic, no offence! So let's just agree to disagree here, okay?
In terms of the big picture, definition of success is an equal to or lower rate of accidents than people. In other words just make something that is statistically as good or better than the average human which there are lots of good benchmarks for, the most important probably being fatalities.
OK. Since it's clear that you have the standard biases and blinkered view of those engineering types who work in a field and want to sweep the underlying problem specification issues under the rug -- no offence, I've seen it a million times, it's a completely standard error made by engineering types -- it's clear there is no point in discussing this with you. We'll agree to disagree.As I'm working in this field, I highly doubt it
There's little point in discussing this with you as you seem to have a very limited understanding of this topic, no offence! So let's just agree to disagree here, okay?
Good luck on your work.
Good start. But it's more complicated than that.
First issue:
Fewer fatalities and accidents is actually very easy to achieve -- we can achieve zero, no problem. And hey, that's a success specification which *I* like.
But it quite consistently means people get to their destination a bit slower, or don't go out in really bad weather, etc. Other people (not me) constantly deliberately make that tradeoff, threatening other people and themselves in order to get to their destination fast in bad weather.
Care to rethink your success specification?
I think that assumes that the car gets from A to B in a reasonable amount of time considering the weather/conditions. Again the benchmark is probably as fast or faster that the average person.
I'd really recommend watching the Sterling Anderson youtube videos, they were pretty informative and probably a year or two old now which is a lifetime in the tech world.
to follow what @neroden said, realizing I am not well versed in this, so I probably shouldn't even be typing:
CNN can sort of be analogous to the subconscious/ muscle response, whereas verification requires the known/ conscious level.
I don't know how I recognize words (even when upside down or on an angle, or when reflected), but I can tell you exactly how letters look.
I don't know how I recognize language (even with accents, pitch changes or static), but I can tell you if a sentence is (mostly) proper grammar.
I can recognize if my interpretation of the auditory or visual stimulus is inconsistent with known patterns.
The matrix of coefficients when applied to the adjusted input data produces an output that can be mapped to a desired response. But that block box does not have a direct mapping to what the data means. It is infeasible to apply all sets of values to the input to validate the output. During training, only the stimulated inputs get tuned. All you can know is that for the cases you have tested, it works.
This unknowingness can be helped by reducing the problem to specific things, what is a lane, what is an obstacle. Then apply rational guidelines to those outputs along with the rules of the road.
While on the subject, can anyone point me to a good article on how they deal with time variant data? Only thing I've read is Karpathy's pong NN which was fed the difference between two frames of the game (which seems to me to just be a feed forward table)
.
- Status
Similar threads
- Sticky
