5 ms·

I agree entirely with the sentiment of the article, and I think Elon and Tesla are hopelessly deluding themselves with belief in their superior engineering and

I agree entirely with the sentiment of the article, and I think Elon and Tesla are hopelessly deluding themselves with belief in their superior engineering and product design. But... Just because a few companies failed at factory automation in 1980 doesn't mean it couldn't succeed in 2018. It's patently absurd to knock Tesla for trying it again in the era of ultra-high resolution cameras and advanced machine learning algorithms.

std_throwaway 9y ago
It's not that other auto-makers were sleeping all this time. They build new factories every now and then and increase automation where appropriate. "ultra-high resolution" doesn't mean much if you can't see the problem. "advanced machine learning" sounds dangerous once the machine damaged itself or killed a bunch of people trying to figure out things.
Robotbeat 9y ago
This is true. But a key point: Tesla doesn't actually have to beat state of the art in automation to be profitable. Their cars are already great, electric cars are simpler, they've streamlined the sales process in a way that already gives them at least 10% greater cost advantage if everything else is equivalent (while also eliminating a major pain point for consumers). But it goes beyond just the fact that technology has advanced: Tesla is willing to pursue alternative approaches where others have long stopped (and this is a point the author of this article makes later on). The fact that Tesla doesn't have the conservative culture of other car makers means that they'll make mistakes like this, but also that they'll find new solutions (or recognize existing solutions in similar fields that haven't become industry standard) whereas others would not. You cannot automate EXISTING car designs much. And Model 3 does make improvements, but in many ways is built similar. It's like trying to automate soldering components by hand. Automation of electronics required a change in the fundamental way electronics were built. Through-hole is hard to automate but relatively easy to build by hand. Surface mount is simple to automate and ultimately better in several ways, but is super annoying to build by hand. Tesla needs to find the surface mount of car manufacturing. And I think they're trying. Something particularly hard to automate for automobiles is the wire harness. There are many degrees of freedom; it's kind of like tying your shoe: easy for humans, hard for machines. The Model 3 uses less wire than the Model S/X, but only by a factor of 2 or so. Model Y, on the other hand, is supposed to use like an order of magnitude less. This is like optimizing a circuit board for automation by only using a few through-hole components, relying mostly on surface mount. So Tesla is going to have to continually redesign their vehicles to be more and more amenable to automation. They can't do this as a step function, and Musk has realized that now. (This is another good point the author makes.)
- gamblor956 9y ago
  Tesla has yet to sell a car at a profit (using standard accounting practices, known as GAAP, for the auto industry or even in general). They're still losing money on a per-car basis, so the more they sell...the more money they lose. And don't give me that crap about R&D blah blah blah. Other car makers include R&D and other related capex in their per-car profit accounting.
  - nugga 9y ago
    This doesn't sound right. As far as I heard they make something like 20% or more profit on sold cars but due to constant heavy investment they burn through cash reserves.
    - gamblor956 9y ago
      Tesla makes a 20% profit per car if you ignore standard auto industry accounting practices and use magic voodoo accounting and ignore the capex that goes into developing and building those cars. Notably, the only sites claiming that Tesla realizes per-car profits are tech companies which defend the use of non-standard accounting practices. Note: Most other car companies make hundreds of thousands of cars at each of their factories, and sell hundreds of thousands to millions of cars each year, which substantially reduces the per-car capex. Tesla sells comparatively very few cars, so its per-car capex dwarfs the rest of the industry.
      - Robotbeat 9y ago
        Except that Tesla is now producing at a weekly rate equivalent to hundreds of thousands of cars per year at their Fremont factory, comparable to other car companies. But they're still ramping up. The factory will be outputting something like 500,000 to 700,000 cars per year after they're done ramping. They'll need Model Y and/or Semi and at least another factory before they're going to be comparable to other mid-range manufacturers. They'll need at least half a dozen factories to be a major manufacturer. Or they'll need to change how cars are made entirely. If they can get millions of cars from a single factory, then the high automation that they're shooting for (and which Model 3 had to back off of) might make sense (and is required for getting that kind of production from a single line). But even that will likely need multiple factories to justify the R&D into the factory line itself.
agumonkey 9y ago
Agreed. It's been long enough to revisit the limits. Let's review SpaceX mistakes and compare them to the 80s, they probably did a lot of them. Nobody thought it would be possible to make reliable reusable low cost launchers. And here we are. ps: not a Musk fanboy, just a middleground spectator.
- buvanshak 9y ago
  > reliable Is it really time to say this?
  - agumonkey 9y ago
    They've reused multiple stages more than once already. Considering the constraints on these devices it's already quite a proof. Sure we'll know after more samples but .. well I said it.
- ummonk 9y ago
  The space industry isn't really comparable to the automotive industry. The former has been stagnating on government contracts while the latter has been heavily innovating for many decades due to massive competition and markets.
talltimtom 9y ago
The failure isn’t in trying to automate. The failure is in expecting automating everything to go perfectly and therefore making the line troublesome to carry out when specific automated sub tasks fail.
- std_throwaway 9y ago
  The whole line stops if a single process has problems. Given a 99% chance of working for every single process for hundreds of processes you end up with... well... not much output.
  - phkahler 9y ago
    I've known a few people who accidentally stopped an automotive production line. The plant manager will be there in minutes to see why the F$%* production is falling short to the tune of 5 figures per minute. When I was younger I had a less severe but equally enlightening experience. A transmission case line with 24 CNC stations had our UI software running on PCs to make them easier to use. They could run fine without the fancy UI but it was how the operators had come to run things. I went in one afternoon to load a software update from floppy disk. The plant manager came out. "What are you doing?" "I'm updating this UI software." "You work for $company?" "Yes." "Is there gonna be a $company guy on site for 3rd shift tonight if something goes wrong with that software?" "No." "You aren't gonna update it then." I went home. Go fast and break things doesn't work in that environment, and yes the critical processes work far more than 99% of the time because as you say there wouldn't be much output. It sounds like that's the exact lesson Tesla is learning now. [edit] the reason all the processes on the line have to work is because it's basically impossible to store a bunch of cars (or doors or engines) anywhere until things get fixed. Ad-hoc fixes are just not feasible at that scale.
gwern 9y ago
I don't think it's absurd either, but I think Tesla was the wrong company to try to do it in 2016-2018. Musk and everyone else agrees that Model 3 is a do-or-die time for Tesla, and the financial picture for Tesla looks dire as it is sitting on a sea of red ink from an enormous burn rate financed by bond debt. That is not a situation in which to take a likely profitable but unnecessary systemic risk which imperils the entire Model 3 program by its nature. If GM or Toyota tried an experimental lights-out factory, it's not going to potentially kill them. For Tesla, that's a possibility.

std_throwaway 9y ago

It's not that other auto-makers were sleeping all this time. They build new factories every now and then and increase automation where appropriate. "ultra-high resolution" doesn't mean much if you can't see the problem. "advanced machine learning" sounds dangerous once the machine damaged itself or killed a bunch of people trying to figure out things.

Robotbeat 9y ago

This is true. But a key point: Tesla doesn't actually have to beat state of the art in automation to be profitable. Their cars are already great, electric cars are simpler, they've streamlined the sales process in a way that already gives them at least 10% greater cost advantage if everything else is equivalent (while also eliminating a major pain point for consumers). But it goes beyond just the fact that technology has advanced: Tesla is willing to pursue alternative approaches where others have long stopped (and this is a point the author of this article makes later on). The fact that Tesla doesn't have the conservative culture of other car makers means that they'll make mistakes like this, but also that they'll find new solutions (or recognize existing solutions in similar fields that haven't become industry standard) whereas others would not. You cannot automate EXISTING car designs much. And Model 3 does make improvements, but in many ways is built similar. It's like trying to automate soldering components by hand. Automation of electronics required a change in the fundamental way electronics were built. Through-hole is hard to automate but relatively easy to build by hand. Surface mount is simple to automate and ultimately better in several ways, but is super annoying to build by hand. Tesla needs to find the surface mount of car manufacturing. And I think they're trying. Something particularly hard to automate for automobiles is the wire harness. There are many degrees of freedom; it's kind of like tying your shoe: easy for humans, hard for machines. The Model 3 uses less wire than the Model S/X, but only by a factor of 2 or so. Model Y, on the other hand, is supposed to use like an order of magnitude less. This is like optimizing a circuit board for automation by only using a few through-hole components, relying mostly on surface mount. So Tesla is going to have to continually redesign their vehicles to be more and more amenable to automation. They can't do this as a step function, and Musk has realized that now. (This is another good point the author makes.)

gamblor956 9y ago
Tesla has yet to sell a car at a profit (using standard accounting practices, known as GAAP, for the auto industry or even in general). They're still losing money on a per-car basis, so the more they sell...the more money they lose. And don't give me that crap about R&D blah blah blah. Other car makers include R&D and other related capex in their per-car profit accounting.
- nugga 9y ago
  This doesn't sound right. As far as I heard they make something like 20% or more profit on sold cars but due to constant heavy investment they burn through cash reserves.
  - gamblor956 9y ago
    Tesla makes a 20% profit per car if you ignore standard auto industry accounting practices and use magic voodoo accounting and ignore the capex that goes into developing and building those cars. Notably, the only sites claiming that Tesla realizes per-car profits are tech companies which defend the use of non-standard accounting practices. Note: Most other car companies make hundreds of thousands of cars at each of their factories, and sell hundreds of thousands to millions of cars each year, which substantially reduces the per-car capex. Tesla sells comparatively very few cars, so its per-car capex dwarfs the rest of the industry.
    - Robotbeat 9y ago
      Except that Tesla is now producing at a weekly rate equivalent to hundreds of thousands of cars per year at their Fremont factory, comparable to other car companies. But they're still ramping up. The factory will be outputting something like 500,000 to 700,000 cars per year after they're done ramping. They'll need Model Y and/or Semi and at least another factory before they're going to be comparable to other mid-range manufacturers. They'll need at least half a dozen factories to be a major manufacturer. Or they'll need to change how cars are made entirely. If they can get millions of cars from a single factory, then the high automation that they're shooting for (and which Model 3 had to back off of) might make sense (and is required for getting that kind of production from a single line). But even that will likely need multiple factories to justify the R&D into the factory line itself.

agumonkey 9y ago

Agreed. It's been long enough to revisit the limits. Let's review SpaceX mistakes and compare them to the 80s, they probably did a lot of them. Nobody thought it would be possible to make reliable reusable low cost launchers. And here we are. ps: not a Musk fanboy, just a middleground spectator.

buvanshak 9y ago
> reliable Is it really time to say this?
- agumonkey 9y ago
  They've reused multiple stages more than once already. Considering the constraints on these devices it's already quite a proof. Sure we'll know after more samples but .. well I said it.
ummonk 9y ago
The space industry isn't really comparable to the automotive industry. The former has been stagnating on government contracts while the latter has been heavily innovating for many decades due to massive competition and markets.

talltimtom 9y ago

The failure isn’t in trying to automate. The failure is in expecting automating everything to go perfectly and therefore making the line troublesome to carry out when specific automated sub tasks fail.

std_throwaway 9y ago
The whole line stops if a single process has problems. Given a 99% chance of working for every single process for hundreds of processes you end up with... well... not much output.
- phkahler 9y ago
  I've known a few people who accidentally stopped an automotive production line. The plant manager will be there in minutes to see why the F$%* production is falling short to the tune of 5 figures per minute. When I was younger I had a less severe but equally enlightening experience. A transmission case line with 24 CNC stations had our UI software running on PCs to make them easier to use. They could run fine without the fancy UI but it was how the operators had come to run things. I went in one afternoon to load a software update from floppy disk. The plant manager came out. "What are you doing?" "I'm updating this UI software." "You work for $company?" "Yes." "Is there gonna be a $company guy on site for 3rd shift tonight if something goes wrong with that software?" "No." "You aren't gonna update it then." I went home. Go fast and break things doesn't work in that environment, and yes the critical processes work far more than 99% of the time because as you say there wouldn't be much output. It sounds like that's the exact lesson Tesla is learning now. [edit] the reason all the processes on the line have to work is because it's basically impossible to store a bunch of cars (or doors or engines) anywhere until things get fixed. Ad-hoc fixes are just not feasible at that scale.

gwern 9y ago

I don't think it's absurd either, but I think Tesla was the wrong company to try to do it in 2016-2018. Musk and everyone else agrees that Model 3 is a do-or-die time for Tesla, and the financial picture for Tesla looks dire as it is sitting on a sea of red ink from an enormous burn rate financed by bond debt. That is not a situation in which to take a likely profitable but unnecessary systemic risk which imperils the entire Model 3 program by its nature. If GM or Toyota tried an experimental lights-out factory, it's not going to potentially kill them. For Tesla, that's a possibility.