Lots of folks here are talking about the latency and throughput, but nobody is amazed at what the robots are doing. Plugging in cables and putting down tape is fiddly and hard. I'm super impressed that they elected to automate those things based on cost merit. Massive kudos to the manufacturing engineers who managed to pull that off.
I had invested in a semiconductor factory early last decade, and it had started making bitcoin miners, it had reminding me of Zion from the Matrix because it seemed like the first time that a machine could make another machine which could immediately start earning a negotiable asset and then transacting for resources.
Early in the game, you'd hack machines to run mechanical-turk style jobs to earn money. Later you'd build a factory to manufacture more processors... ;)
Some of the first bitcoin faucets had captchas to “reduce spam”, as in prevent humans from getting too much free bitcoin, but really the captcha was a real captcha that a bot was stuck on, and was just paying humans to solve them to access some greater bounty
Morally ambiguous, but I actually implemented exactly what you’re describing back when bitcoin was at $50-100. I found some service that would pay me $1 for every 1000 captchas I solved, and then I scraped that and built a BTC faucet site (I think I called it captchacoin.com, or captchabit.com) that would essentially pay the user $0.50 in BTC for each captcha they solved.
It went from 0 captchas per day to 10000 captchas/day over the course of about a month with just word-of-mouth. Then the upstream service just quit paying me with no explanation. My accuracy rates were good. Maybe latency was high? I don’t know why they terminated it. But I struggled to keep the users while finding a replacement, and in the end I shut it down.
On the one hand, I’m bummed, because that thing could have mostly run itself and made a tidy profit for someone still pursuing a degree. On the other hand... maybe it was for the best: 90% of captcha solving work is paid for by spammers who make the internet worse.
Have an upvote for the story, but don't pretend it was "morally ambiguous". That was plain evil with one level of detachment, and helped enable spammers and scammers.
Not only scammers and spammers are restricted by CAPTCHAS. Helping bypass them would only be immoral if they were never used to prevent scraping or other illegitimate purposes.
Thats a huge misconception actually, the information you have is the amount of energy being used, but you don't have information on the source of the energy.
Would you be willing to accept or at least corroborate that cryptocurrency mining is one of the cleanest sectors and a boon for sustainability?
70-80% of that energy use is renewable energy or reducing pollution - specifically hydrocarbons. So existing energy is not being wasted or reallocated, and additional unclean energy is not being ramped up to facilitate mining, and previously wasted energy that was going into the atmosphere is now being used.
The educated discussion is to make sure it stays this way. As nation states are the only actors that could mine at a loss with inefficient ways.
I actually utilized an API that did just that in some automated software I wrote with a friend a few years ago. Was actually sort of interesting how it all worked with sending the captcha then checking for the solution string after the person had solved and inserting that back into the page before carrying on.
Pretty sure at the time the cost was something like .013 cents per solve. I always wondered who the people sitting there doing that were or what their situation was.
The combination of being played on a real world map and covertly growing/spreading reminds me of Plague Inc. I wonder if any inspiration came from this.
There is a very interesting and creepy story about this called The Gig Economy [1]. Machines end up sending gig workers to do inane tasks, like photographing a certain building at a particular time, to take advantage of bizarre arbitrage opportunities too complicated for a human to notice. Eventually, everyone is just a peripheral of this enormous, inscrutable system skimming money off the real economy.
Reminds me of a short run podcast called The Program (https://programaudioseries.com/). It's a historical documentary about the rise of "The Program" which is basically an AI directed gig economy becoming the economy. It asks people to do strange things for seemingly arbitrary reasons but it makes sense when you realize what all it is doing.
Early lights out production. The problem became obvious a little later, for NeXT, which was that these things only make sense at very high scale for products that change infrequently.
these boards werent consumer producs though. playstations are produced for years with no alterations.... millions of times. i.e. playstation 4 was sold ~100million times. thats most definitely at a scale where you can automate with margins.
Game console manufacturing definitely is low-mix compared to cars, but it's not like they're stamping out spoons: depending on how you count, there were something like 19 revisions to the PS3 motherboard, as Sony worked to reduce chip count and shrink PCB size: https://www.psdevwiki.com/ps3/Motherboard_Revisions
Specifically, in the context of assembly, as long as dimensions and locations of each connector stay the same, it does not really matter for the automated assembly line.
I was trying to clean the dust out of my PS4 a little while ago, and had to browse through quite a few YouTube videos before I found a disassembly tutorial that matched my model.
Years ago I worked for a company that had a PCB production line, and even for relatively low volumes (compared to consumer electronics) they used basically that exact process. The PCBs would be printed elsewhere and delivered in flat stacks, the boards would go through pick-and-place, solder bath, reflow, through-hole, wave solder, sometimes xray inspection, etc. Also had a tour of an Iomega plant when I was much younger with similar processes.
Not sure if NeXT was a pioneer in those techniques, but they seemed standard by the late 1990s and 2000s.
The top tier "chipshooter" style pick and place machines are going our of favour these days.
Their golden days were during the first cellphone boom. Back then, the level of integration was lesser, and you had more discrete components on more smaller boards, and volumes were of course very high during the boom time.
Now, you can have a "dumphone" made with just 30 parts on the pcb, and very few passives.
From my experience over the decade, people running factories came to love having multiple, cheaper mounters, and more lines.
The "superboard" concept is also seem to be waning, as you see more, and more individual boards in products like smartphones. It makes for less manufacturable designs, but additional labour expense is not dramatic.
I'm surprised they didn't design out stuff like tape. Although perhaps they did and the article actually means handling things like ribbon cables, which is super impressive, as humans can barely handle those.
My guess is that they have flex cables (and not so much ribbons) and tape. Other electronic devices - including modern ones - that I've taken apart lately include quite a lot of both.
Tape takes a number of forms. Holding down flexes to manage pull-out forces is pretty common. You can do that with traditional single-sided tape if you have the space on something to adhere it, but increasingly common is a double-sided tape on the flex itself. You peel off a liner and stick it down. It wouldn't surprise me if there are liner-less heat-activated tapes too, but I don't know of them myself. That's outside my domain.
I changed the board on my 3D printer today, the biggest hurdle was all the wire and plug manipulation. I play guitar, piano, type all day, you'd think I'd be good at it. It was like trying to pick up a grain of rice with two cucumbers.
When I work with this type of thing I use different kinds of tweezers (chip pullers, flat/round end, angled, etc). The ends of most FFC/FPCs are reinforced partly to provide good contact in their mating connector and also to make it easier to insert the cable with enough force without damaging it.
The spiky-ended tweezers will damage an FFC though.
All the manufacturing engineers I've talked to hate doing fluid dispense. Glue is in a class all its own because if you take the line down it tends to cure in the dispenser or nozzle.
I took apart my PS4 Pro several times since I own it. It’s very easy to do, nothing is glued, and the plastic clips that hold the cover in place don’t break when you unsnap and snap them back.
The only thing is an obnoxious warranty void sticker that delaminates if you don’t lift it carefully but even then I don’t think it’s legal to void the warranty with a sticker (Switzerland, bought from a EU retailer)
There is a interview of Elon Musk with MKBHD. He was precisely referring to this as being one of the issues when they were trying to automate the pipeline. The machines just couldn't plug a loose tube in its place. Maybe they should have consulted with Sony.
Tesla tried it because the technology was mostly ready back then. The difference is a PlayStation has vastly fewer cables, far simpler routing, and problems can more easily be fixed. People can physically pick up a PlayStation diagnosis the problem and then put it back, try to do that with a car.
Also, Sony has to hit a much lower price point, and they made about 10x as many consoles last year as Tesla made Model 3s. Both of those seem like they'd make the automation ROI better for Sony.
What does "lights out" mean here? Is it a way to describe a nearly fully-automated production line?
If that's the case, then aren't they using it for Model 3/Y already? If yes, then I have to disagree with you, because neither of those cars look anything like Cybertruck, even stylistically.
EDIT: thanks for explanation in the replies, it makes sense. If anyone has any additional insight as to how it affects production lines that are, let's say, 90% automated and 10% manual, your contribution would be heavily appreciated.
The pure definition of “lights out” refers to a fully automated facility where human hands never touch a product during the entire manufacturing process. In theory, a true lights-out plant would operate 24 hours a day, seven days a week, with downtime only for routine maintenance or repair.
I think what Elon learnt from that experience is that manufacturing things is hard, and that building that knowhow and ability to both build and troubleshoot robot factories like this one is not something you do overnight, even if you hire the best engineers.
Basically it’s unlike software and that might be where he stumbled.
Andy Grove used to talk about this a lot - that manufacturing know-how does not come easily, and that continuous improvement and innovation in manufacturing depends on continuing to actually do it.
Exactly. Things built for vertical assembly, like the Sony Walkman and the Apple Macintosh IIci, are easy to assemble automatically. All the assembly moves are straight down.
This has those annoying cable connectors that have to be slid in edgewise. Yet they got that to work.
I'm amazed that Sony let out detailed pictures of the cable connecting operation. Looks like they use both a camera and a force feedback wrist.
> Cables and tape are not as uniform and have a tendency to move in random directions.
I'm sure someone's probably already thought of this, but couldn't you deal with that by having two manipulators that hold the cable by both ends while keeping a small amount of tension on it? The cable could be picked up by dispensing it through a narrow U-shaped hole (to stop the loose end for grasping), or by gluing one end to a dispenser reel then having the one of the manipulators follow the cable to the other end like a person would pull a rope through their fingers.
Pick 'n place are usually only XYZ fixed-envelope cartesian systems with additional part rotation about one axis. Sometimes the XY is replaced with dual rotary units (SCARA), or XYZ are replaced with delta systems. See https://www.valin.com/resources/articles/five-types-of-indus... for a good overview here. Plugging in cables requires fine and reliable gripping at two ends plus additional degrees of freedom, and a clear path through a more complex motion envelope for access. This generally necessitates a move to arms, which come with additional issues (play at extreme positions, repeatability constraints, higher cost, power and space requirements, non-square envelope, slower motion, etc.) Having two of these work together on a flexible part means very fiddly parallel control of two systems with these drawbacks. Basically it's expensive because good end effectors cost a lot, parallel control exceeds conventional industrial requirements, industrial control systems are purposefully non-standardized, sensors have to be added in an application-specific manner and subsystems prototyped as requirements are identified, and experienced talent is hard to find.
I think it’s in part due to the pressure thresholds to put cables in, as too little pressure means the cable isn’t connected but too much could damage either the cable or thing it’s attaching to.
Maybe part of the success is in pre-packaging the wire and ribbon to act predictably, similar to how ribbons of surface mount components come on uniform tape reels instead of loosely in containers. You can crease or bend things to be predisposed to movement in certain directions.
> I'm super impressed that they elected to automate those things based on cost merit.
Seven years production run (or more), that should explain a lot. Even if hardware revisions changed significantly in that time (I wouldn't know) that's still a completely different calculation than something that changes annually, or even twice a year, as Sony's own Xperia phones used to.
Could you elaborate the cost merit thing? It was my main question and I couldn't find anything on it. How much do they save using robots instead of people to attach some cables? Is the benefit in failure ratio or do they actually expect to save some money by designing and building robots?
This is only tangentially related, but the dictionary isn't helping me distinguish between these two concepts. In networking what is the difference between bandwidth and throughput, if there is a difference?
The other comment regarding latency/throughput is a good place to start because those can often be in contention.
But bandwidth is not quite the same as throughput either, so I'll answer your question genuinely. Bandwidth typically refers to the actual channel capacity for a single link. Maybe that's your PCIe lane, your wifi channel or your HDRadio channel. The bandwidth is often measured as the spectral width of the channel in frequency. But those channels include some coding for error detection or compression.
Ultimately what matters most is what kind of throughput the channel can deliver. This is the metric that is often measured in bytes or bits per second. A one gigabit Ethernet card is intended to deliver one gigabit/sec of throughput with the Ethernet channel coding. So if you hooked up a test device to your NIC it should be able to drive frames through an otherwise unoccupied channel at one gigabit/sec. But to keep the example interesting, end-users would generally measure throughput with all the added layers of coding and protocol provided by a network stack. The term "bandwidth" is much more nebulous when someone uses it to describe traffic spanning various buses and network media. I'd take context clues and assume that this usage of the term is somewhat like throughput. Though perhaps you could consider the independent physical channels as a single logical one and extrapolate some kind of conceptual bandwidth?
Sorry, starting to ramble a bit at the end there but I think you may get the gist.
