96 replies
1d4h
The machine also automatically outputs a report that details any deviations during printing, removing the need for postfabrication qualification.
Anyone who's previously worked with 3D printing knows that this simply does not pass any kind of a sniff test. Both preventing and detecting internal defects is one of the, if not the hardest problem in 3D printing. There are many large companies trying to find ways to reliably solve just this problem alone. Saying that this method doesn't require any checks after production is simply false.
Titan submersible sound detection system vibes.
I'm not sure how many people know about their 'sound detection system'.
Carbon fiber is a fickle beast and is prone to such failures so that alone worried me, but it was incredible for me to learn that they expected some sort of early warning from cracking.
Well there were probably some early warning signs
The problem is that warning sign comes at around 10ms or less before the actual disaster
If it would be a Tesla, just enough time for the auto-pilot to disengage and claim it was the fault of the user.
Tesla autopilot is not supposed to be unmonitored, so any accidents under its control are entirely the fault of the driver. If a driver disengaged 0.1s before impact, they were derelict in their driving.
OSHA has a nice pamphlet regarding hazard identification and hazard controls. The least effective method of protecting workers is to put the risk on the worker to protect themselves. Your view has a similar vibe to "well they weren't wearing their hard hat and so it's their fault," a view that has been rejected across the board for safety in favor of the view that even letting the situation get to that point is a failure.
https://www.osha.gov/sites/default/files/Hierarchy_of_Contro...
Driver control IS making the driver responsible, by definition. Tesla is legally required to put the driver in control.
This is a driver assist program. There is no way such a program, that is subordinate to the driver and depends on the driver being in control, can protect the driver from not doing their part, and driving.
First, Tesla is not legally required to put the driver in control - they are free to indemnify the driver completely and shoulder all of the liability themselves.
Second, who do you think was ultimately the one at fault for the excessive radiation doses caused by the Therac-25 machines: the machine technician operating the machine or the machine manufacturer? If it isn't the technician then I don't understand your argument because you can just find/replace every instance of driver in your post with technician.
https://en.wikipedia.org/wiki/Therac-25
Therac-25 is more like a stuck accelerator pedal. The operator did not command the machine to deliver too much radiation, that happened outside their control.
And you think that a driver commands their Tesla to run into an object when they enable autopilot or FSD?
One is a failure to act as an agent - to control the car and make decisions. Another is a failure to even be a reliable tool - to do what the operator commands. Very different.
Creating Agents is a lot harder.
Coming up with new words and terms in order to escape the comparison smacks of This Time It’s Different. Therac-25 was an Agent for performing radiation therapy - it controlled the radiation machine and made decisions. Autopilot/FSD is supposed to be a tool for the driver and it fails to be a reliable tool by driving into things.
This Time It’s Not Different At All
Autopilot is NOT advertised as being able to drive perfectly on its own, and your disingenuous analogies won't change that.
Tesla is legally required to put the driver in control. They do not have permission to sell a self-driving vehicle for use on public roads.
Secondly, it is already established and disclosed that the driver assist cannot self-drive. This would be comparable to a Therac-25 machine being sold as being incapable of limiting radiation output, and then not limiting radiation output.
There is absolutely no other company in the world where a predictable and disclosed shortcoming of the driver assist combined with the failure of the driver to fulfill their responsibility to be in control, would be blamed on the driver assist program.
Both Musk and Tesla are far from perfect, but the lengths that people go to to attack them is obviously agenda-driven and a result of emotions.
Shame about the "Full Self Driving" branding
They can put whatever disclaimers in the manual but their branding is giving a different message. It's a message Tesla wants the customer to hear: "sit back and relax, the car drives for you."
The branding does not communicate that the driver needs to be just as aware and engaged as they would be if they were driving on their own, and be ready to take control of the vehicle at any moment.
Compare this to GM's "Super Cruise" branding. The message I get from that is "cruise control, but better." Cruise control is a long established feature, drivers have plenty of experience with it, and they know that it is definitely not going to drive the car itself. They know they're still going to have to pay attention because the car is going to do some of the driving tasks but not all of them. The car is making no implicit or explicit claim that it will drive for you.
"Full Self Driving" and related features like "Summon" make implicit claims in how they're named and presented. The driver absolutely has responsibility but Tesla is trying to play both sides of the coin with their branding vs their actual liability.
What the name implies if interpreted without any context pales in comparison to the repeated and explicit instructions and warnings given to the driver that clarify that the driver should always be in control
There's a reason "it does what it says on the tin" is generally seen as a positive aspect of a product.
When a product's naming and branding is well aligned with its actual utility, it builds trust with the customer. The customer doesn't feel like the seller is trying to pull a fast one on them.
Tesla chose their branding direction and Elon chooses to make his "optimistic" predictions that have made it sound like true self driving is right around the corner for years. That people take this impression away is no fault of anyone but those that put the impression out there in the first place.
If a driver is using any driver assistance feature they need to be paying attention all the time. Not only is it stated in all vehicle manuals, it is the intelligent way to use the features given that automated driving is still far from perfect.
Your analogy makes no sense given that the risk is always on the driver whether there are driver assistance features or not.
Unfortunately it is very difficult to remain vigilant when the automated system seems to be doing a great job. Eventually you will become distracted.
Even in accident cases, remember that the Teslas involved are not new - they have had dozens or hundreds of successful drives before the accident.
This is true, but it also falsifies Tesla's naming, promotion, abd advertising of the capability as "Full Self-Driving".
("Self-Driving" alone could be reasonable in certain contexts, but insisting on "Full Self-Driving" is a flat-out lie in plain language. Saying " Alan is fully capable of driving the car." means that he requires zero monitoring and/or intervention; same for the "Fully..." phrase.)
What the name implies if interpreted without any context pales in comparison to the repeated and explicit instructions and warnings given to the driver that clarify that the driver should always be in control.
Yup.
It's "we'll take the profits from selling it as something that it is not.".
While simultaneously they take every step to ensure that when things go wrong when it inevitably turns out to NOT be what they claimed, the entire burden and responsibility is not on them, but on you.
More conspiracy theories. The driver assist disengaging had nothing to do with a ploy by Tesla to shift responsibility. A driver taking control of the car 0.1s before impact is a result of inattentive driving.
Like I said in the other comment:
There is absolutely no other company in the world where a predictable and disclosed shortcoming of the driver assist combined with the failure of the driver to fulfill their responsibility to be in control, would be blamed on the driver assist program.
Both Musk and Tesla are far from perfect, but the lengths that people go to to attack them is obviously agenda-driven and a result of emotions
No, not even close, dead wrong spurious insults. For many years, I held Musk in very high regard, and you can find I've even posted extensive defenses of him here on HN. My changed views are a direct result of observed behavior.
No, the fact that the driver assist disengages is NOT what I am pointing to as the attempt to shift responsibility.
Tesla is selling and loudly insisting that they can use the term "Full Self-Driving". Tesla advertises and charges(ed?) 5-figure-USD amounts for fully autonomous self-driving capabilities that would allow the cars to be used in competition with Uber/Lift but without drivers. This money was charged for features that were supposed to be available in the previous decade. Yet now, in the middle of the 2020s, there are no such features available.
Moreover, they are so far from delivering such features that your argument is that any accident is fully the fault of driver attention being less than 100%. So, clearly, the features related to "Full Self Driving" or "FSD" do not exist.
Moreover, everything about the actual implementation, including instructions to be always vigilant, UI warnings about no hands on the steering wheel/yoke, alerts to re-engage, auto-disconnect, live data collection so Tesla can prove what mode was engaged, what level of control, etc., etc., etc. all point to the requirement for the DRIVER TO BE FULLY ENGAGED. They also ensure that the situation is such that if the driver fails to be fully engaged, it is the driver's fault/responsibility.
When the driver MUST BE fully engaged
(as you argue and I just documented),
the car is NOT "Fully Self-Driving".
The two conditions are by definition mutually exclusive.
The fact that Tesla is trying to have it both ways is exactly as I said:
They want the profits from the false advertising that it is "Fully Self-Driving", but ensure that when it turns out to be false, all consequences are on the customer.
There's nothing emotional about that fact pattern. I've even said above, that I think it'd be perfectly OK for Tesla to say "Self-Driving" as that does not directly imply that you can fully disengage because the FAD is "Fully" engaged.
I'm not the only one who thinks that language means something, even though Musk does not, and you think that calling out a blatant lie and possible fraud must be emotional.
Sorry for the length, if I'd had more time, I'd have written a shorter post
the lengths that people go to to defend them is obviously agenda-driven and a result of emotions
You don’t just get a pass, especially not when you actively are promoting it as soon™ being able to drive you without any input whatsoever. Robotaxis and the like have been promised for years now, it’s not like Tesla isn’t actively claiming the technology is basically around the corner.
In short, saying it’s “Full Self Driving (Supervised)” is not really enough.
You get a pass because nobody who drives a Tesla can reasonably believe the vehicle can drive itself. That it can't fully drive itself is fully and adequately disclosed.
So, what is the difference between your argument and this more concise version?
The falsehood of the "Full Self-Driving" claims are generally known to existing customers, therefore it is ok to lie to new customers.
You're completely right, but unfortunately people place too much trust in it.
It reminds me of the data entry days. You can have someone type a million table rows into a form and catch all typos but if you give the same person the same data and the same time without having them type it they find non of them.
Is there data, that indicates Tesla did this, or something like this?
Kind of? Autopilot has a habit of disengaging right before crashes [1]; which may not be a bad thing, your seat belt also has a habit of not being adjustable in a crash (dunno if an ICE engine will turn off).
Mix that with Elon had a habit of commenting on crashes [2] to keep good marketing about FSD. And like who cares if it was "Auto lane control" vs "Autopilot" that let somebody drive the car from the passenger seat but Elon made sure to let everybody know "Autopilot" wasn't engaged.
[1]: https://www.motortrend.com/news/nhtsa-tesla-autopilot-invest...
[2]: https://www.consumeraffairs.com/news/elon-musk-says-autopilo...
Your second link makes it clear that no version of self-driving was on anywhere near that particular crash, so that does not support the above comment at all.
If the only true part is that the system disengages at some point, then "kind of" is much too generous for that kind of rumor-mongering.
But if I ever see some real proof I'll spread it far and wide.
There were plenty of early warning signs. In a previous dive back in 2019 they had professional submersible designer Karl Stanley on board, who later wrote an email to OceanGate about the worrying cracking sounds he heard.
It's more impressive that the sub continued to work while giving warning signs for 4 years.
https://abc7.com/titan-submersible-2023-incident-titanic-oce...
They seem to have totally recreated the hull using a different manufacturing method in late 2020/early 2021: https://www.wired.com/story/titan-submersible-disaster-insid...
They seem to have run fewer tests on the new hull, though. From the outside, it looks like one of the lessons they learned from earlier tests was that tests can create bad news, so if you're optimizing for the best reports back to investors you should stop running tests.
I don't know how many 'test to failure' tests would have been required before I would have any confidence in the models but probably so many tests that the titanium alternative would have been far cheaper.
The other problem is that it cracks all the time and they get louder as they get deeper, so it's not just if it cracks it is if it cracks enough or more than expected. Which straw will break the camels back. It is just such an insanely awful metric.
On the hypothetical assumption that I had some faith in the models at what point will I have the life or death fight over whether or not that last loud crack was statistically significant.
Do you have a link? Google just gives results about the US Navy detection of the implosion…
https://www.youtube.com/watch?v=6LcGrLnzYuU
As I remember he talks about the sound detection thing.
https://www.businessinsider.com/oceangate-2020-boasted-incre...
https://www.wired.com/story/titan-submersible-disaster-insid...
Recent one-year-lookback story by Wired.
Exactly what I thought.
What does "vibes" mean here? Are you saying that this is a comparable level of negligence, or that it simply personally reminds you of the former case, and nothing else? There are better words available than "vibes".
To steelman the argument, maybe they meant "the monitoring system is good enough at catching defects that if none are reported, the engine will probably pass an all-up hotfire test of the engine".
Of course you could do a water or air-pressure leak test on the plumbing pretty easily, and you would likely do that on the first 30 engines...
But if you have confidence in your build process, maybe the juice isn't worth the squeeze on (say) a direct contact ultrasound void check on every square millimeter of the part.
It's all about "how expensive is it to run the test" vs "what is the likelihood the test catches an issue" vs "what's the cost of failing while everyone is watching?"
Same reason SpaceX went from dry-dress-rehersals to wet-dress-rehearsals to separate-static-fire-before-launch to hold-down-for-three-seconds-before-launch... The hold-down -before-launch is an integration test that covers everything the previous tests do, so eventually you can start removing redundant tests.
In rockets/aerospace, where the failure of any one of a thousand different parts means instant disintegration, checks are always worth the squeeze. Everyone talks about building rockets on the cheap and accepting a slightly-higher failure rate, but in reality that doesn't work. Even a tiny increase in component fault rates translates to total mission failure once multiplied across thousands of vital parts. The answer isn't to not check but to find ways to more efficiently and more thoroughly check each part. This is only more true if one considers reusable rockets where components will be expected to participate in multiple launches.
this might be 20th-century thinking. if you can build enough copies of a rocket cheap enough, maybe disintegrating a bunch of them isn't a showstopper
this works for munitions, but not for payloads that anyone cares about.
Not sure exploding rockets on their launch platform is such a good thing when they're carrying a bunch of highly explosive / fragmentary warheads (in addition to the rocket itself, which is plenty dangerous).
Unless this was something like a cruise missile dropped at altitude where a failure isn't a big deal.
If failure isn't a big deal, then the weapon should no have been used. An ALCM costs millions. The destruction it causes is part of a larger battle plan. Should it not work properly then friendly forces may die. Should it work properly then enemy forces may die. The effectiveness of such a weapon is never not a big deal.
If it’s that critical then they aren’t launching one cruise missile.
If system A costs 1k quid but only works 50% of the time, while system B works 99% of the time but costs 10k quid, system A actually makes a lot more sense. On average you are going to spend a lot less money for the same outcome.
What's the price on weapons technology falling into enemy hands? A dud is a gold mine of data for an enemy's reverse-engineers.
I'm sure even in this case, risking the air-frame of the bomber or friendlies on the ground below is not ideal. Bear in mind cruise missiles are usually launched from friendly territory.
Also, failures might reduce the accuracy of the missile, leading to potential civilian causalities.
The problem with that idea is that you won't be legally allowed to launch again until you root cause and fix the failure, which can take months (or years if you're Blue Origin). Also, your insurance rates tend to go up a lot when your rockets blow up regularly, which tends to push customers away.
In practice it doesn't work.
Notes: Astra said they were going to pursue this strategy. It was not well received by potential customers and they basically had to walk it back.
Doesn't work? It seems to be working great for the industry leader.
What are you talking about? SpaceX's Falcon 9 is arguably the most reliable rocket ever made. They've launched hundreds of times in a row without failure.
If your point is that they're blowing up Starship prototypes, well... they're in the middle of a development program and they're not flying customer payloads.
They blow up a few of the early Falcons, and the explosions got a lot of news coverage.
Now they are launching an landing[1] safetly, but it's boring and the news show a cute puppy instead.
So for most people the rate of explosions vs launching is skewed.
[1] It's crazy that landing rockets is boring now.
Only if you want to launch from US. I guess with a 160 countries there will some with the right area for a launch pad with way laxer requirements.
Good luck with that.
Exploding rockets are a very serious public safety risk since they're >90% propellant. Even Russia grounded Soyuz until they root caused its launch failure. From what I can gather China has the same policy, although they're very quiet about any launch failures that happen, so it harder to tell.
But it isn't about destroying a bunch of them. Cut corners on checks and you very quickly blow up all of them. Any slight increase in the failure rate of individual parts, saving a few pennies, multiplies exponentially across the entire rocket into total system failure. So the money-saving approach is actually to test test and retest, to cram down the failure rates so low that the cumulative rate become acceptable (about 1%).
Right, because as one of your rocket customers, I'm totally fine with you disintegrating my one-of-a-kind payload
from the rocket perspective, sure, but not so much from the perspective of your cargo.
In many cases the cargo is more precious than the rocket so you need reliable rockets.
Who pays when the payload disintegrates?
The issue isn’t the rocket: it’s what you have affixed to it (the payload).
The rocket itself is purely a delivery system for a payload after all.
what about payload?
ready to build many Webb telescopes?
So what you're saying is instead of inspecting the parts, you just try to launch the rockets and try again when they explode. Production is your testing ground. Statistically speaking, if a specific design succeeds a few hundred times in a row, it's probably sound? If it crashes, you just push a patch?
Day 1 patches for space travel?
Thanks, I hate it.
Well, if you build it robust enough, you can test less. Not saying testing is worthless, but sometimes a one-piece that used to be 45 pieces held together by rivets is just, much much more resilient.
It’s a bit of chicken and egg to know if you built it robust enough without meticulous testing of your robustness
Robust enough to not need testing == too heavy to be cost-effective.
My dad works in aerospace. In no field or world does anyone actually check every little part.
You create a process, you test that process so that you understand its limits, and then you make sure to follow that process.
Now I don't know if 3D printing rocket parts actually works and I have my doubts but this startup is currently testing the process and they will figure out its limits. That's the whole point of R&D.
For context, 3D printing rocket parts is incredibly common, even student teams often use printed nozzles and such. The linked article is about printing a whole engine as a single piece, which is a different beast.
This is simply inaccurate. You can spend months and hundreds of millions of dollars running scanning electron microscopes over all of your parts - no one does this, because it's not "worth the squeeze". The question is where to draw the line, and I have no idea what your opinion is there.
To make matters worse the failure modes don't only affect the launch vehicle itself. A failure of a rocket likely means a total loss of the payload. It also runs the risk of damage/loss of the launch pad, support structures, and hapless down range victims.
Rockets contain a significant amount of stored chemical energy, enough to get the payload mass into a stable orbit of the Earth. If you release all of that energy at once as an explosion it will cause a significant amount of damage. Rockets aren't something to goof around with and make assumptions about safety.
I’m curious about why that is. Naively, it seems like printing layer by layer would allow for a lot of inspection. Maybe even photograph each layer as you go?
And then test and go back to see what kinds of defects were apparent in the photos.
It's just that it's very easy for deviations from what the sensors see to occur. In traditional 3d printing, this can be stuff like a sensor switch wearing out, maybe physically moving slightly, being temperature sensitive, maybe the frame has changed shape slightly due to heat, moving the sensor a little, maybe something in the microcontroller happened to cause a slight delay in reading the sensor, or looseness developing in the motion system, or something being slightly out of alignment, or some component in the extrusion system experiencing momentarily higher friction and so on.
The layers are really thin, so manually inspecting them would slow down the printing process drastically. Then, ultimately, what even can you do if there's a defect? The layer has been laid already. If material is missing somewhere, you could have the machine go back and add it, but if there's excess material somewhere, or it's in a form that the machine can't fix, there's not a lot to be done, particularly in applications like rocketry, where your structural strength tolerance are very tight.
right, but this is a problem in any modern precision machining, and it has been (mostly) conquered to a degree that we can produce very precise things in an almost entirely automated fashion.
laser sintering is easier to audit than a normal fdm style print in a lot of ways if you care to take the time to do it. The process can be paused fairly easy with the right machine and right environment, the product can be weighed mid-process, it can have all sorts of vision and laser metrology done to the product midway through production; whatever -- and the mid print failure rate is astronomically lower than extrusion based methods.
it doesn't seem that unbelievable to me.
I know for certain that defects in the powder layer can be fixed in binder jet by redoing the recoater. There has been talk in the research world about being able to fix errors in L-PBF but I’m not sure they’ve gone past the research stage. The big point is that you can know a part might be out of whatever your acceptance criteria might be.
We aren't talking about deposition 3d printing for metal parts as it needs to get to the oven and precision is...not it's strong part
The most common form for metal precision production is powder bed fusion
You deposit a whole layer of powder and a laser melts the desired parts
Differential cooling is an issue, and is one that isn't apparent until the layers have already been printed.
If their process actually works, whatever they are doing isn't trivial.
Peregrine from Oak Ridge National Lab does exactly this. Lots of other research papers about it too.
You are correct, and with the current metal AM techniques (DED, SLM), you can also take thermal imaging of the melt pool throughout. From this you can a pretty accurate picture of the weld quality across the whole volume.
I'm not sure that's sufficient to eschew any other non destructive testing, but it is great information.
If any system could measure its own signal to noise accurately, it wouldn't have any noise in the first place.
This is a "truthy" statement that sounds right but is, in fact, false. Radio systems have an easily measurable SNR, but the noise cannot be eliminated.
You can measure is from outside the system itself if you know the original signal. Of course its possible to measure the noise externally if you know what was transmitted.
AFAIK this is incorrect. There exist modulation schemes where the received signal power is evident. For Wifi the signal is usually > 10-100x the noise floor and it is modulated with phase-shift keying. So there is no technical problem with seeing a noise floor, then receiving a much larger signal with constant envelope and calculating an SNR.
ETA: Just to sharpen the point, this is clearest when considering digital noise and error correction. It is easier for error correction codes to indicate a corrupted message than it is to provide the correct message.
That only works if you have an ultra-high-resolution record of isolated noise.
An accurate measurement of signal-to-noise is just some power levels and basic knowledge of the shapes of the signals. It's easy to figure out and doesn't help you read back your data. It helps you configure appropriate transmission settings and doesn't do much else.
If they’re not doing post fabrication validation, then the passengers are…
what's wrong with testing in prod? /s
Nothing. Always Be Testing.
But _only_ testing in prod is bad.
... test dummies.
You may go further and say the rocket would be built on a proper rocky foundation.
I too am doubtful. The market is littered with failed 3d printed products, all failed because the designers know nothing about real-life product design or because their deigns are too brittle or melt too easily in heat.
Check out Roboze on this. (disclosure: investor). They do a fantastic job with this, and in fact their 3d printing machines are widely used in aerospace and defense.
It might sound a little wild, but a huge amount of research has been put into getting metal AM parts to be “born qualified.” L-PBF is getting to be a fairly mature technology.
Yes and innovation curve tends to go down in large companies because of barriers which a startup is not bounded by. This is how future big companies are created.
Very Indian approach. Print a report, and call it a day. :D
Just out of curiosity, is your 3D printing experience on the consumer side? Or the professional/commercial side?
There are various types of non-destructive testing that could be integrated into a larger scale 3D printer for post-production analysis, and it's a direction I'm surprised there hasn't been more movement in.
They may be skipping small portions of the post-print qual. Or more likely, they're just more confident that the pieces being created are of a certain level of quality. It's a hype-y statement but they're probably trying to highlight that they're doing some sort of software-defined quals.