6,105 karma · joined March 13, 2011
I have yet to find a way to retract a misclick on the "flag" button, much to my annoyance.
* Cars, because those have millions of units with distributed inattentive owners that can't do stop-the-world refits
* Rockets, because they're disposable and built-and-used serially rather than in parallel
* Hydroelectric dams and skyscrapers and bridges, because those must be 100% bespoke (due to terrain or artistic flourishes) and have no replaceable parts
* Datacenters, because those are either non-safety-critical (AWS et al) or bespoke (ADP et al)
* Hospitals and corporations and governments, because those are 100% bespoke (due to organizational "terrain" i.e. individual leaders)
* Satellites, because those have *no* maintenance
* Oil refineries, because those are 100% bespoke
* Roads, because those have no replaceable parts and no catastrophic cascading failures
Large, high-production-count ships might have worked, things like container ships and warships, but those have very few catastrophic cascading engineering failures - ships that are big enough to be useful (that is, that work like reactors rather than cars) are lost in wars or because they get driven into rocks, not because a single turbine blade failed while the ship was doing a triple reverse backflip over the international date line.Semisubmersible mobile offshore drilling rigs (i.e. oil platforms) might have been another good comparison. I'll have to go look up stats on those; they look to be in serialized production and with double-digit unit counts per design, and vulnerable to the same kinds of subtle and complex cascading failures that nuclear reactors are vulnerable to (e.g. https://en.wikipedia.org/wiki/Ocean_Ranger#Causes_and_effect...). That said, I don't know how many of them sink, and it looks like they don't operate under anything even vaguely similar to the same safety and reliability regimes that reactors do, nor does the oil industry improve at all in response to failures.
For the purposes of safety and reliability as a function of serialization, nuclear reactors and airplanes are more alike than they are different.
(and, like, really, materials are your argument? They both use turbines, they both use hydraulics, they both care critically about fluid flow, they both require active stabilization and fail deadly if uncontrolled, they both require continuous maintenance from dozens of trained specialists, they both have corrosion and fatigue as primary concerns, etc etc etc.)
I went and did looked up some stats from aviation, since that's an industry that would be able to distinguish them and operates in a comparable safety-and-reliability regime. Here's a graph of count of active Boeing 474s by year: https://www.iba.aero/insight/evolution-of-the-boeing-747-fle.... Pay attention to the bright blue showing the count of active hulls. Now compare that to the rate of 474 hull losses: https://en.wikipedia.org/wiki/Boeing_747_hull_losses. Note that the rate of hull losses is basically flat - there were just about as many 474s lost in the 1970s, with only 200 hulls active, as there were in the 2000s with 1200 hulls active, and as there were in the 2010s with 600 hulls active. At worst logarithmic, and I wouldn't be surprised if that's the actual way that failures are probabilistically exposed by increased hull count. I also didn't do any work to exclude "ate a SAM" and other causes that are independent of the airframe. Either way, in a healthy, scaled-out fleet, failure rate is a function of the design much more than it is a function of the fleet size.
Plus, y'know, the part where 1 GW of current real world nuclear power at the current real-world failure rate is safer, quality-adjusted-life-years-per-megawatt-wise, than 1 GW of coal or oil, even before you take higher-order effects of greenhouse gas emissions into account.
Similarly, the cost to decom a reactor is paid once per design, and scaling up would amortize those costs across the fleet rather than across a single unique reactor.
For an illustrative example, imagine that a part on your tractor suddenly becomes a bit of twisted, wrecked metal.
These days? You ask someone else with the same tractor to send you some pictures, pull their part out and measure all the holes and their positions and tell you what the bearing surfaces are made of, you scrape off some slivers of your wreck and assay them to figure out what alloy it's made out of and how it was heat-treated, you go and machine a test piece or two on your mill and try it, you buy a bit of bronze bushing off mcmaster and it works, etc. Maybe someone else has already done this and published a CAD model and instructions! We have solutions for this.
Back then? You have no designs. You have no way to communicate with other people who have a complete working part. You might be lucky enough to remember what bits the part connected to if you'd looked at it before. You may have some idea what the intended geometry was, but you're never going to figure out that the geometry is actually just a few degrees off right angles because that keeps you out of a complicated kinematic singularity that blows up the part every time you turn left. You can guess at some of the materials, but systematic classifications of iron-carbon alloys don't exist so all you can say is "it's some kind of steel". Even if you figure a lot of this out, there's no real way to write it down or share it with people because a lot of the terminology for representing this stuff simply doesn't exist. Even if you make a functional part, its expected lifetime is months rather than decades.
The original creator, by comparison, can just buy a new block of their supplier's "Alloy Number 5", throw the forging dies back on the presses and stamp out a blank, slap the blank into the jig that presents every hole to the drill press at exactly the right angle, put the part through their proprietary heat-treat process, and then ship you the result. Those dies and jigs and heat-treat protocols are exactly the same kind of proprietary protection that a git repo full of source code is today.
If you look at hydraulic and pneumatic fittings, you can see elements of this lock-in still extant in the modern world. Everyone had their own thread sizes and thread geometries for their own tools, getting them to interoperate was impossible unless you had the right adapters, and you could not make the right adapters unless you were a dedicated specialist in hydraulic fittings. You were locked in to your tools provider's pneumatic toolchain in exactly the same way that people are currently locked in to their current power tool brand because they own $50k in tools and battery packs that agree that they're made out of a specific manufacturer's 18650s with specific voltage drop and internal resistance and etc etc etc. (go watch the torque test channel's segment on battery adapters a bit, it's enlightening! - https://www.youtube.com/watch?v=WgJI8Ikrd6Y).
(if you watch enough of Forgotten Weapons, you'll also hear things like "X bought the design for Y gun but it took them 3 years to transfer the tooling and get it up and running" - that's the stamps and jigs and heat-treats and everything that I was talking about earlier, and if those tools were destroyed a gun generally became totally unmanufacturable and had to be redesigned from scratch to continue production, and those redesigns frequently failed to perfectly reproduce the original!)
- Improved FEM makes modern crumple zones deeper and smoother
- Heavy use of composites dramatically improve the raw amount of energy absorbed by the car instead of the passengers
- Better manufacturing methods mean parts can be shaped/placed to be less likely to enter the cabin
- Better placement of more airbags means you're more likely to hit something soft
- Better seats and better placements of better seat belts means you're less likely to snap something due to whiplash
- Better brakes and steering geometry reduces your chance of losing control
- Better brakes and steering geometry improves your ability to avoid other vehicles
- Electronic stability control _massively_ reduces chances of rollovers
- Electronic stability control _massively_ improves your ability to dodge other cars in emergency maneuvers
- Electronic stability control _massively_ improves your ability to handle bumps and potholes and waves and black ice and other disruptions
- Electronic stability control _massively_ improves the car's ability to go where you point the wheels regardless of speed or condition
Seriously, especially if you're driving a truck or SUV, you _need_ electronic stability control.https://en.wikipedia.org/wiki/Electronic_stability_control:
> According to the U.S. National Highway Traffic
> Safety Administration and the Insurance
> Institute for Highway Safety in 2004 and 2006
> respectively, one-third of fatal accidents
> could be prevented by the use of the technology
> Additionally, SUVs with stability control are
> involved in 67% fewer accidents than SUVs
> without the system
> The IIHS study concluded that ESC reduces
> the likelihood of all fatal crashes by
> 43%, fatal single-vehicle crashes by 56%, and
> fatal single-vehicle rollovers by 77–80%.
> ESC is described as the most important advance
> in auto safety
Not "the most important since seat belts", not "the most important since air bags", not "the most important current advance". Electronic stability control is THE most important safety feature in modern automotive vehicles, the single biggest factor in whether you will die because you got in a car _period_, more than air bags, more than seat belts, more than crumple zones, more than _headlights_ or _tire compounds_ or _safety glass_. Preventing an accident entirely is better than any mitigation and ESC is incredibly good at preventing a huge class of accidents.And those numbers are from 2005-ish. Modern active rollover prevention and traction control systems and anti-understeer/oversteer controls are _even better_.
At our current level of logistical sophistication, the benefit-to-the-patient is that they can get their medications at all.
I agree that in a perfect world every prescription would start with a computation, based on body weight and historical susceptibility and suchlike, to determine precisely the right dosage. However, I also don't think we're there yet. I don't think that you can reasonably prescribe "take 6 milliliters of this" or "take thirteen of these"; patients would mess that up reliably even if they were in perfect health and had perfect vision and perfectly stable hands. And we can't stock thirty different sizes of pill; each size of pill takes up already-limited shelf space, increases cost of packaging and logistics, and increases the likelihood of dosing error. So doctors can't yet prescribe precisely the dosage they should be able to. I'm sure that there are cases where they do - drugs with narrow therapeutic ranges administered in hospital settings where they can be precisely dispensed by IV - but for stuff that's being sent home with patients, we're just not there yet.
(Sadly, even on-demand services like pillpack won't save us, because doctors can't depend on the availability of custom pill-stamping when they make a prescription - they have to assume that the lowest-common-denominator pharmacy is being used to fill the prescription.)
(In fact, going by the rate at which hospital mortality is attributed to dosing errors during administration, we haven't solved precision pharmacology even under ideal circumstances and fundamental breakthroughs are required that will obsolete any current approach to dispensation of prescription medication.)
I wouldn't be surprised by AGI tomorrow, which means I have to consider the possibility.
We have, at this point, an extremely good idea of how matter and energy behave. We also know roughly how long it takes to build physical objects and to do R&D cycles with physical prototypes in the loop and how much progress you get out of each prototype cycle. We have a number of good ideas for what a ray gun would look like and we know where we need advances. We've probably already made all of the necessary fundamental breakthroughs, things like laser diodes and capacitors and optical sapphire. We even have some deployed, just at artillery size and less-lethal energies rather than handheld blaster power densities. If you asked DARPA how long it would take them to field a ray gun for the US Army they could give you a date. So I would be extremely surprised if deployment was announced today. I would go looking for fundamental breakthroughs in battery tech and laser media that I had missed.
We can't even agree what AGI is, even on an "I'll know it when I see it" level. We have no idea how to get there. R&D cycles are days or weeks long and produce unpredictable and sometimes significant advances in observed functionality. Fundamental breakthroughs come unpredictably but probably about once every two to five years and their effects are even more unpredictable than the effects of individual R&D cycles. Even worse, we have no idea what required breakthroughs are left or if there even are any. So, yes, I believe that I would not be surprised by - and therefore must consider a future containing - one of these AI researchers glueing the output into the input in exactly the right way or something and producing a living, thinking being in their computer tomorrow.
I cannot concretely refute the possibility of AGI Tomorrow the way I can argue, in specific detail, against Ray Guns Tomorrow.
(There is, of course, the possibility that I've missed some breakthrough that would enable Ray Guns Tomorrow, the same way the people who pooh-poohed heavier-than-air flight the week before the Wright Brothers' fight missed gasoline engines and aerofoils and did their calculations based on a dude flapping wings with his arms. This is why I say that I would be extremely surprised, not why it's impossible.)
Also, if your auto-keyboard tool was hooking in using X debug extensions like xdotool did, your tool was broken and wrong in the first place. The correct way to submit synthetic input events is to build a synthetic input device. uinput makes this shockingly straightforward: https://kernel.org/doc/html/v4.12/input/uinput.html#keyboard.... There are even high-level bindings (via libevdev) for high-level languages: https://gitlab.freedesktop.org/libevdev/python-libevdev/-/bl....
And speaking of that, let's not forget about downgrade attacks! Removing HTTP is the least worst option in much the same way that you disallow RC4 when negotiating encryption for an SSH connection.