I think about these things every time I cook, but it never occurred to me to just try it. I always assumed that the "reaction rate doubles for 10C change" rule of thumb meant it would take much longer, plus presumably cooking is endothermic.
I remember using something similar a long time ago - basically fuzz testing, I suppose you could call it property based testing where the property is "whatever edits the user does, we should be able to save and reopen the document without crashing".
It found so many bugs: file corruption, crashes, memory leaks, pathological performance issues. The kind of issues that standard unit testing doesn't find.
I would argue that the software quality of ADAS systems is very different from Infotainment.
Infotainment systems are a race to the bottom on BOM+SW price point. ADAS OEM's understand that there is a human cost, liability, and reputational cost for failure.
The real risk with these monoliths is when companies start to remove the distributed/redundant nature of safety critical systems, in order to reduce hardware costs.
There are multiple very good reasons for a distributed system in a car. However, irrespective of how clever your architecture is, there is only one good reason for centralized systems in a car and that is cost. It benefits no one but shareholders and C-suite.
OTA updates are sold as a key benefit but again it's marketing, they only reduce costs for the manufacturers and effectively remove a lot of the penalties of recalls. I would argue that difficult/costly recalls put pressure on manufacturers for 'first time right' design, OTA favours happy-go-lucky software.
I believe you, that these systems work. However, I watched too many videos about cars on which it doesn’t. Many of them were expensive cars. Phantom breaking is really scary to me. I’d rather have full control of the car, than letting some system randomly emergency brake the car for no reason.
ABS and ESP one can anticipate. ESP usually can even be turned off.
Statistics work on generic population but mush away a lot.
People are careless and inattentive beast of animals in our modern societies. Things are done for them, expected this way, they do not need to pay attention that much, which has lot of merits and advantages for the advancement of humanity. Dumb solutions doing as told and need to be handled expertly can be dangerous for modern people. Developing automation right (emphasis is here, big emphasis!!) is very necessary.
But unfinished and sloppy developers are killing careful people. Not show in the statistics, saving more bad drivers than killing good ones overridden by shit software cars.
Need to do it right with no collateral casualties.
I believe the tone of the conversations are into this direction anyway: please, pretty please, do it right! Not the current sloppy way! This is a dangerous game not mobile messaging platform, needs different mindsets than average software development approaches.
I have a new EV9 and the software on it regularly causes near miss accidents. I’m waiting to see accident statistics after it’s been on the market for a while. They’ll be dismal.
Examples: there are pedestrians nearby and the car is in reverse so it wildly swings the acceleration curve around, cycling between “slam on brakes” to “1-2mph” to “why am I on the other side of the street and standing on the break pedal?”
Even without pedestrians it constantly changes the acceleration curve and this cannot be disabled in a way that reliably survives turning the car off and back on.
Once, a motorcycle was lane splitting, so the adaptive cruise control tried to race it and accelerated at the car in front of me. I had to slam on the brakes.
This is by design: There’s a chapter in the manual explaining the dozen different reasons it’ll fail to regen brake at a stoplight (or, from what I can tell, rear end someone on the freeway) because of low visibility. (Including being on a hill, not going straight, pedestrians, vehicles in other lanes, and motorcycles lane splitting.)
Changing from drive to reverse to drive disables one pedal mode. I’m sure that’s caused at least one collision (when you take your foot off the pedals it automatically accelerates, especially in parking lots).
I’ve had it override steering too. Sometimes it tries to force me out of the lane. Sometimes it wants me to stay in the lane so it overrides emergency maneuvers when other cars try to merge into me.
The beeping is constant, and alert fatigue has set in. There’s even an undocumented alert icon that looks like “car is about to explode”. We don’t know what it means.
Even if they fix the safety issues, we plan to sell it for that reason.
Note that all of this idiocy is possible because they wired together the transmission, throttle, brake and vision systems more deeply than “we need emergency override”.
Anyway, mark my words: The accident rates on this model will be high.
You'd probably use some kind of MILP or CLP based model for that kind of thing, wouldn't you? The constraints define the search space and the solver algorithm then explores it.
I actually think a lot of people will want to talk to an AI that's available any time to listen to their problems and give them validation. Whether that's ultimately good or bad, I don't know (I suspect bad).
That's not really friendship, or at least it's just a part of friendship, but I think that's the part that AI is most capable of.
reply