The Apollo Guidance Computer had 4KB of RAM and 16KB of ROM. Can you write a scheduler that would execute 8 programs related to landing on the moon in 4KB?
An asynchronous scheduler with task priority, interrupts, and real-time response. That goes on a spaceship with less than 70 Watts of power available for the computer. That weighed about 70 pounds. When there has never been that sort of a multi-tasking computer controlled avionics before.
At a time when the contemporary computer systems that used asynchronous, priority-scheduled multitasking operating systems were things like IBM OS/360 mainframes and DEC PDP-6/PDP-10 minicomputers.
It was like having an Apple II in mid 1960s with double the word size.
And with a ROM that had to be hand sewn. No flash updates, just "little old ladies" working in a former bra factory for a defense contractor (and huge tax payer expense if any mistakes needed to be debugged).
Every 1970s videogame has some sort of scheduler that moves actors, draws the screen, and play sound. Depending on the computer, it might be very easy (some could interrupt when the VDP got to a specific spot) or require you to count cycles. If, for instance, your game is for an Apple II and you have background music, you need to hit the speaker at the frequency of the currently playing note WHILE you move things, do physics, and redraw the screen. And you need to count cycles, because there is no reliable counter you could use (I remember how awesome it'd be if I could write 0 to a memory location and read it later to know how many clock cycled passed, or one I could read to get which scan line of the image the video hardware was outputting).
I mean, the implementation can be incomprehensibly impressive, but as an example of “man-rated” — being sufficiently robust — rejecting the request would be sufficient and I suspect this was an obvious place to have some kind of handling for. If there were no handler of any kind in this position, it’d be unforgivable in modern, far less robust software.
The kind of example for man-rated I was expecting would be dealing with really nutty edge cases (perhaps like correctly operating under arbitrary bitflips) or the LISP recovery system from Deep Space 1.
Handling an expected error condition under normal operation is not, in my mind, a good example of man-rated robustness. It might be a good example of other things.
You make it robust by making it simple. Simple to implement, simple to reason about, and simple to debug. They were not only programming an existing computer, they were also more or less defining the computer and this allows some hardware-software co-evolution. In cases like these, the hardware is frozen well before the software is close to ready, so you might need to do some workarounds.
> Handling an expected error condition under normal operation
Nothing in the Apollo program was "normal operation" by any standard. The thing is the alarm elegantly - and correctly - handled a situation that should never have happened in the operation, that only happened because the simulators the procedures were developed against did not accurately account for the workload of having both docking and landing radars on at the same time (IIRC, this was the issue).
This kind of robustness in the face of unexpected misuse is required if you want to qualify your hardware as "man-rated".
That’s the missing bit in the original post; if exactly 7 tasks was intended to ever be possible to run (as in, it logically doesn’t make sense for an 8th to exist), and they tossed in a scheduler anyways, then that’s man-rated thinking.
I had read the setup as the 8th task being user error (you have 20 things to do, I can only run 7 at a time, and it’s an user error to submit the 8th task while the 7 are running)
Necessity is the mother of invention. The shortsighted protections put on chips, etc., by the US has forced Chinese AI industry to adapt or die. Guess what their response to this fitness function has been? Kudos to Z.ai on their inventions and excellent write-up, which reads like humans wrote it.
Wouldn't it be refreshing if OpenAI and Anthropic were this open, and spelled out how they were using their own models during development and rollout?!
All I can recall reading from OpenAI about what they have actually done in the name of "RSI" is using one of their models to help automate the training process.
True - OpenAI did at least say they used their models to help design their Jalapeno chip, but AFAIK zero details on how they are using their models in their software development process other than to automate some part(s) of training.
Ziphu seem much more matter of fact about it. To me it' a shame that they've decided to the use this "RSI" name, but at least they are being fairly specific about what they mean by it, while the western companies seem to want to invite you to think it's more than just dogfooding and automation.
I intellectually understand the retailpocalypse replaced by online shopping, but I will always want to try on clothes before buying them, and that goes double for shoes. After trying out New Balance for a year or two in adulthood I ended up going back to Nike running shoes (I still remember my pair of Air Icarus from high school, so light), just seemed to fit my feet better.
Hopefully most people here getting ribeyes aren’t getting them from the same factory farms that supply McDonald’s and sell other cheap kinds of beef. A ribeye is an occasional luxury and best bought from the farmer.
The industry has mostly consolidated packing and distribution. McDonald's is mostly selling trimmings from cows that also produce high-end steaks.
Surprisingly enough, they hold most beef processing plants to higher standards than the rest of the industry (they're still buying cheap trimmings, don't get me wrong), because they do so much volume, basic probability tells us that most food poisoning outbreaks would hit them otherwise.
Sure, you could find a farmer that raises cows and butchers them on site, but that's rare, and the easiest way to actually achieve that is to buy a whole or half cow at a time.
Most small farms don't bother with that: they raise it from calf to adult on grazing land, then send it to a feed lot where the miracles of modern science significantly increase the cow's weight before slaughter. Cows that are still producing calves get retained to grow the herd next year.
If you want to pay extra for this, look for "grass-fed, grain-finished" beef. Most unlabeled stuff in the US goes through that funnel though.
Back when McDonalds was first rolling out their standardized restaurants and menu, it was pretty common for restaurants to buy fresh ground beef from local sources, some of whom were less clean and less careful about temperatures and selling old meat. It was also common for them to mix in old meat with fresher meat to try to "hide" it. Since it's all ground up, it's hard to tell.
So they developed standards and demanded that all their hamburgers be formed, frozen, and packaged by the processor and kept frozen until cooking. It's really the safest way to handle that volume of ground beef through the distribution chain.
Apparently their 1/4 lb burgers now use fresh beef, not frozen. Not sure how that's handled but it sounds more risky.
Too bad. I get mine from the small town butcher/locker when I’m in the Midwest in summer, and the neighborhood farmer’s market on the east coast the rest of the year. A couple of ribeyes a year and I’m good, so I see them as an affordable luxury where I can pay for good prime meat from a short supply chain.
Yes. And there are a ton of companies that sell just loose stones, so you can take it to a locally owned shop for the ring and setting, thereby keeping more of the money in your own community.
reply