3 ms·

Yeah, it's not great. I'm kind of in the middle of this kind of thing at the moment, and explaining it all the way - from where I am, at least - would require s

Yeah, it's not great. I'm kind of in the middle of this kind of thing at the moment, and explaining it all the way - from where I am, at least - would require suspension of disbelief at the sheer amount of waste that exists today. AI's a perfect tool for generating a lot of the paperwork because the paperwork's just MIL-STD talky words in the first place. Unfortunately, 5% of it isn't, and that's kind of a problem, because like with civil engineering it's not fault tolerant at all. This is the part that's taking a lot of work right now, just the identification of critical bits. To do so we need a complete transparency from industry CAD and MATLAB environments all the way down to how they pack up the mechanic toolbox and what sort of staff we got. The LLM training works hand in hand with simulation, see? It writes and tests in concert. But that's a tall order because, well, an awful lot of companies don't know how to make solid models right, or what they pack into the box doesn't have anything to do with the lists in their ERP, or they send out the janitor as a Level III Tech. This "universal simulation" thing all rides with the fashionable but reified ideal of "Digital Twin", which hopefully doesn't end up providing a real-world proof of Goodhart's Law. As if you couldn't guess by now, I'm skeptical about this thing, but I'm also extremely hopeful. If there's one theme to my professional career, it's been to clear out as much meaningless crap from MIL-STD work as humanly possible. These LLM "AI" people are at least on the right side . . for now. They might eventually get co-opted into the big Goodhart Machine so that they can make powerpoints about powerpoints about powerpoints about powerpoints about . .

swader999 3y ago
The irony is palpable, we will end up with a Powerpoint presentation meticulously outlining how to avoid falling victim to Goodhart's Law, only to realize that by making it the focus, we've just proved Goodhart's point.
- Nevermark 3y ago
  So we just need a task force to come up with a system to simplify the system… And since we don’t know how to do that, we better bring in the consultants… — What all these things have in common is reactive top down control of processes, with imprecise measures & slow feedback. Instead of continuously handing down responsibility to individuals & teams for pushing up process efficiencies & change - based on ground level truth, needs, opportunities, and creativity. With real time feedback.
ooterness 3y ago
Where I have seen digital models work well is in the initial planning phase, where you don't even know what you want to build yet. Iterating on abstract requirements in a simulation is faster than iterating on prototypes, but it can still get you (some) of the necessary intuition. Where it goes off the rails is when management drinks too much of the kool-aid, and gives all the R&D funding to the model-based-system-engineering experts. Then they start building models for every other domain without consulting those domain experts. (They're the experts on models so obviously those other folks will just slow things down, pointing out ways that theory and practice aren't quite the same.) Then it's all garbage-in, garbage-out. Turns out that the map is not the territory.

swader999 3y ago

The irony is palpable, we will end up with a Powerpoint presentation meticulously outlining how to avoid falling victim to Goodhart's Law, only to realize that by making it the focus, we've just proved Goodhart's point.

Nevermark 3y ago
So we just need a task force to come up with a system to simplify the system… And since we don’t know how to do that, we better bring in the consultants… — What all these things have in common is reactive top down control of processes, with imprecise measures & slow feedback. Instead of continuously handing down responsibility to individuals & teams for pushing up process efficiencies & change - based on ground level truth, needs, opportunities, and creativity. With real time feedback.

ooterness 3y ago

Where I have seen digital models work well is in the initial planning phase, where you don't even know what you want to build yet. Iterating on abstract requirements in a simulation is faster than iterating on prototypes, but it can still get you (some) of the necessary intuition. Where it goes off the rails is when management drinks too much of the kool-aid, and gives all the R&D funding to the model-based-system-engineering experts. Then they start building models for every other domain without consulting those domain experts. (They're the experts on models so obviously those other folks will just slow things down, pointing out ways that theory and practice aren't quite the same.) Then it's all garbage-in, garbage-out. Turns out that the map is not the territory.