3 ms·

> oh so messy and built by trial and error Not only built by trial and error, but also continuously adapted in near real time to deal with new sources of error

> oh so messy and built by trial and error Not only built by trial and error, but also continuously adapted in near real time to deal with new sources of error. The most complicated aspects of semiconductor manufacturing utilize statistical process control to determine the best course of action by relying on large sample sizes. You probably couldn't start up a modern manufacturing line without already having a manufacturing line due to this. Finding viable "hyperparameters" for a photo tool makes training an LLM look like a tutorial. Bootstrapping all of this required direct human involvement with ever-so-careful incremental offloading of these concerns to automation over a period of decades.

cxr 2y ago
> Finding viable "hyperparameters" for a photo tool makes training an LLM look like a tutorial. There's generally an unstated (and occasionally explicit, as in this case) reverence from software people for the kind of mythical engineering that goes on in fabs. In reality, if you've had any direct experience with the manufacturing process—and I'm talking about current- or next-gen processes for the most sophisticated mass market devices like those going into flagship smartphones, mining ASICs, GPUs, and critical applications like use in EVs—you know that a bunch of it is in the hands of folks whose most desirable asset in a prospective worker is that they'll accept low pay to eventually get the necessary work done to the prevailing standard best described as "adequate". Valley types especially, but even other software folks would be really surprised by how much of what goes on in fabs is basically the sort of thing that you would expect to see from people plucked from amateur hour. I've posted about this before on HN. Where improvement to existing chipmaker operations is concerned, the fruit hangs so, so low. Elon's biggest, dumbest misstep is not just buying Twitter; it's buying Twitter and not putting an equal or lesser amount of resources instead into gaining control over how his own (and others') chips are made—doing the same thing for the industry that he did with SpaceX for aerospace. Again, because it cannot be emphasized enough: what passes for acceptable in fab operations is bonkers.
- Teever 2y ago
  This sounds quite interesting. Can you link to some of your previous comments on this subject?
  - cxr 2y ago
    Here's one: <https://news.ycombinator.com/item?id=31065392 https://news.ycombinator.com/item?id=31065392>
CamperBob2 2y ago
Training an LLM is a rudimentary exercise at this point, so maybe not the best example.

cxr 2y ago

> Finding viable "hyperparameters" for a photo tool makes training an LLM look like a tutorial. There's generally an unstated (and occasionally explicit, as in this case) reverence from software people for the kind of mythical engineering that goes on in fabs. In reality, if you've had any direct experience with the manufacturing process—and I'm talking about current- or next-gen processes for the most sophisticated mass market devices like those going into flagship smartphones, mining ASICs, GPUs, and critical applications like use in EVs—you know that a bunch of it is in the hands of folks whose most desirable asset in a prospective worker is that they'll accept low pay to eventually get the necessary work done to the prevailing standard best described as "adequate". Valley types especially, but even other software folks would be really surprised by how much of what goes on in fabs is basically the sort of thing that you would expect to see from people plucked from amateur hour. I've posted about this before on HN. Where improvement to existing chipmaker operations is concerned, the fruit hangs so, so low. Elon's biggest, dumbest misstep is not just buying Twitter; it's buying Twitter and not putting an equal or lesser amount of resources instead into gaining control over how his own (and others') chips are made—doing the same thing for the industry that he did with SpaceX for aerospace. Again, because it cannot be emphasized enough: what passes for acceptable in fab operations is bonkers.

Teever 2y ago
This sounds quite interesting. Can you link to some of your previous comments on this subject?
- cxr 2y ago
  Here's one: <https://news.ycombinator.com/item?id=31065392 https://news.ycombinator.com/item?id=31065392>

CamperBob2 2y ago

Training an LLM is a rudimentary exercise at this point, so maybe not the best example.