To riff off of Margret Atwood, men go to AI chatbots because they won’t laugh at them. Women go to AI chatbots because they won’t kill them.

Yeah, I was referring to spicy autocorrect in the more general sense of something that uses a faster statistical model to replace a slower theoretically derived exhaustive calculation.

The regular old FEM based models can be quite misleading and when I had the chance to dig into them some years ago, it made me vaguely anxious. Except that nobody trusts the existing CAE solvers, there’s always a process to verify that actually the structure does what you think it does.

Aerodynamics, at least the coefficient of drag, is actually really good for this because you can’t cheat the air and it’s mostly obvious when you screw it up. Which isn’t true for flutter or the more structural details.

So, yeah, there is that risk, that they’ll get high on their own supply. But thankfully the management already thinks that the current crop of CAE solvers are magical and so the credentialed professional engineers already know how to fight that battle for a lot of the structural details. (The long-suffering assembly line folk who are trying to assemble the airplane properly are, of course, a different matter and have had a lot less leverage)

Although, I’d also propose that there’s a second risk, which is that the current validation process is oriented towards the ways with which the existing FEM models screw you up and it’s likely that when the large physics model screws you up, it won’t be the way FEM models do.

Yah, I have some vague experience in the space and, without getting into things covered by NDAs, I guess I can say…

First, The popular media talks about the classic style of physics solvers as these magical black boxes but my experience is that they are sufficiently unreliable that I would never trust my life solely to the answers of a solver. They do provide very valuable feedback for refining a design without an endless hardware-rich cycle of destructive testing. Thus, I think that a large physics model is probably going to be the same sort of useful tool.

Second, while the CAE engineers can be very very protective over the time they spend on the two week cycle the article talks about, it’s fucking drudge work and a waste of a good mind. At the same time, the article does not really talk about some of the nitty gritty details. Aerodynamics is a great place to start because there’s less setup but the coefficient of drag is only one problem that needs to be considered.

Third, the good engineers can “see” things intuitively because things do operate with a pattern. Vorticies from protruding features… stress fractures from square holes in a beam… etc. This does feel like an area where spicy autocorrect can spicy autocorrect you to a useful answer.

Finally, cycle time for real world engineers is just like the cycle time for software engineers. Nobody wants to go back to the world where programmers submitted a deck of cards and got the printout back a week later.

The only real risk here is that somebody gets high on their own supply and decides that a large physics model is actually predictive and we don’t need the same set of actual physical tests that validate the models.

Well, think about it this way…

You could hit AGI by fastidiously simulating the biological wetware.

Except that each atom in the wetware is going to require n atoms worth of silicon to simulate. Simulating 10^26 atoms or so seems like a very very large computer, maybe planet-sized? It’s beyond the amount of memory you can address with 64 bit pointers.

General computer research (e.g. smaller feature size) reduces n, but eventually we reach the physical limits of computing. We might be getting uncomfortably close right now, barring fundamental developments in physics or electronics.

The goal if AGI research is to give you a better improvement of n than mere hardware improvements. My personal concern is that that LLM’s are actually getting us much of an improvement on the AGI value of n. Likewise, LLM’s are still many order of magnitude less parameters than the human brain simulation so many of the advantages that let us train a singular LLM model might not hold for an AGI model.

Coming up with an AGI system that uses most of the energy and data center space of a continent that manages to be about as smart as a very dumb human or maybe even just a smart monkey is an achievement in AGI but doesn’t really get you anywhere compared to the competition that is accidentally making another human amidst a drunken one-night stand and feeding them an infinitesimal equivalent to the energy and data center space of a continent.