00:00Thank you, Representative Ross. I now recognize my colleague from Illinois
00:04member of the full committee, Dr. Foster, for five minutes of questions. Thank you,
00:07Mr. Chair, and to our witnesses. You know, first off, it goes without saying
00:12that our hearts go out to all the recent victims of the recent floods. You know,
00:16it's a, you know, back in 2021, a tornado ripped through Naperville, Illinois, the
00:22town I live in, and by a matter of just a couple of minutes, lives were saved
00:26because of the prompt warning. That was only possible because the close
00:30coordination between the local government, units of government, and the National
00:34Weather Service that, that, you know, keeps an eye on things. And I was grateful
00:38actually last year to have an opportunity to achieve a childhood dream where I
00:43climbed up the weather, the tower into the radar dome of a National Weather
00:48Service facility at Lewis University Airfield, and see what's actually inside
00:53that, which is, for those of you who have never been inside those little
00:56spherical domes on the top of a tower, there's a parabolic antenna that's
01:00spinning around, and down below they have a klystron in a building, and talk to
01:04the guy that maintained the klystron. These are, you know, these are great
01:07facilities, and, you know, they're most of the time not used, and then they save
01:12lives from time to time, so it's just, it's incredible. Now, I'm also, I guess, one of
01:17Congress's few actual AI programmers, so I've just been fascinated by the progress
01:23that's made, made. And so, Dr. Gupta, you know, there, there's sort of two
01:26different approaches that have been very effective in this. The first one is just
01:29say, give me all of the data, all the sensors we have out there, and predict
01:33tomorrow's sensors. And you can do that with almost understanding no physics, no
01:37chemists, atmospheric chemistry, or anything. Or you can go to a hybrid approach,
01:41where you use the, what's known about the physics of a little blob of air the way
01:47they, they used to solve them on supercomputers. And it seems like, if you
01:51look at things like protein folding, the best approach is to do a hybrid of both.
01:56Where you have AI supervising it, or are you an AI purist, where you just say,
02:00forget about the physics, you know, um, you know.
02:03I would say this is, this is the thing with, like, respect to, like, the scaling laws in
02:07artificial intelligence. Even in the protein folding domain, if you look at the new
02:12research that has come up, the share of, like, physics goes down, and share of data,
02:17and the modeling goes up. So in general, this is true, I would say, even in the
02:23weather modeling space. You don't, I mean, there is definitely a role and
02:27opportunity to use these models themselves to understand the physics better.
02:32There is a lot of interpretability research that needs to be done to
02:36understand what these models are able to learn from data that we don't fully
02:40understand from a physics perspective. That's happening with respect to
02:43language and other domains as well.
02:46Now, and one of these, the things that this, um, committee has oversight over, is
02:50the split in effort between, sort of, state-of-the-art AI machines, which are
02:53going to, to smaller and smaller data widths. You know, as small as single-bit data
02:58paths, but certainly less than 8-bit data paths, which is contrary to what we've
03:03traditionally been doing, you know, 32, probably 64-bit floating point. And the
03:07question is, what should our investment be shifting down to these narrower data path
03:11architectures for weather applications? Or will there be always a use for the
03:17high-precision part of it?
03:20I think the high-precision stuff comes from, like, the physics approaches, because you
03:25can't solve those equations unless, and if you lose position, because you kind of lead to catastrophic errors.
03:33AI machine learning approaches, the current ones actually, I mean, the high-precision is good for them as well, but it's just, like, much cheaper to scale it up.
03:42It's a waste. Yeah, it's a waste. And so the trend in pure AI for commercial applications is toward the low-width data paths.
03:48Yes.
03:48So the question is, are we going to get two branches of supercomputing, one of which, the high-precision one, may have to remain federally supported, because it won't survive commercially. Do you have any...
03:59When you said that the trend is more toward use AI and not use mathematics and physics, is that something that we should anticipate should be reflected in government and supercomputer investment?
04:11I would say this is true. I have to accept, like, my point of view, but even the previous chairman of ECMWF, the European agencies, Peter Bauer, does agree with the fact that operationally the AI methods are likely going to continue to improve and are the future of how operationally the systems are going to work.
04:32But physics still has a role. We need to understand how these things behave, what other observations do we need, how do we actually improve our understanding.
04:41So these things are likely going to remain in parallel for a long time, but operationally, what we actually need is provided by the AI models already.
04:52Yeah, well, it's a brave new world. You know, everything you're, you're so proud of, of understanding the physics turns out to be irrelevant. You just throw it into machine learning and get a better answer. So, um, now I'm out of time here, but really appreciate this discussion.
05:07Thank you, Dr. Foster. And I'm not sure what else may be on your bucket list, but if flying into an eye wall is, uh, is on there, I'd love to have you down to our district to ride out with a hurricane hunter sometime.
05:17Okay. I'll ask my wife about that.
05:19Yeah. So I spent an entire career in the Navy trying to avoid storms when I was flying, but they, they go charging right into it, which I admire.
05:25Well, yeah, I'll pilot a drone into an eye.
05:28There we go. Let's see. All right.
05:31Let's see. All right.
Comments