America’s Last Top Model | Hacker News Reader

America’s Last Top Model

99percentinvisible.org

157 pointspanic43 comments

On a related note, there is a model of the Bay Area by the US Army Corp of Engineers located in Sausalito:

http://www.spn.usace.army.mil/Missions/Recreation/Bay-Model-...

Ah, I remember this model from the episode of Mythbusters where they escape from Alcatraz. They used the model to test potential routes across the bay.

Yes, that's still open. It's now more of a museum than a working hydraulic model, though. The Corps of Engineers maintains it, but most of their analysis of the SF bay is done on computers. One of the big issues is salinity, as in how far up the Sacramento Delta salt water will penetrate during a drought. The physical model can't do salinity; the mixing doesn't scale linearly.

I don't have a source, but someone said that we don't crash cars for crashworthiness, we crash cars to validate the computer models we have for crashworthiness.

I don't completely agree with that statement, but I do think it's interesting.

It's more than just validating the computer models, it's also the manufacturing process, and the car company etc. In the end when a third party is crashing a physical object that's much harder to fake.

Exactly, that's why I don't completely agree with the statement.

If you model with high strength steel and the manufacturer or supplier make a mistake or substitution, then your model doesn't represent reality.

The crash tests validate reality and the model at the same time.

That's a classic example of Goodhart's law. Regulations call for extremely specific tests to measure crashworthiness, so vehicles are engineered to pass those tests (front-end impact with a specific wall shape at a specific velocity and angle, etc), so they get a "five star crash rating" but at the expense of real-world crash performance.

What else could regulations call for than a defined set of tests? You need comparability between tests or the tests are worthless when it comes to informing customers. Real world crashes differ wildly, so it's impossible to use those as a measure. There are also many different crash tests carried out around the globe and a car performing well in many of those will perform well in the real world.

Goodwin's law outlines a paradox. Practically, there is no perfect metric; every metric has a limited shelf life.

Use a defined set of tests along with real world metrics. [1]

Ensure your tests match reality as close as possible.

Ensure you test for the most important events found in the real world. In the case of crash tests, make sure you test for crashes that often cause death or serious injury.

Real world metrics get really tricky and expensive, but when speaking about life and safety, it's probably important enough to measure.

It's also very important to be smart about measurement and reporting of events, statistics, and trends.

1. http://lesswrong.com/lw/1ws/the_importance_of_goodharts_law/

But that's exactly what happens: crash tests get validated against real world accidents and refined all the time.

Is that a problem? That sounds like a solution... but it's a continuous solution, not a solution that you apply once and forget about.

> at the expense of real-world crash performance.

Citation? I'm pretty sure that the crash ratings system has resulted in better real-world crash performance.

It's theoretically possible that crash-tests overshoot the optimum and thus effort that could be used for better real world performance gets wasted.

It's possible. I'm not sure it's likely.

Even if it is true, we're still better off having cars optimized for crash survival even if the optimization parameters are imperfect.

> so they get a "five star crash rating" but at the expense of real-world crash performance

Do you have any source for that?

I worked a company what made a popular CAE (Computer Aided Engineering) package. It was really fun to see the connection between physics, material science, mechanical engineering, and software development all in one place.

As an intern I remember being fascinated by the ability to build a part, generate a mesh, assign a material to it, assign forced to some parts of it, "solve it" and then watch it deform.

It was also fun to use Irix, AIX, HP UX, and SunOS, but the switch to Windows had already started. It is funny to think that at some point, Unix workstation were the top graphics workstations available...

Surpising that these physical models can outperform computers! I would have thought surface tension and gravity would mess up the scaling in the physical model. Is the problem in computer simulations that we don't have the equations, or that we don't have the computing power?

Both.

In terms of computing power, think of the Butterfly effect: small changes in initial conditions produce enormous differences in outcomes.

Also, the models aren't constrained by what we think the equations should be. In a sense, the equations are just a model of a more abstract kind rather than physical.

Of course for the reasons that you listed physical models aren't perfect either.

Scaling is a very hard problem[0], but it's easier in these sorts of fluid systems where the materials have a simple Newtonian viscosity and the temperatures and other parameters don't change dramatically throughout the model. Contrast this with solid Earth or (I think?) atmospheric modeling, where those changes must be represented correctly by different materials. Nonetheless, for lots of problems in earth science, new computational models are still benchmarked to analog models.

The other problem with analog modeling is that can be pretty hard to measure the full 3D velocity field, stresses and strains, and that sort of thing, which are explicit in computational models. Computer-aided particle tracking has made this far better for analog modelers in the past decade or two, though.

[0]: A great reference for this is M King Hubbert's 1945 paper in which he considers whether it's possible to pick up the great state of Texas with a crane: https://www.depts.ttu.edu/gesc/Fac_pages/Yoshinobu/4361_5361... (Hubbert is most famous among non-geologists for peak oil theory, and most famous for geologists in considering the role of pore fluid pressure during faulting).

We have the equations, its the computing power that is the limiting factor. You can do a full DNS (direct numerical simulation) of fluids on the order of a few cubic meters. That will simulate flow to the particle, and fully turbulent, with Reynolds numbers (Re) on the order of 10^6 (IIRC). Re goes up proportional to length scale, so if you want to do a km or so, multiply this by 10^2.

This is way ocean models parameterize turbulence (large eddy simulations (LES), and Reynolds Averaged Numerical Simulation (RANS) for example). It's also why observational oceanography is still done.

I was recently at an applied maths seminar. The speaker had been working on the main "over the horizon" radar in Australia for over 40 years. It was amazing to me that he could map ocean currents with radar!

Thinking about what your wrote is aesthetic to me for some reason.

scientists have still not invented equations that can fully capture the nuances of water moving through a river.

-- America’s Last Top Model (http://99percentinvisible.org/episode/americas-last-top-mode...)

We have have the equations [1], we don't have the computer power for a river. Observe the nuances of water moving around a square:

https://www.youtube.com/watch?v=c8zKWaxohng

[1] https://en.wikipedia.org/wiki/Navier%E2%80%93Stokes_equation...

Be awesome to see someone buy up the derelict Mississippi one and turn it into a mini golf course or something. Sad to see so much work going to waste.

Absolutely fascinating in any case. I had heard about the one New York used previously, didn't know there was ever a model of that scale.

I've wondered for a while about how good computer models can be when modelling complex systems. We can't model water flow well, what fundamentally differentiates this from modelling climate?

Another comparable flood was the 2008 Kosi flood in india[1]. It must be as hard, if not harder, to simulate a river like that because of factors like gradient shifts as the river merges into Ganga and tectonic forces.

[1] https://en.wikipedia.org/wiki/Koshi_River

> “The physical model will stimulate the processes on its own,” says Gibson.

Probably a typo, s/stimulate/simulate

The is the most egregiously clickbait title I've ever run across.

This is not what clickbait is. For those unfamiliar with 99PI, this is a thing they do with all of their episode titles. They're often a pun of some sort, and make the subject of the episode a fun surprise which is made apparent at the start.

One of my favorite titles is "Reefer Madness", which is about refrigerated shipping containers – a.k.a. reefers.

Yet, wouldn't you agree that it's a bit deceptive? I'd bet I wasn't the only one to click expecting something completely different.

And that’s a problem … because?

I get that clarity can be a goal, but it’s not the only goal worth achieving. Clear and unambiguous titles aren’t some self-evidently true virtue, especially not in a medium like podcasts, especially not with a podcast like 99PI that’s very much about storytelling.

I do agree that for the medium of Hacker News ambiguous titles aren’t optimal, but 99PI has not special obligation to optimize its titles for Hacker News. Here I would rather have Hacker News adapt the title than wanting 99PI to change it.

I dislike the “clickbait” accusation very much. It’s so lazy. “Clickbait” is really not the only reason that can lead to ambiguous titles, far from it.

Also, I enjoy ambiguous and creative titles, especially from someone I already trust to put out great content (like 99PI). Under those circumstances I honestly fail to see the problem and if there is a problem then it’s that the title isn’t optimal in circumstances where clarity is a virtue, but under those circumstances it seems better to just manually adjust the title or to add a little abstract instead of wanting 99PI to change.

It is more of a bait and switch more than a click bait type title.

The titles for 99PI are done in a way where your expectations are set up one way, with all the baggage the title implies, and they then turn it around on you. I feel like when they do it well, like the Reefer Madness episode, it expands your understanding of the world. One, by expanding your understanding of what a Reefer is, but it also expands upon a topic that you probably had no interest in so that you now will understand the madness that concerns reefers.

> I dislike the “clickbait” accusation very much. It’s so lazy. “Clickbait” is really not the only reason that can lead to ambiguous titles, far from it.

But it leads to more clicks, doesn't it? You're saying that they don't have that intention?

Isn't that the intention of every article title? Is everything clickbait?

Only the deceptive ones.

Aka a creative subversion of expectations.

There are more than one ways to view this. Also, to me something is not clickbait if the actual content might be different from what I expected, but just as interesting or maybe even more interesting.

Here's to the honest writers, who hope fervently that no one will ever read theirs works!

Replacing First with Last implied a pun to me. So, I assumed they meant probably meant the other kind of model. Sure, it might be unclear but not really deceptive.

> This is not what clickbait is. For those unfamiliar with 99PI, this is a thing they do with all of their episode titles. They're often a pun of some sort, and make the subject of the episode a fun surprise which is made apparent at the start.

And that differs from clickbait how?

https://en.wikipedia.org/wiki/Clickbait

> Clickbait is a pejorative term describing web content that is aimed at generating online advertising revenue, especially at the expense of quality or accuracy, relying on sensationalist headlines to attract click-throughs and to encourage forwarding of the material over online social networks.

Is this article low-quality or inaccurate? (No.) Is the headline sensationalist? (No. A silly pun in not sensationalist.) Is this title designed to get you to send it on Facebook? (Maybe? This is kind of a weak criteria since most authors would like you to share their work.)

The entire point of a title is to catch your interest. Having a good title alone isn't clickbait. The fact that you missed a pun doesn't make this clickbait.

The title was clever.

Reply on news.ycombinator.com