mark

mark, 4 months ago

@lauren And they didn't ground it after pressure warnings?

Geezus. I'm looking forward for their justification of that decision being "Well, those pressure sensors are known-unreliable, so..."

mark, 4 months ago

@rusozoll @lauren I'm glad they took steps, but I still have questions about the overall policy of letting a plane that's check-failing somewhere in the pressure system fly at all.

My thinking on the topic is "No pressure events are okay. Even if most are survivable, they all involve the whole cabin getting roller-coastered into breathable atmospheric density, so most will also guarantee you have a planeload of passengers who never want to fly with your airline again. Maybe don't risk it period."

mark, 4 months ago

@rusozoll @lauren You raise an excellent point. Alaska Airlines continues to enjoy a zero-fatalities safety record, which is impressive. So I think it's fair to believe they take safety seriously and are competent.

... all the more reason they should respond to this failure by being more careful in the future. There's definitely a risk of NASA effect here; the Space Shuttle was a fatality-free technical marvel until it wasn't.

I don't doubt given their record though that they will take multiple pressure events over several days more seriously in the future.

mark, 4 months ago

@ErikUden It's useful to frame these technologies in terms of plagiarism but, ultimately, if the underlying tools have utility the question of plagiarism will be short-lived.

If it's valuable to a Disney, or a Universal, or a Paramount, or any one of the dozen Madison Avenue ad agencies to auto-generate unlimited visual content at the cost of text prompts, they will hire a small army of artists for a one-time fee to sell their work for training data. All nicely tucked away with proper pedigree and copyright i's dotted and t's crossed. How many millions do we imagine it would cost them to do it? Doesn't matter; it'll be worth it as a spot-cost, and they will find artists willing to trade their catalogs for that fee.

The plagiarism concern is real with the current iteration of the tools. It won't be relevant with the next iteration. Not when companies that own stock photo archives and B-roll databases can choose to feed those archives into their own tool.

My prediction is the next shape of the terrain will look primarily like this: publicly-accessible tools that output content that's plagiaristic, and privately-owned tools that allow mega-corporations to output unlimited media content for nearly free from training data they can show receipts on if someone hauls them into court. With a smattering of self-built assistants used by individual artists / small collectives trained on the datasets they themselves created ("Self-plagiarism machines," if you will).

mark, 4 months ago

@lauren Hmm. :)

I'm reminded of the old chestnut from the history of AI where one of the MIT profs was talking to a student who was building a tic-tac-toe AI and started with random weights. Prof asked why they are random. Student said (not actual quotes, paraphrasing) "I don't want the system to have any preconceptions," and he responded "Well it does; you just don't know what they are."

(That having been said... You now have me thinking whether machine learning could be described as "It breaks the GIGO rule by finding pattern in data that's otherwise garbage." But it's still up to the human to decide if the pattern matters, and there's a big understanding gap in a lot of the currently-available tooling).

mark, 4 months ago

@lauren Exactly.

There's a term I can't remember right now that I picked up from an intro to machine analysis a few years back. The idea is basically that as the number of variables you're tracking approaches infinity, the odds the statistical regression you're using will key into a pattern approaches one.

... but whether that pattern is real is unknown. It's just correlation, it doesn't imply causation. This is why "fishing" for data in medical analyses is such a problem and causes reproducibility issues; every data sampling operation creates a snapshot, and if you stare at the snapshot long enough, some pattern will fall out. But whether it's real or a "shape in the clouds," you don't know until you find causation.

mark, 4 months ago

@mekkaokereke This is excellent news! Do you have any data on it?

One of the things that honestly startled me about higher ed is how profoundly bad some of the profs were at communicating information relative to the high school teachers I'd had. Some were also excellent, but I hadn't actually wrapped my head around the fact that a tenured university prof's main job isn't teaching, so they can be awful at it.

Meanwhile, some of the best teachers we had were only "instructor" title because they'd never gotten a Ph.D, but the school pulled them in because they'd made a career in industry doing comp. sci and clearly demonstrated they knew their shit. But academia can't reward, y'know, "just demonstrating a lifetime of knowing how to do things." Not without that thesis.

mark, 4 months ago

@mekkaokereke

Ah, excellent. Carnegie Mellon went through a similar evolution in their process at some point: they saw in their admissions data that much of the gender gap in performance might be correlated to how young men tended to have picked up the basics by "nerd osmosis," but young women hadn't (purely because of how hobbies in high school were shaking out: women didn't tend to get into the [at the time, I remember, pretty damn sexist] online hack-your-Linux-box scene so they didn't have as much of that 'journayman knowledge' hitting freshman year).

They increased the offerings for core skills and the gap started to close.

mark, 4 months ago

@karabaic @mekkaokereke In my case, I seriously questioned my decision to go CS track because I was comparing myself to graduates from the dedicated math-sci school in Northern Virginia and was worried that I wasn't going to measure up because they had a functional one-year head start on the core discrete mathematics concepts (a topic my high school didn't touch on at all).

I'm glad I stuck with it. It was touchy for awhile there. I had help.

mark, 4 months ago

@lauren I thought they tried that in the, what, '60s and it didn't go great?

mark, 4 months ago

@mcc Yeaaaahhhh.....

It turns out that even for all its flaws, YouTube still has some of the savviest people working on the technical problems and the algorithm, though it has bad corner cases, is very good in the common case.

mark, 4 months ago

@lauren Is this an actual new decision, or are you taking the piss on the Twitter account hack?

mark, 4 months ago

@lauren Sorry; I didn't see a link. I'll follow up on the search engines. Thanks for the heads-up!

(ETA: chalk it up to the eventual-consistency of Mastodon... I can see your "Congress needs to get" post, but not the post immediately before it that included the link until I navigated directly to your feed on your server. 🤷 )

mark, 4 months ago

@mcc We had cached a compilation of the tzdata database in our build chain instead of building it. As a result, we were slowly falling out of date with political changes and weren't aware of it.

Moscow changed the law about daylight savings, and all of a sudden we were off by an hour for all our users in that city when the new law went into effect. That was... Mildly embarrassing. 😉

mark, 4 months ago

@ayo Political campaigns, technological innovations, or "That's what she said?" 😉

mark, 4 months ago

@chrisamaphone I think "Greatest Show on Earth" notes that even in a non-life-consuming environment, cross-individual competition occurs.

Trees don't need to be so damn tall, and in fact it's expensive / risky to be so; height increases the risk of weather events uprooting you.

... but if the neighboring tree is a foot taller than you, their canopy spreads over you and cuts off access to some light, so to maximize light you want to be the tallest thing around, but that means your light-maximizing neighbors want to be taller than you, but that...

(... in short, "cooperation requires more complex interlocking and signalling than competition, so even among members of the same species competition is anticipated").

mark, 4 months ago

@chrisamaphone If they haven't, you could be admired by lesswrongers everywhere if you publish that paper. 😉

mark, 4 months ago

@RickiTarr Oh, absoluelty "The Midnight Pals." Created by @bitterkarella, famous horror authors as if they were members of the Midnight Society gather around a campfire to tell each other their ghost stories. Each one much different from what one would expect.

Top link:

https://www.midnightpals.com/audio-series

Recommended ep: "The Tale of Dr. Jekyll and Mr. Hyde." https://open.spotify.com/episode/2FUk1UBu19xlojIxLDbxjT

mark, 4 months ago

@accretionist @adhdeanasl This is actually the first thing my wife said when we were talking about this. I mentioned "So in some countries they're seeing improvement with COVID after taking animal de-wormer..." and she immediately said "Did they test the patients for worms first?"

mark, 4 months ago

@donw Yes. The short answer is "process capture; the parties had the political will to bend the public coffers towards their semi-private process so they did."

(The longer answer is way more complicated and has a combination of "in all practicality, it likely is most convenient for the most citizens if we don't have the parties buying and maintaining their own election equipment when we already have all this election equipment" and "as consequence of using the public machinery, the state law places constraints on party operations that tend to make those operations more open than if they were allowed to run their affairs full-on Tammany Hall-style, so it's better for the health of the republic than the alternative").

mark, 5 months ago

@ErikUden It's a little funny how in that last post he trips over structural racism and sexism without realizing he has. I believe the name of the Facebook tag group is "Conservatives walking backwards into the point."