Interesting Links for 13-02-2023

[andrewducker]

1. Jews, bicycle riders and trans folk

(tags:bigotry Jews lgbt transgender Russia )

2. Trove of spices from around the world found on sunken fifteenth-century Norse ship

(tags:food history boats trade Sweden )

3. Protesters clash at Tate Britain during drag queen kids story-telling event held to mark LGBT+ History Month

(tags:lgbt history UK art bigotry )

4. ChatGPT Is a Blurry JPEG of the Web

It's not clear that it's technically possible to retain the acceptable kind of blurriness while eliminating the unacceptable kind
(tags:ai knowledge text )

5. 84% of studies find that Adenovirus 36 causes people to put on weight

(tags:Weight virus obesity )

6. What if Judaism advertised the way that Christianity does?

(tags:religion judaism funny )

Comments

[cmcmck]

1. As you know, it's the 'how can you be sure?' that exercises my mind!

People meet me every day and they have no idea because I don't spend all my time on the topic. Why would I?

People on here only know because I'm willing to share the fact on here.

[simont]

#4: I remember that 2013 Xerox incident – it was discussed on Mono at the time. The thing I remember thinking about it¹ was that it was a problem of a mismatch between the nature of the lossy image compression algorithm, and the cost function used to assess whether a given degree of lossiness was acceptable.

Suppose you started with one of the most obvious cost functions for a lossy compressor: measure the difference between each input pixel and the corresponding output pixel, then take the sum of their absolute values, or the rms, or some similar kind of aggregation.

The problem with that cost function is that not all pixel errors are equal. The one-pixel change that turns a (low-resolution) 6 into an 8 is a very bad error compared to a nearby one-pixel change that would have merely given the same 6 a cute little cedilla that everyone would ignore anyway. A better cost function (for preventing that kind of error) would take into account the meaning of the image, as a human will interpret it – so it would know which pixels are more critical than other pixels to preserving the meaning, and would weight errors in those pixels more heavily.

Now if your compression system was based on something like rounding Fourier coefficients, so that its error profile consisted of introducing a higher and higher quantity of white noise to the image as you reduced the bandwidth, then the deficiency in the cost function wouldn't be that important. Maybe you'd get the occasional mistranscribed digit if a random error pixel happened to hit just the wrong spot, but it would happen rarely enough that people would understand.

But when that meaning-oblivious cost function is combined with a compression technique that deliberately spots and exploits large-scale similarities in the image – i.e. spots things that look like meaning – then that's just not good enough any more, because a meaning-aware compressor will generate meaning-correlated errors at a much higher rate, and you need a meaning-aware cost function to prevent it.²

I'm not completely sure what conclusion you get if you map this argument through this article's analogy between the Xerox bug and ChatGPT. The cases aren't exactly parallel, because ChatGPT isn't trying to be a lossy-but-acceptably-so mirror of reality; it's not clear to me that it's really trying to be anything at all beyond a PR stunt. (Which, for very similar reasons to Zaphod Beeblebrox, it's very, very good at.) But if there's any conclusion to be drawn, I suppose it's that systems like this do better at idea generation than correctness filtering, which, fair enough, doesn't seem that much like news.³

---

¹ ok, one of the things. Another was "If this device is behaving as a photocopier, why do you even need to compress the image data – lossily or not – between scanning and printing?" My best guess was that the photocopier in question might have also been able to operate in some fax-like mode which sent the data over a network, so that compression was needed for that mode, and then nobody bothered to disengage the compression pipeline stage in plain photocopier mode. But that's a separate issue, because even if they had, it would surely have made the same error in fax mode.

² An analogous case is the accidentally-quadratic errors that can arise in the use of hash tables. Hash tables always have a theoretically possible nasty worst case, but it happens rarely enough not to worry about, under the fundamental assumption that your hash function is random-looking enough to be uncorrelated to any property of the input that you care about, in particular the order that records or queries are processed during program execution. But as soon as you iterate over one hash table in hash order and add the results to another hash table (without re-randomising the hash function so the two tables are uncorrelated with each other), you've introduced exactly the correlation you were trying to avoid.

³ drswirly pointed me to a hilarious attempt at a chess match between ChatGPT and Stockfish which illustrates this very well...

[channelpenguin]

The photocopier story was from Germany. I'd guess that your guess about the fax is spot on. Even 5 years ago, the hospital chain I worked for STILL regularly used faxes. For whatever reason, I think they were thought of as more secure. Germany still is more likely to use a paper form at points in a process than anything online, and still more likely to respond to a phone call than an email, ESPECIALLY in the public sector. (Emails regularly just seem to disappear into black holes).

[simont]

A friend of mine who went into pub management said a few years ago that cold-calling sales people were persisting with fax beyond reason, and that this turned out to be a great way to get them to put the phone down on you, along the lines of

"... so just give me your fax number and I'll fax you all the information."
"Oh, I don't have a fax machine."
"What? Get with the programme. Everyone has a fax machine."
"Nonsense, fax is 20-year-old obsolete technology."
"Whatever." <slam>

[liv]

1. The 'why the bicycle riders' joke has become a lot less funny since fascists started directly attacking active transport schemes as evidence of a "globalist" conspiracy. There do seem to be a lot of people out there who violently hate cyclists.

Also I disagree with the thrust of the article: the original Nazis, and Putin, and the resurgent far right in America, didn't just pick arbitrary minority groups that just happen to be able to be safely attacked where The who doesn’t really matter. I mean, it's true that the reasoning of the haters is vacuous and there is no realistic threat from these groups. But who gets picked as a target is not just random, it's a consequence of European colonialist history since at least the Middle Ages. Antisemitism is the long hatred, and the violent policing of binary gender has long been intertwined with racism and psuedo-scientific justifications for why white men should plunder and rule the world. It is a mistake to contend that everybody was just fine with Jews until Hitler came along, or everyone was just fine with trans people until Putin or Trump. Yes, coordinated attacks on trans people have got far worse in recent years, but that doesn't mean the conspiracy theories suddenly materialized out of nowhere for no reason. No good reason, true, but I don't think we can address these problems separate from an awareness of history.