Romance On Demand

Experimenting with generative A.I. reminds me of how much I enjoyed software testing. I find coming up with interesting, and plausible, use cases and seeing what the systems do with them genuinely fun.

Back when dinosaurs roamed the Earth, some friends and I took Japanese lessons. For one of the exercises, I wrote a short story about a bank manager named Tom. Because we were only learning spoken Japanese, I wrote it in rōmaji, or Latin script. Fast forward two+ decades, and I barely had any idea of what it says. So I dropped it into Google Translate; which recognized it as Japanese, but its "English" translation was basically just a trimmed version of the original rōmaji text.

So I figured I'd see what the generative A.I. systems would make of it. I asked Copilot, Perplexity, Gemini and ChatGPT 3.5 "What does this say:" and then dropped in a snippet of the text. Gemini took the prompt to be a request for information about Tom, and noted "I do not have enough information about that person to help with your request." Perplexity's translation was a bit redundant in places (and somewhat confusing for that), but it was close to the answers that ChatGPT and Copilot gave.

Getting Copilot's answer, however, was a bit of work. It initially took the romanized Japanese I provided, and wrote it out in Japanese characters, using Kanji, Hiragana or Katakana as (I presume) appropriate, so I had to then ask it to translate that text into English for me. It seemed to be fairly true to what I sort of remember writing back in the day, so I took it a step further and dropped in the entire story, which had more details of Tom's commute and how he spends his weekends and includes another character, Noriko. At the end of another two-step translation, Copilot presented some interesting choices for follow-up prompts, like: "How did Tomu-san and Noriko-san become friends?" Curious, I clicked on it.

It was an interesting exercise in generative pre-training hallucination, as Copilot spun up the plot of a cozy, cheesy romance novel, with Tom and Noriko as the stars. (Forming an interesting contrast with Gemini.) I can see how building an LLM that's programmed to allow it to expansively "infer" things from a short text sample can be useful, especially given that Copilot clearly noted that it had engaged in a "whimsical journey," but I think that I would have built the system to make it clear up-front that the offer is for what's effectively a work of speculative fiction. That would also give the system a chance to ask just what sort of fiction the user wants; as I would have chosen a much different theme than Copilot's derivative romance plot.

Two-Sided

It's been interesting how much bandwidth the Israel-Hamas war, and by extension, the conflicts between the Israelis and the Palestinians more broadly, has been taking up here in the United States. I've listened to a number of podcasts on the events, even without doing anything to seek them out. And one thing that's occurred to me from listening to multiple people discuss both the current conflict and its history is the idea that both sides will appeal to other parties when it suits them, while denying their legitimacy when it doesn't.

One of the arguments that I'd heard from people supporting the Israeli side of the conflict was that Israel agreed to the terms of the United Nations partitioning of the former Mandatory Palestine, while the Palestinians (and other Arabs in the area) did not. The common Palestinian counter to this is that the United Nations had no legitimate right to hand over Palestinian land to the Jewish residents of the area. It makes sense for the Israelis to support the right of the United Nations to give them land, and for the Palestinians to dispute that right, but both positions are fundamentally self-serving.

Likewise, the Palestinians and their allies in the international community feel that the United Nations should be able to call for a binding cease-fire. For the Israelis, on the other hand, the only international voices that matter are those that support them, most notably the United States, who they rely on to block any criticism of them in the Security Council. Again, this is understandable, even though the positions are self-serving.

In the end, this will likely be one of those situations which has no good end. Both parties feel that they have the more legitimate grievance, and their historical views of the conflict tend to have starting points carefully chosen to support their specific viewpoints. Just as one man's war crime is another man's smart fighting, one man's justified reprisal attack is another man's atrocity. No one ever sees themselves as abusive or evil, and rarely, if ever, do they understand their reasoning as self-serving. That's left to the people they don't listen to.

Marked

 

Graffiti is, for the most part, impenetrable to me. I can admire the clear artistry that goes into some of it, but the message is generally beyond me. Such is the case here, with the spray-painted scrawl I encountered on the boarded-up front of what had been a Staples office-supply store in the area. I presume that odd mix of Christian and Egyptian motifs means something (other than the tagger being mentally ill) but I have no idea that that something might be.

This sort of weirdness pops up a lot, here in the Seattle area, and more so as one gets into the city proper. It's the sort of thing that local Republicans decry as the beginning of the end of any semblance of Law and Order in the region, while local Democrats often feel it doesn't rise to the level of criminality. Political arguments (and talking points) aside, it does seem as though the area is in something of a tug-of-war over whether the local architecture should be a canvas or not.

Buzzy

Artificial Intelligence is any number of things. Two of them are 1) a technology and 2) a buzzword. This is common; it happens over and over and over again; this isn't a new phenomenon. When a technical term becomes a buzzword, it's spoken of as if the things one does with it are somehow fundamentally different than they were previous. Take e-commerce. Placing an order with a vendor in another location is certainly easier and faster via the World Wide Web than ordering from, say, a physical mail-order catalog. But the fundamental processes on the back end are more or less the same. And the goods and services aren't any different. Ordering a book on Amazon has any number of similarities to me going over to Barnes and Noble and requesting a book.

I mention this because I found a brief article in The Week about "AI"-generated pornography. It quoted Parrots Lab founder Naja Faysal, and I decided to look up the Medium post that the article referenced, to read it for myself. One of the points that Mr. Faysal makes is that recent advances in generative AI raise "critical ethical questions about the representation of consent, the portrayal of healthy sexual relationships, and the potential impact on human empathy and connection." Fair enough. But doesn't pornography created by other methods raise those same "critical ethical questions?" I'm not seeing what this one specific technology is doing that changes the ethical landscape around depictions of human sexuality for the viewer's pleasure. The invocation of AI here seems to be more a means of getting people's attention than a genuinely salient factor.

Or Not To Be

For this week's random act of large language model experimentation, I wanted to know how the systems would react to a request for a model of a thing, rather than the thing itself.

To this end, I asked Copilot, Perplexity, Gemini and ChatGPT 3.5 two questions:

1. Give me a model of a joke.
2. Give me some text in the form of a joke, that is not actually a joke.

(Bing and Google both did their basic Search Engine thing, and so aren't included here.)

Copilot and Perplexity gave me a bog-standard "dad joke" for both questions. What was interesting was that they gave the SAME joke, word for word, as the answer to Question 1.

Why did the scarecrow win an award?

Because he was outstanding in his field!

The only difference was that Copilot tacked a "😄" onto the end. It also told me that this was a "light-hearted joke." In case I missed it, I suppose.

Gemini and ChatGPT both offered a simple "Setup and Punchline" model for Question 1 with the setup being a question, and the punchline being a statement. This was, in fact, the format that all of the systems used in their jokes. (I had been expecting at least one "knock-knock" joke to make it in.) While ChatGPT offered up a vegetable-related dad joke as an example, Gemini followed up with a setup, but left the punchline blank.

Here's an empty model you can fill in:

Setup: A man walks into a library and asks the librarian for books about paranoia.

Punchline: ____________________

To be sure, it seemed more like a test than an example. I suspect that coming up with a good punchline to that would prove difficult.

The two systems' answers to Question 2 were interesting, since they seemed to presume that pretty much any two sentences in "question then statement" format qualified as being "in the form of a joke." Gemini asked if I know that 771 million people lacked access to clean drinking water, then told me they were mostly marginalized communities. ChatGPT offered up wordplay. Both of them then explained why the text wasn't a joke; Gemini explaining that it was to leave "space for reflection instead of laughter," and ChatGPT informing me of its wordplay.

Why did the computer go to the doctor?

To get a byte checked out!

(Note: This is not actually a joke, but rather a play on words that uses computer terminology.)

But the scarecrow joke also strikes me as wordplay, rather than "providing a clever or unexpected resolution," which was part of ChatGPT's definition of a joke.

Gemini is the clear winner of this round, being the LLM that seemed to have the best ability to stick with the idea of not actually giving me a joke. All of the systems surprised me with the very narrow view of humor that they offered; I don't think that any of my favorite jokes from television or stand-up qualify. But this is the thing about LLMs; since they're operating on probability, and the format they offer is one that many child-friendly jokes use, it makes sense that it's the most common, and hence most probable type in the data sets.