ChatGPT as a blurry jpeg of the web

by | Sunday, February 12, 2023

Ted Chiang is one of the greatest, insightful writers working today. I had written previously about one his short stories in a post titled: Truth of fact and feeling: Unpacking McLuhan (2/3) about his short story The truth of fact and the truth of feeling. (If you haven’t read it – please do so).

In a recent piece in the New Yorker, (ChatGPT is a blurry jpeg of the web) Ted Chiang focuses on ChatGPT, and offers one of the best descriptions of the technology, making an analogy with lossy compressions systems (such as the Jpeg file format that we usually use to compress images). The essay is worth reading in full, but here are some key quotes:

… Think of ChatGPT as a blurry jpeg of all the text on the Web. It retains much of the information on the Web, in the same way that a jpeg retains much of the information of a higher-resolution image, but, if you’re looking for an exact sequence of bits, you won’t find it; all you will ever get is an approximation. But, because the approximation is presented in the form of grammatical text, which ChatGPT excels at creating, it’s usually acceptable. You’re still looking at a blurry jpeg, but the blurriness occurs in a way that doesn’t make the picture as a whole look less sharp.

This analogy makes even more sense when we remember that a common technique used by lossy compression algorithms is interpolation—that is, estimating what’s missing by looking at what’s on either side of the gap. When an image program is displaying a photo and has to reconstruct a pixel that was lost during the compression process, it looks at the nearby pixels and calculates the average. This is what ChatGPT does when it’s prompted to describe, say, losing a sock in the dryer using the style of the Declaration of Independence: it is taking two points in “lexical space” and generating the text that would occupy the location between them. (“When in the Course of human events, it becomes necessary for one to separate his garments from their mates, in order to maintain the cleanliness and order thereof. . . .”) ChatGPT is so good at this form of interpolation that people find it entertaining: they’ve discovered a “blur” tool for paragraphs instead of photos, and are having a blast playing with it.

Given GPT-3’s failure at a subject taught in elementary school, how can we explain the fact that it sometimes appears to perform well at writing college-level essays? Even though large language models often hallucinate, when they’re lucid they sound like they actually understand subjects like economic theory. Perhaps arithmetic is a special case, one for which large language models are poorly suited. Is it possible that, in areas outside addition and subtraction, statistical regularities in text actually do correspond to genuine knowledge of the real world?

I think there’s a simpler explanation. Imagine what it would look like if ChatGPT were a lossless algorithm. If that were the case, it would always answer questions by providing a verbatim quote from a relevant Web page. We would probably regard the software as only a slight improvement over a conventional search engine, and be less impressed by it. The fact that ChatGPT rephrases material from the Web instead of quoting it word for word makes it seem like a student expressing ideas in her own words, rather than simply regurgitating what she’s read; it creates the illusion that ChatGPT understands the material. In human students, rote memorization isn’t an indicator of genuine learning, so ChatGPT’s inability to produce exact quotes from Web pages is precisely what makes us think that it has learned something. When we’re dealing with sequences of words, lossy compression looks smarter than lossless compression.

Do read the actual piece… it is worth your time: ChatGPT is a blurry jpeg of the web

A few randomly selected blog posts…

Shin, et. al. wins Outstanding paper award

Just found out that Shin, T., Koehler, M.J., Mishra, P. Schmidt, D., Baran, E., & Thompson, A.,(2009, March). Changing technological pedagogical content knowledge (tpack) through course experiences. Paper presented at the 2009 International Conference of the Society...

Manoranjan ka baap

The Indian Premier League, Twenty20 cricket championship was a great success. I had a chance to watch a few games (including the finals and semi-finals) when I was in india and it was a blast. This posting however is about an extremely creative commercial for the IPL...

Photos from SITE08

Matt has Flickrd photos from SITE08. Some of these photos are taken by me, but the rules are that the owner of the camera automatically gets the bragging rights 🙂 and since I didn't take my camera along, he takes credit for all the pictures. Given that a bunch of...

Mirror, mirror

Imagine you are standing in front of a bathroom mirror; how big do you think the image of your face is on the surface? And what would happen to the size of that image if you were to step steadily backward, away from the glass? A recent article in the NYTimes (Mirrors...

Why don’t students like school… In 30 mins!

One of the challenges faced by all instructors is ensuring that students actually read the textbook. This summer we came up with a innovative assignment to address this issue. The book in question was  Daniel Willingham's Why Don't Students Like School? A cognitive...

The making of “Editing is Cool”

I had posted about this really cool video I recently found (see Life is about editing). Behold my surprise when one of the comments on the blog was from none other than Allee Willis (see her wikipedia page here, and personal website here). It was just great to hear...

It Takes Two: A scientific romp using AI

It Takes Two: A scientific romp using AI

Dark 'n' Light is an e-zine that "explores science, nature, social justice and culture, through the arts and humanities." It is a labor of love by a small, dedicated team led by Susan Matthews, former legal and policy wonk, turned editor and podcaster. I came to know...

Links of interest

During Dr. Jalaluddin's keynote I took some time to search online for some reports, prompted by what he had been saying. (Yes I was listening not just browsing). The first is an European study: ICT in Schools: Trends, Innovations and Issues in 2006-2007. You can...

0 Comments

Submit a Comment

Your email address will not be published. Required fields are marked *