[ftxbro]

when i see ted chiang i think of his blurry jpeg LLM take in the new yorker. i'm not a fan of that take. i think it was one of the worst LLM takes, competing with the stochastic parrot take. i heard he wrote some amazing science fiction stories though i should read them some time.

[brianpan]

What don't you like about the article? A lot of people don't understand the limitations of ChatGPT and I think it's worth talking about. I like "blurry jpeg" a lot better than "hallucination".

[ftxbro]

> "What don't you like about the [blurry jpeg] article?"

It's valid at some technical level but I dislike it so much. There are so many things. If I had to pick only a few, it's that it encourages to see LLMs as a worse and static rendering of an 'ideal' piece of data that we already have. This is not how the LLMs are useful. An LLM can be at its best when it's explaining or imagining for example something like 'how could x be combined with y in the context of z.' The LLM uses its imagination to answer it. Some would say 'hallucination' which isn't so good description. But in the metaphor of 'blurry jpeg' its answer could only be described as a 'jpeg artifact'. Which is bad. The most useful help that an LLM provides is analogous to the worst aspect of a blurry jpeg. Ughhh I hate it so much.

I don't want to be only negative so I'll put an example of an LLM take that I think is good, it's the simulators essay https://generative.ink/posts/simulators/

Also I can easily believe that Ted Chiang is an excellent science fiction story author I have read almost none of his stories but I heard they are great and I want to read them soon.

[photonthug]

On a related note: Tchiakovsky's novel Children of Memory has some characters of dubious sentience that almost seem based on LLMs. Book is 3rd in a trilogy, but anyone who likes Chiang will like the series.

[flir]

I've read both, and honestly Chiang's on another level. Each story is a jewel constructed by a watchmaker, in the "change one parameter and think through the consequences" tradition of SF. I think you can see this most clearly in What's Expected of Us.

The Children novels are from another tradition - space opera. They focus on big canvases, rather than big ideas. That's not to imply the big ideas aren't there, but they're not the focal point in the same way. They're comparable to Alastair Reynolds or Iain Banks, not Ted Chiang.

[capableweb]

It is very possible (and I'd wager even common) to be a master in some things and suck at others. If you find a good author, it's OK not to agree with what they say or do while still thinking they write good books.

[skybrian]

The stories are good.