Andrew Chatham

San Francisco. Helps make self-driving cars.

[Andrew Chatham @Jumbosaur]

You can see A Christmas Carol at https://t.co/aAQ8nsiktl. It cost me $41!

[Andrew Chatham @Jumbosaur]

My kids love books with color pictures, so I made a tool to illustrate arbitrary ebooks. My 6th attempt in 2 years, but it finally works well enough that I've gotten through most of "The Voyage of the Dawn Treader "and "The Westing Game" with them. Here's A Christmas Carol:

[Andrew Chatham @Jumbosaur]

Like most AI, it's amazing-but-imperfect. Here's the sea monster from Dawn Treader, a near-perfect match for the text. But the ship's stern isn't supposed to break until later in the chapter.

The kids are *very* enthusiastic about pointing out inconsistencies.

[Andrew Chatham @Jumbosaur]

@emollick People keep claiming the open models are always 6 months behind the frontier, but that's based on very few data points, most of which depend on DeepSeek coming up with the next miracle. Everyone else's wins seem dependent on that?

[Andrew Chatham @Jumbosaur]

@KelseyTuoc LLM fiction is pernicious: it looks good but isn't. I've tried it a few times with my kids, and they'll happily listen and not realize what they're missing, but I do. Retelling classic myths and illustrating are the best use case I've found.

[Andrew Chatham @Jumbosaur]

@JigarShahDC So far netork-basee inference latency doesn't seem that big a problem. Maybe for live voice chats? For everything else you're spending multiple seconds waiting for the answer, so a few dozen milliseconds doesn't matter much.

[Andrew Chatham @Jumbosaur]

I don't know Quantum Field Theory or children's literature, but it feels like 'everything is made of ripples in invisible, universe-spanning fields, for real' ought to be a beautiful children's storybook.

[Andrew Chatham @Jumbosaur]

@TheStalwart 3D Gaussian Splats (3GDS) are a current version of this idea. You take a bunch of photos, "train" a network, and then it makes a photo you didn't take but plausibly could have.

[Andrew Chatham @Jumbosaur]

@binarybits True, though we have building automated Adverse Selection machines at this point, so they may break faster than usual.

[Andrew Chatham @Jumbosaur]

@KelseyTuoc @__hand_banana When we looked in SF I only found one that had 7th grade algebra and was in SF proper, so we went there.

Worse was the feeling that I shouldn't ask directly about this stuff because you want them to pick you and not be a problem parent.

[Andrew Chatham @Jumbosaur]

@joseph_h_garvin @TheStalwart There are also severe hardware level exploits mean that "probably secure" sw isn't necessarily so. Examples are rowhammer and spectre

[Andrew Chatham @Jumbosaur]

@fchollet And the most visible token consumption being under flat-rate subscription plans makes this harder to figure out. You also have Adverse Selection tilting things, like most all-you-can-eat deals.

[Andrew Chatham @Jumbosaur]

@emollick I look forward to when someone figures out using LLMs in the gameplay loop itself, perhaps as the scoring mechanism. I briefly played around with a MTG-style game where your cards were words, and an LLM assessed damage. I'm bad at this so it wasn't fun, but maybe someone else!

[Andrew Chatham @Jumbosaur]

@socialwithaayan Bloom's two sigma claim hasn't been replicated because it was probably fudged. The benefits of 1:1 tutoring have been replicated, but not at 2 sigma.

https://t.co/vPngTxPQbv

[Andrew Chatham @Jumbosaur]

@RakeshSFNYC @reed @agarwal I was going to call Reed out, but I think he means "so the cars are less likely to kill you on a bike", not "take a Waymo instead of a bike".

[Andrew Chatham @Jumbosaur]

@TheStalwart Current image generation is also very tied to language understanding, and they use very similar methods of turning text into a numeric representation for their input.

[Andrew Chatham @Jumbosaur]

@TheStalwart OpenAI was the one who published the Diffusion paper which led to the proliferation. The preceding technology, GANs, produced images that were *much* worse. So at least a company-level connection.

[Andrew Chatham @Jumbosaur]

@CaleMcNulty @lymanstoneky I think you switched definitions of "selection" there and maybe are both right? Alpha absolutely selects for parents who are very interested in academic progress, moreso than other "selective" schools. That community is surely some factor.