It’s been like three years, and small local models keep getting more capable. Photorealistic video rendering is already here. Writing code from high-level goals largely works. A whole lot of “computers will never do [blank]” went out the window.

These are from models designed for denoising and autocomplete. It’s ridiculous that they work at all. ‘What’s the next word?’ should not be the right question for recalling trivia, translating foreign languages, or answering riddles, yet this gimmick manages a sloppy approximation of whatever you want.

Right now is the worst it will ever be again.

The underlying technology could be trained for literally any goal with examples provided. A better question will inevitably perform deeper witchcraft. Meanwhile: more data and more parameters have not fixed goofy models. Everything they’re capable of keeps appearing in tiny versions that will run on a laptop. Which means better models can be trained for mundane quantities of money, on consumer hardware, with experimental differences. It’s gonna get fuckin’ weird.