😍

Sounds and seems to behave very similar to Moshi from Kyutai, another duplex audio model that came out last year

Just wait until you have the choice between 4o and o4

It's live on the App Store today.
apps.apple.com/us/app/tinyc...

I used Cursor and Lovable to write the code and for the most part because these were idle games, React worked well, with what ended up being a “computed value” engine for complex relationships between numbers (with upgrades etc)

Damn that’s rough. Pray Operator doesn’t remember.

I did see one team successfully fine tuning GPT-4o with images + coordinates for significant improvements in performance, which would be an interesting experiment to perform as well!

See case studies:
openai.com/index/introd...

I’ve also tried strategies like scaffolding which may help but it seems models have been pretrained with very specific coordinate systems and you must use them for it to work. E.g with Gemini 1.5 you must use ymin, xmin, ymax, xmax which must be ranged 0-1000.

Very interesting, thank you for sharing! I was surprised that Claude with computer use (10/22) was unable to use the picture better since it is able to click screens. Did you try other resolutions / coordinate formats? I’d also be curious to see how Gemini 2.0 does (even 1.5 is good at coords!)

We talked about this back when we implemented AI for @framer.com almost two years ago - a model that can’t reason about what it’s doing visually can’t make good looking websites. Hopefully this will change with this new ability in models.