We’ve since re-enabled it after addressing the underlying issue.

Source: I’m the pm who launched the feature.

What was the underlying issue? Dylan said there was no training on existing apps which…I mean pretty obviously a lie.

(Or worded in some way that only he knows is “technically true” but functionally a lie.)

There was no training on existing apps. We launched the feature in May and we didn't start training on data until a couple months ago. We used off-the-shelf GPT4o for the feature with no fine tuning. How it works is we feed in a design system with premade components, and the LLM decides how those premade components are arranged, then populates them with data and chooses how to theme them. The issue was the underlying design system we contracted to have made had components that too closely matched other screens.

Details are here: https://www.figma.com/blog/inside-figma-a-retrospective-on-m...

Ah, so that’s the technicality. Contractors built components based on existing apps and Figma used that for training. It’s one step removed, but, like, it didn’t not happen.

No need to do it in Figma. Soon your OS will be able to do it, and it will be able to use all apps on your system to create a design, not just one app.

Interesting. Are you thinking of an agentic system here?

Yes they are releasing AI features: https://www.figma.com/ai/

Great idea to allow prompt based design!

Isn‘t vercel v0 doing this?

Somewhat tangential but I've found Claude 3.5 to be ridiculously good at doing this kind of prototyping with React, and you get the preview immediately in the artifact pane to the right of the chat. Mind you I don't ever use React (I'm writing a desktop app in Rust), so I'm probably as unbiased an audience as it gets.

Here are a couple of examples of things people asked for help with and I just iterated with Claude for 5-10 minutes on (code upfront and gif demo at the bottom):

https://gist.github.com/airstrike/7ae444de207e679adca7be6faa...

https://gist.github.com/airstrike/e0e47eaab733277b537923c4d3...

I think given the sheer amount of training data on React, these models tend to really hit it out of the park when they are allowed to use it. Part of me thinks the best approach is to have it model in React first and then translate it to whatever other language/framework you need it in

Definitely, AI for coding makes sense. In the Veryfront Studio, you get a similar experience with live previews and AI coding capabilities, with the added benefit of being able to instantly deploy or export your code.

Uizard does exactly this: https://uizard.io/

Disclaimer: the founder is a friend of mine

Any inside scoop on why they sold the company to miro?

I like that Figma is not required. Can it dynamically model horizontal scrolling, such as album art and info?

Recently needed to develop a React Web solution for hackathon. I had 0 previous experience with web, but was able to create a good looking (but with 0 logic) website, using Vercels V0[0]. It did great both working from text suggestions and from figma screenshots. It also was great at implementing changes after initial generation. Not an add btw, haha

[0] v0.dev

nice one!