I've been wondering for a long time about when this more Bayesian approach would become available alongside an AI. Really excited to play around with this!
Is this kind of like a Karpathy 2nd brain for investing then?
Onlook (https://www.onlook.com/) | Founding Engineer (Fullstack / AI) | San Francisco or REMOTE (US) | Full-time
Onlook is an AI-native visual editor that lets designers, PMS, and engineers collaborate directly in the browser. We launched earlier this year, hit #1 on Hacker News, became the #1 trending repo on GitHub (above DeepSeek + Anthropic), and have grown to 22,000+ stars. We’re now hiring our first Founding Engineer.
We’re a design-led, product-obsessed team backed by YC, and we’re building a next-generation visual IDE used by tens of thousands of designers and engineers.
What you’ll do:
You’ll work across the stack, contribute to one of the fastest-growing OSS projects ever backed by YC, design and ship core parts of the visual editor, work with 100+ contributors, and help define the product with real users. If you’re in SF, we have a beautiful office near South Park.
How to stand out: Share something you’ve built. A project, repo, or portfolio tells us more than a resume.
This is an ultra cool product - stoked someone is building this. We've internationalized our product using ChatGPT O-1 in the JSON files and it's been a real pain to try to decode the strings and otherwise for the text that is in the code.
Definitely going to ask my cofounder to look into implementing this.
- Are there other use cases for this tech besides language translation? Wondering if there are other eng problems that something like this could solve.
- Where are you guys thinking of taking this next? Seems like you're able to handle a lot of languages, so are you just continuing to refine the product or what's the plan for the next couple months and years?
Was it too boring or what was disappointing about the result?
Sonnet 4 has been producing much more reliable outputs, but it's always a challenge to push it to get really creative. I think there are probably great ways we could be prompting for more creative output, but something we'll need to keep experimenting with for sure.
If you use our Desktop Product – https://onlook.com/download - you can import your projects there and use a lot more features like Layers, Components, Pages and more.
We don't have the ability for you to import existing repos for this beta product just yet, but it's something we'll be enabling in the coming weeks for sure.
It's been a challenge to find the right balance for this – on one hand you want to give people who know what code is an easy way to navigate their codebases and be more effective, and on the other hand you want to help people who have never coded before create something they really love.
The visual layer has a lot of polish left to do to be a perfect design tool experience, but we're getting there.
I think for most designers they don't love being limited by traditional web "structures" like flexbox, but that's also how things can get built and scale properly. AI is very good at generating flexbox styled websites but when a designer jumps-in to an AI generated website it's like picking up a complex project someone else has architected. If you know how websites work it isn't super intimidating, but if you don't then it can feel very overwhelming.
One of the main reasons we decided to not have layers and styles immediately visible on the left and right sides of the app was because people who have never jumped into a design tool mentioned they were overwhelmed with the UI. Moving these tools into a "secondary" interaction layer cleared up the interface but made it easy for pro-users to still find them.
I'm excited to keep refining this editor experience so when you're able to import projects you have all of the tools you'd expect from a visual editor for your codebase.
I think there are some issues with the apply model – Kiet's working on a solution for it.
In the meantime, can you try asking Onlook again? I know it's inconveniant.
If nothing loads on the page, try clicking the text on the page, then prompting from the chat asking it to create a part or section of the website or thing you've prompted. Sometimes there's just a ton of code that gets rendered and doesn't apply, so shortening the length of the output manually with a smaller prompt may help.
Ah that's unfortunate – been having some issues with the rendered code being applied.
Can you try it again and see? Or, click on the page, then add your screenshot to the chat with your prompt.
Honestly doing 3D may be a bit tricky in Onlook but super curious to see if you get it to work. It works just like a normal browser would, but the complexity of knowing how to setup a 3D site like that may be too advanced for Onlook.
Is this kind of like a Karpathy 2nd brain for investing then?