It's implied by the fact that early in the post they say:
>"To test this system, we pointed it at an ambitious goal: building a web browser from scratch."
and then near the end, they say:
>"Hundreds of agents can work together on a single codebase for weeks, making real progress on ambitious projects."
This means they only make progress toward it, but do not "build a web browser from scratch".
If you're curious, the State of Utopia (will be available at https://stateofutopia.com ) did build a web browser from scratch, though it used several modules/packages for the networking portion of it.
Thank you for the detailed feedback, though we would prefer for you to comment on the announcement threads where you see it. We really appreciate the feedback.
You're referring to State of Utopia's[1] web browser, currently available here:
That livestream demonstration is side-by-side with Chrome, rendering very simple pages.
It compiles, renders simple web pages and is able to post.
The differences between cursor's browser and our browser:
- Cursor's long-running autonomously coded browser: over a million lines of code and a trillion tokens, which is computationally intensive and has a high cost.
- State of Utopia's browser: under 3000 lines of code.
- Cursor's browser: does not compile at present. There's no way to use it.
- State of Utopia's browser: compiles in every version. You can use it right away, and it includes a fun easter-egg game.
- Cursor's browser: can't make form submissions
- State of Utopia's browser: can make form submissions.
I'm submitting this using that browser. (I don't know if it will really post or not.)
We are taking feature requests!! Submit your requested feature here:
This visualization lets you watch a fruit-fly-inspired neural system learn in real time using several different approaches. This architecture is capable of navigating a small maze, also included in the visualization.
Click on any fly to see a visualization of its brain. You can select a fly at the start to follow its progress over time.
Although the term "evolution" is used, this refers to a genetic-algorithm process: selecting high-performing brains and combining them through crossover, without learning from experience within an individual lifetime.
The visualization includes different training methods: pure evolution (genetic algorithms only), backpropagation (gradient descent only), a hybrid approach (evolution plus backpropagation), and a small attention-based model (three flies). Be sure you click a fly and check its brain at the start and after a few generations. The visualization is excellent.
Adobe added embedded javascript to pdfs. Its an option to turn it off but its enabled by default. I turned mine off a long time back and never notice any problems but I don't use a lot of pdfs with interactive forms.
I have yet to see an exploit that can be performed with a .txt file. PDF files can have all sorts of interactive junk and nested files embedded in them - you can get really crazy in that format.
At the same time they were doing this, I also iterated on an AI-built web browser with around 2,000 lines of code. I was heavily in the loop for it, it didn't run autonomously. You can see the current version of the source code here:
I'm posting from that web browser. As an easter egg, mine has a cool Tetris clone (called Pentrix) based on pieces with 5 segments, the button for this is at the upper-right.
If you have any feature suggestions for what you want in a browser, please make them here:
you're not meant to scroll ahead, the intention is to create a time sink in each chapter. I figured this out after reading the first chapter forever, then I was like "wait a second this is the concise book? how can I read the introduction forever?"
I've read the introduction part and I've seen no "time sink." I genuinely don't know what your comment means. Do you really have a hard time reading the introduction part?
you won't believe how powerful Python is with libraries. ChatGPT and Claude made a brand new browser, that isn't based on Chromium or Firefox, and yet still follows many aspects of layout correctly. I read the article we're discussing ("stop designing languages") on this browser and I'm currently using it to post this reply.
it doesn't look like it would be easily derived from Chromium or Firefox, because this code is Python and those don't use Python this way.
By the way is there any feature you'd like to see added to the toy browser? The goal is that one day it's a replacement for Chrome, Firefox, etc. It's being built by ChatGPT and Claude at the moment. Let me know if there are any feature ideas you have that would be cool to add.
>A python-based browser? What are you using for the GUI toolkit?
Great questions.
1. Yes, for the moment. Like the title of this article suggests - we're using a library! :)
It's great to iterate in Python, which has a large ecosystem of libraries. Believe it or not, there is a chance that in the future it would be able to translate the language into a different one (for example, C++) while using C++ bindings for the same gui libraries. This would speed up its actions by 40x. However, not all of the libraries used have C++ bindings so it could be harder than it looks.
you can have a quick read through. Originally it was using tkinter for the GUI toolkit. I believe it is still using tkinter, but the AI might be leaning on some other library. As you read it, is it using anything but tkinter for the GUI toolkit?
These libraries are doing a lot of heavy lifting, but I think it is still ending up drawing in tkinter (not handing off rendering to any other library.)
This is pretty relelevant to a project I'm working on - a new web browser not based on Chromium or Firefox.
Web browsers are extremely complex, requiring millions of lines of code in order to deal with a huge variety of Internet standards (and not just the basic ones such as HTML, JavaScript and CSS).
A while ago I wanted to see how much of this AI could get done autonomously (or with a human in the loop), you can see a ten-minute demo I posted a couple of days ago:
It's only around 2,000 LOC so it doesn't have a lot of functionality, but it is able to make POST requests and can read some Wikipedia articles, for example. Try it out. It's very slow, unfortunately.
Took a quick glance through the code, its a pretty decent basic go at it.
i can see a few reasons for slowness - you arent using multiprocessing or threading, you might have to rework your rendering for it though. You will need to have the renderer running in a loop, re-rendering when the stack changes, and the multiprocessing/thread loop adjusting the stack as their requests finish.
Second, id recommend taking a look at existing python dom processing modules, this will allow you to use existing code and extend it to fit with your browser, you wont have to deal with finding all the ridiculous parsing edgecases. This may also speed things up a bit.
Id also recommend trying to render broken sites (save a copy, break it, see what your browser does), for the sake of completion
thank you for your quick code review and for these many helpful tips! I'll take a look at them and see what I can put into practice.
EDIT: Unfortunately, it seems that the code is getting near the limit of the context window for Claude, so I'm not able to add several of the feature suggestions you added with the present approach. I'll look into breaking it up into multiple smaller files and see if I can do any better.
>"To test this system, we pointed it at an ambitious goal: building a web browser from scratch."
and then near the end, they say:
>"Hundreds of agents can work together on a single codebase for weeks, making real progress on ambitious projects."
This means they only make progress toward it, but do not "build a web browser from scratch".
If you're curious, the State of Utopia (will be available at https://stateofutopia.com ) did build a web browser from scratch, though it used several modules/packages for the networking portion of it.
See my other comments and posts for links.
reply