yuntian

We just released 4.8 million real user–ChatGPT conversations collected from our public chatbot.

- Covers a wide range of topics and languages, all from actual users in the wild.

- Includes 122K conversations from reasoning models (o1-preview and o1-mini) which are long, often involving complex problem solving, and very costly to collect.

- 2.5M conversations from GPT-4o.

Links:

- Non-toxic version: https://hf.co/datasets/allenai/WildChat-4.8M

- Full version (gated): https://hf.co/datasets/allenai/WildChat-4.8M-Full

- Exploration tool: https://wildvisualizer.com

Thanks for the feedback! Yes, the demo is definitely limited. The reason I built NeuralOS is that I'm excited about a future where boundaries between software categories fade away. Imagine converting a movie directly into an interactive video game, customizing app interfaces by talking to it, or sharing the same underlying physics/world model between movies and games. Perhaps someday, movies or even interactive games could just be detailed text prompts describing scenes and characters, with the OS "hallucinating" everything on the fly (maybe movies adapt to user preferences as well so different users watch different "versions" of the same underlying movie plot). This minimizes storage and download times, but also provides maximal flexibility.

Unlike other ML-based OS projects (such as Gemini OS, which generates code and renders traditional UIs), NeuralOS directly generates every pixel. While this makes it susceptible to hallucination, in my opinion the other side of hallucination is full flexibility. In the future, I imagine operating systems running entirely (or mostly) on GPUs, adapting to user intent on the fly rather than relying on pre-designed menus and options.

When we say "powered by a neural network," we mean something fundamentally different from a traditional OS (or even gemini os). NeuralOS is essentially a video generation model that "hallucinates" every pixel on the screen in response to user inputs (mouse movements, clicks, keyboard inputs).

There is no underlying kernel, no function calls, no program execution, and no networking. Everything is purely visual and imagined by the neural model. You can think of it as a safe, isolated container where nothing can actually run or cause harm, since no real code executes. It's essentially an interactive video simulation, conditioned entirely on user inputs.

In response to complaints about the laggy demo experience (due to limited GPU resources), I've now set up a huggingface space version for developers: https://huggingface.co/spaces/yuntian-group/neural-os

Note: The Space is intended as a template, so please duplicate it and run with your own GPU for a better experience. (The default Space has only one worker.)

Recommended GPU: At least an L40, ideally an A100-large. (The original demo at neural-os.com used H100s.)

All code and models are self-contained in the huggingface space.

Thanks for your comment! I completely agree that currently NeuralOS is far from being as reliable as a real OS. The Firefox loading time is indeed a funny artifact of the neural model simulating delay in real OS.

However, my real dream behind this project is to blur the boundaries across applications, not just simulate traditional OS interactions. For example, imagine converting a movie we're watching directly into an interactive video game, or instantly changing the interface of an app (like Signal) to something we prefer (like Facebook Messenger) on the fly.

Of course, the current training data severely limits what's achievable today. But looking forward, I envision combining techniques from controllable text generation (such as Zhiting Hu's "Toward Controlled Generation of Text" paper) or synthesizing new interaction data to achieve greater and customization. I believe this is a promising path toward creating truly generative and personalized interfaces.

Thanks again for your interest!

Actually NeuralOS works very differently from Gemini OS. NeuralOS directly generates each screen at the pixel level entirely from neural networks, while Gemini OS generates code that's then rendered into a traditional UI. This difference is why NeuralOS is much slower and currently runs at a lower frame rate.

Good idea. I'll update when no one is using it, don't want to cause further interruptions...

Thanks everyone for trying out NeuralOS, and apologies for the frustrating user experience!

I coded up the demo myself and didn't anticipate how disruptive the intermittent warning messages about waiting users would become. The demo is quite resource-intensive: each session currently requires its own H100 GPU, and I'm already using a dispatcher-worker setup with 8 parallel workers. Unfortunately, demand exceeded my setup, causing significant lag and I had to limit sessions to 60 more seconds when others are waiting. Additionally, the underlying diffusion model itself is slow to run, resulting in a frame rate typically below 2 fps, further compounded by network bottlenecks.

As for model capabilities, NeuralOS is indeed quite limited at this point (as acknowledged in my paper abstract). That's why the demo interactions shown in my tweet were minimal (opening Firefox, typing a URL).

Overall, this is meant as a proof-of-concept demonstrating the potential of generative, neural-network-powered GUIs. It's fully open-source, and I hope others can help improve it going forward!

Thanks again for the honest feedback.

A generative operating system that directly predicts screen images based on mouse and keyboard inputs, powered by an RNN for state modeling and a diffusion model for image generation.

See my tweet for more details: https://x.com/yuntiandeng/status/1944802154314916331