so what is this project about?
Local Brain — Experiments with In-Browser AI
Basically just experimenting with in-browser AI capabilities. trying to build simple demos around them and see how we can use them in our projects.
New : If you are on chrome. try gemini nano prompt api
spoiler: it is smarter/better than you think.
What is in-browser AI?
In-browser AI basically means Tiny Laguage models that run directly on your machine. instead of sending requests to external servers like openai or anthropic the inference (prediction and generation) is done locally using your on-device GPU or CPU. this is very useful for privacy and security reasons.
On device AI is very early stage and not available on all browsers or on all devices. but it is already very useful for small tasks like summarizing a blog, group tabs, search something in history etc.. these tasks will be well done by browser.
This means users can use power of LLM's without worrying about privacy for tasks like
- parse email to extract contact details
- schedule a calander event based on event details on the web page
- extract or explain content on the web page
- group tabs based on content or purpose
- Go through your browsing history and find relevant links, something you want to search
- and many more...
Since all the inference is done locally, the user can be sure that their data is not being sent to external servers.
How to use in-browser AI?
Currently the user experience is not good. documentation is poor, which browser is supported, how to enable it, how to use it everything is confusing.
Hence this project. will document the process will try and add good demos and write guides on how to use it.
What's Next:
I'm planning to go through WebLLM and see how well our browses can run small to medium size models for example: we should be able to download and run surprisingly models like DeepSeek R1 Distill, Llama 3B or Mistral 7B models run very well on chrome running on macbooks or windows machines. I will be doing some research and will add guides as well as stats or compatibility of these models on different browsers and devices.
I'm planning to go through WebLLM and see how well our browses can run small to medium size models for example: we should be able to download and run surprisingly models like DeepSeek R1 Distill, Llama 3B or Mistral 7B models run very well on chrome running on macbooks or windows machines. I will be doing some research and will add guides as well as stats or compatibility of these models on different browsers and devices.