 Anthropic's https://claude.ai/ grew a new feature today: an equivalent of OpenAI's ChatGPT Code Interpreter mode, where the chatbot can write and then execute code in order to help answer questions (e.g. to run calculations that are beyond a next-token-predicting LLM) OpenAI use server-side Python for this, but Anthropic instead chose to use client-side JavaScript running in a Web Worker. Here are my notes so far on the new feature: https://simonwillison.net/2024/Oct/24/claude-analysis-tool/ 10 comments
@danil or at least burn one available CPU core - Web Workers should get their own threads so in theory it should be be able to freeze a whole tab (in theory...) @simon interesting choice. @mesirii I think it's pretty solid. Browsers are the most widely deployed and tested sandboxes on the planet, so I think the absolute worst that could happen is someone's CPU core gets stuck in a loop @simon would be cool to see Llama 3.2 1B or similar doing it right inside the browser π @prem_k that's VERY feasible - I've considered trying to build something like that myself in the past. https://chat.webllm.ai/ runs Llama 3.2 1B very neatly in the browser already @simon yes, I'm a fan of it. It's just so feature rich already. Just need a way to make them use embeddings of content & API for RAG & #LAM. Imagine if they could be exposed by websites, somewhat like robot.txt. Also on similar lines to WebLLM chat is the code of of chromeai.co & its fork, chromeai.org both of which use the Gemini Nano that comes within Chrome 128+ (though it's a bit of a hassle to get that model downloaded right now). @simon Interesting. Do you know (or does the code show) why the vis nodes for the dependencies are of different size? |
Gregory's Server
@simon >use client-side JavaScript running in a Web Worker.
So you can literally ask - "freeze my webbowser".