Brunch ⎯ an LLM-powered canvas for your thoughts

Brunch is an alternative interface prototype to interact with Large Language Models. You can play with the prototype at brunch.shuvam.xyz

Ever since the advent of ChatGPT, we have gone through endless internet discussions (some might have even ended in fist-fights) on what is the right UX to use LLMs ⎯ and we do have a general consensus that while chat is a good generalized starting point, it also reinforces the notion of these models being a black box, without giving the user suitable granular controls to tweak the output, at least not all the time.

Back in ~2022, I made a rough prototype powered by GPT-3 to explore a branched interface to interact with multiple instances of the model, all of whom were provided with differentiated system prompts, viz. one instance was supposed to answer/respond like a devil, while another was an angel, one was an all-knowing information explorer, while another was a poet who spit out haikus and limericks, regardless of your input. Couple of weeks back, re-wrote the older codebase to swap out the now-deprecated GPT-3 along with adding support for multiple models. I'd admit these new models have not been angel or devil-pilled yet, but if you want to add that enhancement, I'll be on the lookout for PRs to merge.

• Talk to OpenAI's GPT 3.5 Turbo, Claude 3.5 Sonnet, and Gemini 2.0 Flash (Support for more models and granular model controls shall be added over time) by bringing in your own API keys.
• Whenever you write a message, an instance of GPT-3.5 scans the message to figure which model would be the most apt to be the primary responder (essentially a glorified LLM-powered classifier). This optimization was initially made before OpenRouter came into existence.
• You can expand on an idea, get related/adjacent areas of information, and also add more notes/context to the output and basically keep doing this endlessly.
• Personal favourite feature, once you select a thought card, you can visualize the entire chain of thought traced to it's root.
• Voice input, in case you're just that bit reluctant to write, powered by the browser's native TTS instead of adding another additional model.
• Create Clusters to focus on a specific part of the canvas ⎯ there are also instances where GPT 3.5 Turbo scans the contents of the canvas and creates it's own smart clusters. However, the smart clustering behaviour is still flaky at the moment.