rohan

What's happening today? Trying to build webgazer in python xD Got FaceMesh from google mediapipe ready. Now I can get 468 3D landmark coordinate for face, and 52 blendshapes (idk what this means, will fig it out) Ref: webgazer.cs.brown.edu

@koltregaskes @btibor91 @rileybrown @Obsidian i did not know about the extension either. Thanks tibor. i generally just ask claude to open chrome with remote debugging so that it can directly talk with the browser.

@btibor91 @rileybrown @Obsidian OK, I've not figures that out yet. Thanks Tibor.

Oh my, having a geeky Christmas evening, why have I not tried Claude Code until now!? It's bloody brilliant. I've followed one of @rileybrown's videos about using @Obsidian as a workspace for the agent, and it's completely turned my head around. It's gone from a coding agent I don't understand to a general agent that I can easily use. I love queuing it work to do - so good!! I'm wondering, after all this time, should I be using Obsidian instead of Notion? I've definitely shifted to text-only stuff this year, so it would fit me now. Also, why didn't anyone tell me about @typefully, which looks way better than Buffer? I'll be checking that out. If it can do quotes with media, then I'm sold. :-) I've also been trying Codex, it seems good. I've tried it both on the web and in VSC (I also tried Claude Code in VSC), but Anthropic have better tools around their agent, such as skills and the new browser plugin. The plugin seems to be really hit and miss when you select to 'act without asking'. It also closes the sidebar when you move to another tab. I also cannot select a different model (I'm using it in Edge not Chrome). I tried Gemini Code in VSC but couldn't get it to work. It doesn't seem like it's as easy to use as Claude and Codex. I love that the latter two use my subscription, and I don't need to pay or fiddle around with API credits. I wish Claude Code could work with Claude plugin to control my browser, I wish I could use skills with the plugin). I wish I could edit the teachings for tasks like I can edit macros from a recording. I@m also wondering if the Claude Code UI would be prettier. I've asked Code to research this, but I've just hit my Claude limit. ;-) @simtheoryai, if you can get anywhere close to Claude Code, I cannot wait for it to come to the platform. I know you spoke of skills coming next year. I cannot wait; they are a huge deal!

@threepointone I built something like what you want … it’s crude … GH link: github.com/RohanAwhad/cal… initially it was supposed to be gaze tracking, but hand tracking was a bit more robust

can some ml / opencv nerd build a demo where I can use my webcam to interact with my screen?

verl is such a beautiful giant … love to stand on its shoulder play my heart out 😅😂

@kalomaze ohh this should still be fairly easy for an 8b model … i was testing with tool calling high hopes from the model ig xP

@RohanAwhad x.com/kalomaze/statu…

here's qwen7b instruct falling off a cliff

@ottogin1 ohh this is grok 4 heavy model from grok.com. i don’t know what the interface means though … i dont pay for super grok subscription

@RohanAwhad Noice. Do you know what’s this interface? I keep seeing this agent loading progress bar all the times. Typically it would say agent1, agent2 and have different progress bars next to them.

i dont want to see how much I spent on apis

if you have been living under a rock for the past 7 months and have no clue about RL and want to understand the current state of RL and GRPO and Agents and VERL and how you can do asynchronous GRPO or DAPO you should go watch this small talk

dude my dev setup is goated nothing can come close to this

@ottogin1 this has been the best ai engineering crash course that i have had that anyone can ever have tbh dogfooding my way to success

well in my neovim editor, i usually have a llm.md file which is like my chats with LLM it’s like cursor but for neovim - can search web and answer questions - can search the current codebase using find and grep to answer questions - can make changes to file can take in files as context and streams the output into the editor

@ParejaAldo i loved transformers i still love transformers and i will continue loving transformers till the end of time ❤️

@RohanAwhad so now you love HFFFFFF!!!!!!!!!!! Lol

HF transformers is standardizing variable names for at least decoder-only transformers. this way I don't have to if-elif-else my way through supporting multiple models for tensor parallelism tbh w/o HF I don't think we would have had chatGPT at least not in 2022 love you HF

@ottogin1 x.com/rohanawhad/sta…

@ottogin1 ohhh i built this tool xD … it’s a neovim plugin

literally didn't change anything ... 12 hours later ... script just runs

its a new dawn, its a new day, its a new life, ooh and my tensor parallel works

fu o3 🖕

noob errors dont ask me how to solve them 👀