🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

AI now beats every single human in the hardest college entrance exam in India, the IIT JEE. Bytedance silently published this result this week. The top scorer was Rajit Gupta with 332/360, but Google's Gemini 2.5 Pro was at rank 1 with 336/360.

So much FOMO happening as I can't be there in person. Enjoy all you wonderful Bangalore people :)

Full house at this afternoon's workshop with Prateek Gupte Shubhankar Khare on MCP and Stable Diffusion. See ya'all tomorrow for the conference, and help our editors beat their milestones!! Vedang (fosstodon.org/@vedang) Pramod Biligiri Sid

Oh my - indeed a packed house. And super high density of Macs and Brains in one area 😀

Thanks for pointing it out! I feel that this piece of regulation is too early and stringent. The AI models are still evolving and it's not clear that compute should be the way to measure it. At least they should be looking to setup an EU-Bench designed to test for things

I have offline wikipedia and some LLMs saved to a hard drive in my earthquake kit My feeling was that it could be a doctor-of-last-resort in a major disaster.

To put it mildly, the past week at Windsurf has been crazy. There have been a lot of different rumors and reports, so I want to share a transparent account of how it actually went down. Before I start, I just want to say that Varun and Douglas were great founders and this

Shubham Tripathi covering segmentation using SAM and Grounding DINO models The Fifth Elephant :D

An Insight from Target's talk in The Fifth Elephant is that since the image exports are in SVG format, which is kind of similar to XML, it is plausible to get some pattern matching via VLMs and LLMs to segment these image SVGs and also train smaller models via reasoning traces.

A talk about Zasper, a more performant alternative to Jupyter at The Fifth Elephant, by its creator Prasun Anand.

🎧 Made some progress on Podsumo today! ✅ Added local audio script generation (with Gemini 2.5 Flash) ✅ Integrated text-to-speech capabilities using kyutai TTS model It's fun to sneak in a couple of hours yet achieve a lot thanks to Anthropic claude code 😄

Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at AthenaAgent , in collaboration with Physics Wallah (PW), using custom RLVR training on 130K+ curated JEE problems

Jeff Coggshall thanks to Vibhu Sapra + Runpod for credits to run these experiments to replicate this: arxiv.org/abs/2305.05065 the a100 has just enough memory for this (using sxm cos no availability for pcie) also inspired by this YouTube talk by Devansh x.com/swyx/status/19…

starting by building a merged base to boost your starting point is nice. though i wonder about multi turn implications. another thing is it removes clipping and KLD completely. removing clipping altogether hasn't been explored that well in open source.

NEW: Higgs Audio V2 from BosonAI open, unified TTS model w/ voice cloning, beats GPT 4o mini tts and ElevenLabs v2 🔥 > Trained on 10M hours (speech, music, events) > Built on top of Llama 3.2 3B > Works real-time and on edge > Beats GPT-4o-mini-tts, ElevenLabs v2 in prosody

Papers with Code was sunsetted yesterday. It had a beautiful run over the years. We are excited to carry the baton with a new launch of Hugging Face Trending Papers 📄

This is very useful. pricepertoken.com

Bilawal Sidhu nice doodling! Mine worked with some questionable results x.com/altryne/status…

Paras Chopra 1. Use the plan mode heavily and that will make it easier to work with existing codebases 2. Use the @ notation to mention file names, classes, functions and it'll work even better 3. "Think deeply", "Super think", "Mega think" result in deeper thinking modes which result in