Log inSign up
Image
user avatar
Avijit Ghosh
@evijit
AI Research & Policy @huggingface 🤗 . Leading: @evaluatingevals @huggingscience
Boston, Massachusetts
evijit.io
Joined January 2012
1,615
Following
3,027
Followers
RepliesArticlesMedia
  • Pinned
    user avatar
    Avijit Ghosh
    @evijit
    Jul 1
    Very proud to launch FLARE-AI today, after extensive coordination between cybersecurity and safety orgs. This is just a starting point, and we are working with more orgs and legislative efforts to cement coordinated flaw disclosure practices in AI. Find us at ICML next week!
    user avatar
    clem 🤗
    @ClementDelangue
    Jul 1
    AI systems that are accessible or even better open-source are safer for a simple reason: more people can inspect them, test them, stress them, and report what breaks or harms to keep the builders accountable. "Sunlight is the best disinfectant"! That’s why we’re proud to help
    Image
  • user avatar
    Avijit Ghosh
    @evijit
    5h
    One of my papers is trending on hacker news and a commenter is hating on the large number of authors 💀 truly fascinating community
  • user avatar
    Avijit Ghosh
    @evijit
    Aug 3
    Just realized that the Position and the E&D tracks use different scales -- this is the most confusing neurips review cycle ever
  • user avatar
    Avijit Ghosh
    @evijit
    Aug 1
    The only correct take, it’s generally helpful to consider which problems are fixable before thinking about broad policy. It’s turning out that almost all verifiable problems are moving towards being solvable by GPAI systems.
    user avatar
    Matt
    @matt_oai
    Aug 1
    2022 / 2026. Wild to see how far and fast we've come in less than 4 years
    Image
    Image
  • user avatar
    Avijit Ghosh
    @evijit
    Jul 31
    Forever grateful to 🐳 for turning development attention to efficiency. Where we were 2 years ago was completely unsustainable, cost wise. Which leads me to think 1 year from now people will just be able to download a few GB model on their MacBook and not pay $ for coding 👀
    user avatar
    Ahmad
    @TheAhmadOsman
    Jul 31
    DeepSeek V4 Flash is ~70% smaller in size than GLM 5.2 It also beats GLM 5.2 which was the SoTA model just about a month ago In this video I explained how we will have GLM 5.2 level intelligence on a single RTX 5090 in < 18 months Looks like it’s happening way sooner than that
    Image