Arunachal PradeshAssamChandigarhKarnatakaManipurMeghalayaMizoramNagalandPunjabRajasthanSikkimTripuraUttarakhandTelanganaBiharKeralaMadhya PradeshAndaman & NicobarGujaratLakshadweepOdishaDadra and Nagar Haveli and Daman and DiuJammu & KashmirChhattisgarhDelhiGoaHaryanaHimachal PradeshJharkhandTamil NaduUttar PradeshWest BengalAndhra PradeshPuducherryMaharashtraLadakhAI Safety NITAAgartalaDelhi AI Alignment and Safety CollectiveDelhiAI Safety India BengaluruBengaluru
Boundaries traced from Survey of India maps by DataMeet (CC BY 4.0). Not to scale.

Community

Events and speaker sessions

Live talks, reading groups and city meetups, all on our Luma calendar.

Subscribe on LumaJoin the WhatsApp community
Anusha presenting at an AI Safety India speaker session in Bengaluru.

What’s on

Upcoming

Past

From the room

Aditya Raj giving the opening talk at the Bengaluru meetup.
Keynote · BengaluruAditya Raj opens the evening
Parishrut presenting by video to the room.
Speaker session · HybridParishrut joins by video
Aditya Prasad presenting in an online speaker session.
Speaker series · Online · August 2026Aditya Prasad on how to get into AI safety research
People talking in small groups after the talks.
Meetup · BengaluruConversations after the talks

Write-ups

All posts on Substack
Speaker session · Hackathon winners · 5 August 2026

Do multilingual vision-language models abstain under cross-modal conflict in low-resource languages?

Anvesh Reddy Lankala, Vicky Feliren and Akansh Jain, winners of the Apart Research Global South AI Safety Hackathon (Asia-Pacific track)

What they found
  • VLMs get about twice as text-biased in Hindi and Telugu as in English.
  • High abstention can be a mask: Qwen followed the text 81% of the time when it failed to use the image.
  • Steering vectors trained on English transfer for abstention models, but fail for heavily text-biased ones.
Read the write-up on Substack

Hear when the next session goes up.

Subscribe on Luma

Programmes

Hubs