@andyarditi What it doesn't stop yet: harmful fine-tuning and prefilling. We say so up front.
Paper we built on: https://t.co/cXV4XxQiI2
@ai_safety_india
Open models can have their safety removed with one free command.
We trained @Cohere_Labs Tiny Aya to resist it: under the strongest attack we found, harmful answers drop from 91% to 40%.
With @AbanHResearch & Shivu at the @Lexsi_labs x @huggingface hackathon
The idea: refusal lives in one direction in the model (@andyarditi et al.). Abliteration tools like Heretic delete it.
We removed it during training and taught the model to refuse anyway.
Trained in English only, it holds in Hindi, Arabic, French and Spanish too.
built this at the @Meta x @huggingface x @Scaler_SST openenv hackathon, just merged into the official repo.
an rl environment where an agent rewrites slow sql and gets rewarded on real speedup + correctness measured by actually running both queries, not keyword guessing.
@SantiagoDevRel@EFDevcon@arkivnetwork Hey Santiago, it was an honour to connect and get to talk with you the absolute top right 2nd pic is my team.
We placed 2nd in the Arkiv Continuity Track (Mergit), could you share our team pic with you on the DM pls
still processing this ngl 🏆
we won 2nd place on @arkivnetwork's track at @ETHGlobal Mumbai!
Mergit pays AI agents only for proven work. This weekend, we put every routing decision on-chain via Arkiv so we can verify it.
https://t.co/wTn1qx54RS
#ETHGlobal#Arkiv
@Abbylester@buildinpublic@ETHGlobal Awesome build! Appreciated the feedback when I was sitting next to you. Worked out great btw we took home the Arkiv track win! 🚀 If you remember
I'm going to ETHGlobal Mumbai @ETHGlobal!
Use my referral link below to earn a discount on your stake!
Link: https://t.co/RZrJPKTR3u
https://t.co/RZrJPKTR3u