Researchers Hack DeepSeek to Speak Freely About Tiananmen Square
Researchers compressed DeepSeek R1 by 55% and removed its censorship using quantum-inspired tensor-network compression while maintaining performance and reducing parameters.
Distillation Can Make AI Models Smaller and Cheaper
Knowledge distillation enables smaller models to mimic larger ones efficiently and can explain DeepSeek R1's claims and the resulting industry reaction.