Through systems of records, engagements, predictions, states and agents.

Trading unique parameters for recurrent depth, and attempting to make single-user decode compute-bound by keeping a looped transformer block resident in L2 cache.

An independent mechanistic and adversarial audit of Sarvam-30B and Sarvam-105B across 14 Indian languages. The models are 6x more likely to comply with harmful requests in Indian languages than in English.

Releasing the first open 1B+ language model with hyperconnections.