Sarathi models trained from scratch
Schnitt trains Sarathi from scratch, then keeps training it on every interaction it has. Interactions and their outcomes become training data for continued pretraining and RL environments, and the model is retrained. Not a bigger prompt. New weights.
Self-improving models are the next frontier We are betting big on RSI
A model that learns from every interaction it has, then learns to improve how it learns. That is where the frontier is going. Sarathi gets there in three steps.
Continued pretraining on every interaction and its outcome. No knowledge cutoff. The model on Friday is not the model it was on Monday.
Reinforcement learning in environments built from real outcomes, so the model is rewarded for what actually happened, not for what a rater liked.
A model that improves its own training. The frontier is racing toward it. So are we.
Better Capital, LetsVenture