Update README.md

README.md

@@ -19,7 +19,7 @@ Pretrained weights for the **Disco103** meta-network from [*Discovering State-of
 
 A small LSTM neural network (754,778 parameters) that **generates loss targets** for RL agents. Instead of hand-crafted loss functions like PPO or GRPO, Disco103 observes an agent's rollout — policy logits, rewards, advantages, auxiliary predictions — and outputs target distributions the agent should match.
 
-Meta-trained by DeepMind across 103 complex environments (Atari, ProcGen, DMLab-30).
+Meta-trained by DeepMind across 103 complex environments (Atari, ProcGen, DMLab-30).  Originally in JAX, this is a PyTorch port.
 
 ## Usage