Deterministic by Design: Why "Temp=0" Still Drifts and How to Fix It

Send us a text

Why do LLMs still give different answers even with temperature set to zero? In this episode of The Second Brain AI Podcast, we unpack new research from Thinking Machines Lab on defeating nondeterminism in LLM inference. We cover the surprising role of floating-point math, the real system-level culprit, lack of batch invariance, and how redesigned kernels can finally deliver bit-identical outputs. We also explore the trade-offs, real-world implications for testing and reliability, and how this breakthrough enables reproducible research and true on-policy reinforcement learning.

Sources:

• Defeating Nondeterminism in LLM Inference
• Non-Determinism of “Deterministic” LLM Settings