[TimByte]

The scary part isn't "LLMs doing sums." It's that the same deterministic model, same weights, same prompt, same OS, produces different floating-point tensors on different devices