Summary
Physicist Sabine Hossenfelder examines a counterintuitive finding in recent AI research: the more consistent a large language model’s answers are forced to be, the less accurate it can become. The video explains why even models set to zero randomness still produce varying outputs, since chips round numbers differently depending on hardware and batch size.
She cites several studies. Researchers at Rice University found rounding differences alone shifted a small DeepSeek reasoning model’s math score by up to 9%. An Australian group found that changing a single bit in a small neural network’s starting parameters produced a different network in about 48% of cases, and that the most accurate models disagreed most about the answers they got wrong. Anthropic researchers observed that longer reasoning can make errors more random, and Nvidia demonstrated bit-for-bit reproducible training of a trillion-parameter model across more than 2,000 chips, though only for 800 steps.
Hossenfelder discusses what this means for AI safety, since more powerful models may have errors that are harder to predict and detect. She also considers whether the problem is specific to current LLM architectures or a general feature of AI.
📺 Source: Sabine Hossenfelder · Published November 21, 2026
🏷️ Format: Deep Dive







