Scientists Just Hacked AI’s Mind And The Truth Is Shocking

Researchers have discovered that AI model inconsistencies stem from server-side batching variability rather than floating-point errors, and have developed methods to ensure deterministic outputs, while AI is also proving capable of innovative scientific design and discovery.

• AI output inconsistencies stem from variable server-side batch processing, not floating-point errors.

• Researchers have engineered deterministic AI models by standardizing core operation calculations, crucial for reliable science and training.

• AI is autonomously redesigning complex physics experiments and uncovering fundamental scientific principles and equations.

AI models, even with zero temperature, produce inconsistent outputs not due to floating-point math issues but because server batching processes requests in varying orders, impacting internal calculations. Researchers have addressed this 'batch invariance' problem by redesigning core AI operations to ensure deterministic results, a crucial step for scientific reproducibility and reliable AI training. Concurrently, AI is demonstrating advanced capabilities in scientific innovation, redesigning complex experiments like LIGO and quantum entanglement setups, and discovering fundamental physical symmetries and dark matter equations.

AI Output Inconsistency Revealed

• 00:01:05 Despite setting AI models like ChatGPT to zero temperature to eliminate randomness, outputs can still differ for the same input, a significant problem for scientific reproducibility. Contrary to previous assumptions that blamed floating-point arithmetic quirks in GPUs, researchers found that these low-level mathematical operations are consistent when isolated. The true culprit is the higher-level 'batching' process, where AI servers group multiple user requests, causing internal calculations to execute in different orders depending on batch size and composition, leading to variable outcomes for individual prompts.

Solving Batch Invariance

• 00:03:32 The core issue, termed 'lack of batch invariance,' means an AI system's output changes based on whether a request is processed alone or in a batch. To fix this, researchers redesigned the calculation methods for three core transformer operations—RMS norm, matrix multiplication, and attention—to ensure they behave identically regardless of batch size. This involved forcing a consistent processing order even if it meant sacrificing some speed, making the model's outputs 'rock solid' and truly deterministic, a trade-off deemed worthwhile for inference consistency.

Impact on AI Training and Science

• 00:05:53 Achieving deterministic AI outputs, while slightly slowing down inference, is critically important for AI training, especially in reinforcement learning, where consistent model behavior between training and inference prevents 'training collapse.' For science, this breakthrough is fundamental, as repeatability is paramount for trusting experiments and comparing results. By making AI fully deterministic, this research paves the way for more reliable scientific inquiry, easier debugging, and more robust AI training, representing an 'invisible upgrade' that enhances foundational AI capabilities.

AI Redesigns Physics Experiments

• 00:09:51 Beyond achieving determinism, AI is being used to innovate in scientific design, as demonstrated by its application to complex physics experiments. Researchers used AI to redesign LIGO, the gravitational wave detector, discovering a novel optical ring addition that could have increased its sensitivity by 10-15%, exploiting obscure quantum noise reduction principles previously overlooked by human scientists. Similarly, an AI system called Pytheist generated a simpler, more efficient design for a quantum entanglement swapping experiment, which was later experimentally confirmed, proving AI's capacity for genuinely new scientific inventions.

AI for Scientific Discovery

• 00:12:03 AI is also proving adept at uncovering hidden patterns and generating scientific theories from data. At the Large Hadron Collider, machine learning models rediscovered Lorentz symmetry from raw data without explicit programming of physics principles. In astrophysics, AI generated formulas for dark matter clumping that fit data better than human-derived equations. While these models do not yet explain the underlying theory, they signify AI's potential to spot fundamental structures in chaos, hinting at a future where AI could propose hypotheses and contribute to theoretical physics itself, becoming a true partner in scientific discovery.