Deterministic AI: Why LLMs Aren't, and How to Build Products That Are Anyway (broutonlab.com)

Thinking Machines' recent experiment highlighted a critical gap in the reliability of large language models (LLMs) by demonstrating that even when using a fixed temperature setting aimed at achieving deterministic outputs, variable results still occurred. By prompting an open model 1,000 times with the same input, they received 80 different responses, raising alarms about the unpredictability inherent in LLMs, which don't guarantee the same output for the same input. This non-determinism can lead to significant issues, especially in scenarios where users must depend on consistent decision-making, such as finance or compliance. For the AI/ML community, this reveals a compelling need to differentiate between deterministic and probabilistic AI systems. While deterministic models, such as decision tables and fixed classifiers, provide consistent outputs and are easy to audit, LLMs excel in understanding unstructured data but can introduce variability that complicates decision-making processes. The article offers strategies for developers to integrate both types effectively by using hard-coded rules for critical decisions, employing calibrated classifiers for high-volume tasks, and constraining LLM outputs to enhance reliability. Ultimately, it calls for a nuanced approach to system design that emphasizes repeatability and transparency, fundamental for building trust in AI applications.