Keep Deterministic Work Deterministic (www.oreilly.com)

In a recent exploration of agentic engineering and AI-driven development, the author conducted a blackjack simulation where a large language model (LLM) played hundreds of hands based on English-written strategies. Initial results revealed a concerning 37% pass rate, primarily due to the LLM's erroneous computations, which compounded as each wrong decision cascaded through subsequent moves. This phenomenon highlights the "March of Nines," a term coined by Andrej Karpathy that illustrates the challenge of increasing reliability in AI systems; improving from 90% to 99% reliability demands significantly more engineering effort, while each step introduces new potential failures. The findings emphasize a critical lesson for the AI/ML community: while LLMs can handle complex natural language tasks, they struggle with deterministic operations, such as precise counting and rule application. For efficient AI pipelines—where outputs from one step inform the next—it's essential to minimize reliance on LLM calls for deterministic tasks. The author's experiences underscore the necessity of integrating robust deterministic components into AI workflows, ultimately demonstrating that while LLMs excel at interpretation and generation, they are not always the best tool for every part of a task, particularly in scenarios where precision is paramount.