On LLMs and Quicksort (criptae.substack.com)

Large Language Models (LLMs) mark a significant breakthrough in AI, offering capabilities that extend beyond traditional tools, even though they are far from achieving artificial general intelligence (AGI). Understanding LLMs can be approached in two ways: white-box analysis, which inspects their complex internal mechanics, or black-box usage, which treats them like reliable functional components akin to classic algorithms such as quicksort. While quicksort’s workings are fully transparent and easily implemented, LLMs remain opaque; yet, both successfully deliver valuable, practical results from given inputs. This black-box perspective highlights LLMs as embodying a novel software development paradigm rather than just a step toward AGI. By shifting the focus from whether LLMs will achieve human-like intelligence to how they change programming itself, the AI/ML community is invited to explore new problem-solving possibilities enabled by these models. Tracing back to concepts like Andrej Karpathy’s Software 2.0 and Gwern Branwen’s insights on GPT-3, this view positions LLMs as data-driven programming tools that complement, rather than replace, traditional software, opening fresh avenues for creativity and innovation in AI applications.