9ParmAI — Generative AI Explained

Leonard Brown began his career in mathematical modelling of inertial navigation systems, Kalman filtering, and under-characterized estimation (representing a high-dimensional system using a deliberately simplified model) before moving into quantitative finance, where he worked in derivatives and financial markets at J.P. Morgan. He holds a Bachelor of Science in Electrical Engineering (BSEE) from the Indiana Institute of Technology and an MBA from the UCLA Anderson School of Management.

Throughout his career, Mr. Brown has sought to make complex technical subjects understandable through two complementary approaches: identifying their underlying principles and building working, inspectable computer models that demonstrate those principles in practice.

He taught advanced financial derivatives at New York University’s School of Continuing and Professional Studies (NYU SCPS, now NYU SPS) and served as Head of Exotics and Hybrids Derivatives Education at J.P. Morgan. In teaching derivatives, for example, he emphasized the relatively simple concept of risk-neutral expectations (weighted averages) rather than relying on the complex mathematical machinery of stochastic calculus. He reinforced these explanations by constructing working models that allowed students to observe and explore the underlying mechanisms. Together, these approaches made sophisticated derivatives theory accessible to a much broader audience.

He applied the same philosophy to generative AI. His background in inertial navigation, mathematical modelling, and recursive estimation led him to consider whether the fundamental processes of generative AI could be explained without the complexity of large language models. Rather than beginning with LLM technology, he sought to identify the underlying principles and build a small, fully inspectable system that demonstrated those principles in operation. That approach ultimately led to the development of the Tiny Inference Model (TIM), a nine-parameter generative AI system that does not rely on LLM technology. As with his approach to teaching derivatives, his objective was to make generative AI, its underlying mechanisms, capabilities, limitations, and broader implications accessible to a much wider audience.

Consistent with his approach to teaching, the book combines accessible explanations with mathematical rigor. A detailed understanding of the mathematics is not required to grasp the fundamental concepts of generative AI presented in the book; a conceptual understanding is sufficient. Moreover, the more advanced mathematics is largely confined to the learning process, particularly the recursive least squares (RLS) algorithm. Encoding requires only elementary mathematics, while the generation loop itself is computationally straightforward.

For readers who wish to explore the mathematics in greater depth, the mathematical appendices develop the necessary concepts step by step from first principles, with minimal assumptions about the reader’s mathematical background. For example, the appendices begin with the calculation of simple averages and progressively develop the mathematics underlying TIM’s matrix recursive least squares learning algorithm. Although some readers may find the early steps elementary or repetitive, this deliberate approach is intended to make the material accessible to readers with widely differing mathematical backgrounds. Readers can therefore understand not only how the system learns, but why the mathematics works, examining the complete learning process rather than treating it as a black box.

Additionally, all mathematical appendices are available for download in AI-readable formats, enabling readers to use AI tools to explore the material, request explanations, and work through mathematical concepts at their own pace and level of understanding.

Building on this foundation, the book explores broader questions concerning the reuse and repurposing of AI systems, computational universality, and recursive self-improvement. Underlying these discussions is the conviction that important decisions about the development, use, and regulation of AI should be grounded in knowledge and understanding of the fundamental principles of generative AI, rather than assumptions and speculation.

Mr. Brown makes no claim to be an expert in artificial intelligence, particularly large language models. His aim was not to replicate an LLM's capabilities, but to reproduce, on a deliberately small scale, some of the fundamental processes associated with LLMs—encoding, learning, and inference—using a system whose operations can be directly observed, examined, and understood.

Now retired, he pursues independent research into generative AI and accounting, a discipline whose underlying mathematical structure is far richer and more interesting than its everyday practice might suggest. His accounting research, documented in working papers available on SSRN, has included deriving double-entry accounting from first principles and developing more effective and efficient mathematical representations of double-entry accounting. Consistent with his approach to teaching and research, most of these papers are accompanied by working computer models that demonstrate the underlying principles.

Developing TIM and writing Learn Generative AI Without LLM Complexity proved to be an extraordinary learning experience, and he hopes readers will find it equally rewarding.

Additional materials, interactive demonstrations, and working implementations are available at 9ParmAI.com.