Legendary computer scientist Donald Knuth was recently in for a "shock" when an open mathematical problem he had been working on for several weeks was successfully solved by Anthropic's Claude Opus 4.6. In this episode, we dive into the fascinating story behind "Claude's Cycles," exploring how this generative AI showcased a dramatic advance in automatic deduction and creative problem-solving.

The complex problem, intended for a future volume of The Art of Computer Programming, involved finding a general decomposition of a specific digraph's arcs into three directed Hamiltonian m^3-cycles. After Knuth solved it for m=3, his friend Filip Stappers challenged Claude to find a generalized solution. Guided by strict instructions to document its progress, Claude worked through 31 distinct algorithmic "explorations". Moving from simple depth-first search and simulated annealing to "serpentine patterns" and fiber decomposition, the AI eventually realized it needed "pure math" to discover a working solution for all odd values of m.

Join us as we recount Claude's impressive 60-minute analytical journey, discuss the 760 perfectly valid "Claude-like" decompositions, and see how Knuth rigorously proved the AI's brilliant discovery. Hats off to Claude!

Read Don Knuth's original paper here: https://cs.stanford.edu/~knuth/papers/claude-cycles.pdf

There's an incredible buzz around AI agents, with the prevailing wisdom suggesting that bigger is always better. The industry has poured billions into monolithic, Large Language Models (LLMs) to power these new autonomous systems. But what if this dominant approach is fundamentally misaligned with what agents truly need?

This episode dives deep into compelling new research from Nvidia that makes a powerful case for a paradigm shift: the future of agentic AI isn't bigger, it's smaller. We unpack the core arguments for why Small Language Models (SLMs) are poised to become the new standard, offering superior efficiency, dramatic cost savings, and unprecedented operational flexibility.

Join us as we explore:

• • Surprising, real-world examples where compact SLMs are already outperforming massive LLM giants on critical tasks like tool use and code generation.

  • The key economic and operational benefits of adopting a modular, "Lego-like" approach with specialized SLMs.

  • A clear-eyed look at the practical barriers holding back adoption and the counter-arguments from the "LLM-first" world.

  • A concrete, 6-step roadmap for organizations to begin transitioning and harnessing the power of a more agile, cost-effective SLM architecture.

This isn't just an incremental improvement; it's a potential reshaping of the AI landscape. Tune in to understand why the biggest revolution in AI might just be the smallest.

The research paper discussed in this episode, "Small Language Models Are the Future of Agentic AI," can be found on arXiv:
https://arxiv.org/pdf/2506.02153

It looks incredibly impressive when a large language model explains its step-by-step thought process, giving us a window into its "mind." But what if that visible reasoning is a sophisticated illusion? This episode dives deep into a groundbreaking study on the new generation of "Large Reasoning Models" (LRMs)—AIs specifically designed to show their work.

We explore the surprising and counterintuitive findings that challenge our assumptions about machine intelligence. Discover the three distinct performance regimes where these models can "overthink" simple problems, shine on moderately complex tasks, and then experience a complete "performance collapse" when things get too hard. We'll discuss the most shocking discoveries: why models paradoxically reduce their effort when problems get harder, and why their performance doesn't improve even when they're given the exact algorithm to solve a puzzle. Is AI's reasoning ability just advanced pattern matching, or are we on the path to true artificial thought?

Reference:
This discussion is based on the findings from the Apple Machine Learning Research paper, "The Illusion of Thinking: Understanding the Strengths and Limitations of Large Language Models with Pyramids of Thought."
https://machinelearning.apple.com/research/illusion-of-thinking