Chapter 6. If You Want More

In teaching terms, you just passed the qualification exam. Shifting your way of thinking is difficult and, frankly, mentally exhausting. You have not yet mastered anything. Mastery comes with deliberate practice coupled with close observation. But you now have names for things that you have already mastered. I have found that latter skill remarkably valuable.

These are real accomplishments, and I do understand the effort involved. You already have the skills to evolve what you just learned into true mastery. That means, as you will have guessed, there is more.

You are now standing at the start of the next cycle of mastery. I see mastery as cyclical, not a linear progression from "A" to "B." Completing one cycle of mastery enables you to begin the next. That fact, I find, keeps things fun and the boredom at bay.

But we just hit a barrier.

The Barrier

The next cycle of mastery involves nearly-lost art. This art was taught to children, and in my case, elementary school and middle school.

The rise of Artificial Intelligence usage shows this area of mastery has become surpassingly important once again. I therefore created an apprenticeship for you (as a pair of highly unusual books) that draws you through to experience the skills and techniques (and attitudes) before naming them. That is a continuation of the already-familiar technique you just encountered: experiencing via demonstration before naming what was demonstrated.

You have already taken strong steps in this direction by learning to model your own thinking, and by learning to spot transcendent patterns in the wild. The pioneers of computing learned to do this so automatically and habitually, that none of us thought to write down how we did it. (I say "we" because I worked amongst, and learned directly from, those pioneers.)

Transmitting the craft proved to be a large undertaking. That is the barrier. The result is a pair of books to be read in order (the one is prerequisite to the other). They are relatively large, and that is the challenge.

Capabilities

At Cray Research, when we faced the impossible, or at least something that had never been done before, we took stock of the situation. That is the three-step process I showed you: identify the system and its constraints. The key overriding constraint becomes the point of leverage.

This approach is counterintuitive. When there is a barrier, our first instinct is to remove or soften the barrier. At Cray Research we asked, "how can we turn this barrier into a decisive advantage?"

We did not immediately jump to solutions. Instead, we carefully measured and identified our actual capabilities (or what those capabilities will have to be). If it was outside our capabilities, it was obviously not a viable solution. The capabilities told us where to look for solutions.

Observation. Large Language Models, in my experience (Claude and ChatGPT), are designed to immediately jump to solutions. When you learn to use AI as an accelerant, you will also need to fight this tendency. Understand capabilities before considering solutions.

This "capabilities" approach all sounds simple and logical, and it is, and this is literally how we built the world's fastest computers, and how we stayed in front.

Solutions

After I wrote the two books, I realized I had a challenge: who should read them? How does one transmit a nearly-lost art, when people don't even know the lost art exists?

That barrier suggested a solution: what was needed was the ability (and willingness) to think differently and to look at things differently. You already know that if you made it to this paragraph by reading everything, without skimming, you are one of those people. You qualified yourself.

What is being lost? Holistic and systems thinking, and forming judgement when there are no do-overs or turning back. Naval officers, for example, will recognize this as learning to make operational decisions under adversarial conditions. I find the same ability applies day to day.

A key application of these skills is working with Artificial Intelligence. Why? Large Language Models are systems whose behavior can be observed and understood, if you have the skill and expertise to do so. Chapter 4 demonstrated that exact concept. We learned about AI behavior without needing to have the slightest idea of how LLMs work mathematically.

That is why I designed two books. One uses AI as the mechanism for teaching these skills. The book appears to be about AI, but it really is an apprenticeship. This book demonstrates the exact same Cold War-era skills right now in 2025-2026.

The other book is the historical narrative showing how these critical skills developed in the first place, and why. For high tech people, this is the possibly-unknown background explaining why things work today as they do. It's a fascinating history beginning in 1904. (A specific 1904 event is indeed the correct starting point for the direct path to supercomputing!)

The two books are:

Nobody but Us: A History of Cray Research's Software and the Building of the World's Fastest Supercomputer
The Wizard's Lens: Learn to Think Like AI

If you want more, please do check the first book or both. Take your time. Skimming, by design, will fail. That is not how apprenticeship works. Both books directly state that I am making no attempt to bring you "up to my level." The books exist so that you can pass me by.

You will do well.

Cognitive Design of This Book

Think back over what you just accomplished, and how you accomplished it.

You gained new ways of thinking. Or, more likely, you can now put a name to tacit skills you already had. That is a good thing.

Separate Concept From Implementation

In Chapter 4, you learned that how AI "thinks" is actually quite similar to how humans think. In formal terms, you spotted transcendent patterns that apply to both human and AI. That skill, spotting transcendent patterns that apply to both human and AI, has tremendous value. Do you see why? If you spot the pattern, and since you of course understand the human-side implications, you suddenly have insight into Large Language Models that even AI experts might not have. Simply by spotting the same pattern.

Donald Michie did something similar 1960-1961 with MENACE, a project that enabled matchboxes to learn to win at tic-tac-toe. He demonstrated machine learning without using mathematics or electronics. We just did the same thing by standing at the Breaks Interstate Park overlook. Do you see the invariant, the common pattern? We separated the concept from the implementation. Michie, obviously, did it first. Mine is not an original solution.

Parallel Lineage

There is an interesting lineage at play here. Michie worked with Alan Turing at Bletchley Park during World War II, on the Tunny project breaking the Lorenz cipher. He built MENACE as a non-classified demonstration because he had no access to non-classified computers.

The U.S. had a parallel codebreaking effort, Navy's OP-20-G. After the war, several of those codebreakers founded Engineering Research Associates, which then hired Seymour Cray. Cray later founded Cray Research, where I learned similar skills from within the parallel environment. I believe it is no coincidence that I found a solution similar to Michie's. It was common to model concepts before designing or implementing solutions. We often worked from analogies because the technology simply did not yet exist.

I transcribed two declassified Top Secret documents introducing this tradecraft that originated with Bletchley Park and OP-20-G, and continued through Cray Research:

Appendix A, "Analytical Machine Employment (1952)" distinguishes between labor-saving machines and revolutionizers. This discussion of allocating brainpower strongly applies today.
Appendix B, "Cryptanalytic Machines in NSA (May 1953)" is first-hand revelation from Dr. Howard Campaigne, who worked at Bletchley Park, within OP-20-G, and the NSA.

Novelty

I am unable to find any similar explanation of LLM behavior in the open literature that separates concept from implementation. My explanation of LLM behavior without mathematics might be novel.

Here is my point. I demonstrated a capability: create something possibly novel in line with Cold War-era engineering tradition. At Cray Research, before proposing solutions, we first measured actual capability.

Content as Method

I designed this short book where the content is the method. That approach appears to also be novel or nearly so. I demonstrated this capability so that I can assure you that both large books demonstrate this same capability from the first page to the last.

Appendix C, "HPC Tradecraft Road Map" lays out both series of three books each, plus the companion website allowing you to walk through constraint-based design.

Creating Agency

Finally, I designed the experience as a qualification exam. I designed it in such a way that you know whether the large books are likely of interest, or definitely not. You respected me enough to get this far, and I do not want to waste your time on something you do not want or need. You are now able to make your own informed choices.

Once again, I designed and demonstrated a capability: one book demonstrating all those characteristics simultaneously. Both large books demonstrate these characteristics, cover to cover, 290,000 words between the two. That is a large undertaking, and apprenticeship-style learning is sometimes mentally exhausting. Fortunately, either book can be abandoned at any point without losing what you gained to that point. You can return later if you choose.

The path exists. The route has been traveled. The choice is yours.