The AI Coverup: Have We Been Hallucinating?

Hallucination

Anthropic's Claude Opus 4.7 fails on this course title, just seven words. You can see the Claude output in the sample lesson. The industry calls this result "hallucination" but Claude is working correctly as designed. You do notneed AI internals knowledge to see how this works.

When using AI as a tool, we expect the same input to produce the same output. That is how computers work. We have been using spreadsheets, word processors, power points, and email for decades. We know what to expect. What happened?

Spreadsheets and word processors are deterministic tools. Same inputs produce the same output. We trust the result to be correct. But AIs like ChatGPT or Claude are probabilistic computing systems. They have no concept of what is correct! Instead they produce what is judged most likely to be correct. The experts either forgot to tell us this crucial difference, or we did not hear the message. The experts compounded the error by calling incorrect results "hallucinations" as if the AI were not thinking right.

Mastering any skill, including diagnosing AI failure modes (hallucinations), requires mastering prerequisite skills. This is why that skill is in Course 3 of this series.

Biggest Computing Systems Ever

During World War II, the Allies won partially due to using large computing systems to secretly read Axis communications. That experience with large computing systems continued through the Cold War. Across that 50-year period 1941-1991, the largest computing systems, applied to codebreaking, produced probabilistic results *most likely to be correct*. That same knowledge applies today!

Frontier-class computing means you cannot simply spin up more hardware because you are already sitting on top of the largest computing systems available. You need to understand and work with what you have. In frontier-class computing, the scale does not matter. Holding on to first place matters, and the fact that you cannot just order more hardware matters.

Cray Research famously inhabited this frontier. We repeatedly accomplished what nobody else on the planet could do. Modern AI is beginning to hit this frontier-computing zone where you cannot just add more hardware. What Cray Research accomplished is well known. How we did it was never written down, until now. This tradecraft was originally Top Secret, and became tacit knowledge passed from person to person. This knowledge will be interesting to leading-edge AI and other frontier-computing customers once its existence becomes known. Those skills, and our ways of thinking, are Course 2.

Creating Tacit Knowledge

Cray Research carried an elite engineering tradition. We now know it originated with code breakers at Bletchley Park in the 1940s, including Alan Turing and Donald Michie. Donald Michie's MENACE (1961) taught matchboxes and beads to win at tic-tac-toe. This was the first Machine Learning demonstration without electronics. Michie separated concept from implementation.

Two U.S. Naval officers from codebreaking unit OP-20-G, James Pendergrass and Howard Campaigne, visited Bletchley Park. After the war, OP-20-G members shifted to the NSA and predecessors, and other members founded Engineering Research Associates in Minnesota. ERA hired Seymour Cray. Bill Norris went on to found Control Data Corporation, with Cray soon following, and Cray founded Cray Research in 1972. This was the Top Secret chain of tacit knowledge transmission.

At Cray Research, "it has never been done before" was never a barrier. It was our necessary starting point. Like MENACE, we often separated concept from implementation because the technology did not yet exist. Course 3 constructs functional equivalents to Large Language Models without electronics. While this might be a startling and novel way of doing things, my tradecraft is just one generation removed from Donald Michie. We share a parallel lineage with frontier-computing institutional experience.

Transmitting Tacit Knowledge

Course 1 focuses on how we taught tacit knowledge. One method is verbally modeling expert thinking. Rather than answering a question with a fact, walk through your thought process (out loud) showing how you arrive at that answer.

This three-course series is an apprenticeship in written form:

  1. This course teaches how we teach. An apprenticeship in written form is highly unusual, which is why this first course exists.
  2. Course 2 is frontier-class computing design, analysis, and debugging, beginning with the reason frontier computing became necessary (to win the war).
  3. Course 3 is close observation and characterizing frontier-class computing systems, using AI as the working example.