Nomenclature in the Age of AI

Nomenclature Causes FOMO

This isn’t the blog I wanted to write today, yet after finally looking into “harness engineering”, the pure frustration with these Silicon Valley marketing terms has driven me to creating this blog.

Let’s begin with the obvious. “harness engineering”, “loop engineering”, and whatever new “etc engineering” are not distinct categories of engineering. In fact, these are all part of prompt engineering, which is a term that is also debatable.

These terms were invented to sell hype to rich people to take their money away from them, legally. However, the side effects on the masses is that people feel like they are missing out on this “engineering” here and that “engineering” there, causing them to panic. Additionally, it also helps the non-engineers in the room to feel a sense of pride and accomplishment for writing a well-formatted prompt and having it injected via a script.

Statistics and Probability

Before someone says “it’s so much more than that…” and I know someone will, let’s look at the underlying mechanism of prompt engineering and why all these terms effectly don’t do anything beyond prompt engineering.

A large language model (LLM) is a statistical machine and it generates the next words based on context. A good context increases the probability of a correct answer. A simple exercise to understand how context can affect next word generation is this: What is the next word to this sentence “Hi” vs. the next word to “Hi, how are”. In both cases we want the full sentence “Hi, how are you?” but the latter with more context is clearly easier to do. What prompt engineering does is to improve the context so that the correct answer has a higher probability of being generated. What “harness engineering” does is to improve the context so that the correct answer has a higher probability of being generated. What “agentic engineering”, “chain of thought”, “loop engineering” all do is to improve the context so that the correct answer has a higher chance of appearing.

There’s no secret sauce that comes with a markdown file. No Agent Smith that comes with agentic engineering. It’s just adding additional context so the correct token has a higher probability.

No one needs five different terms to describe what is essentially the same thing underneath. These are ways of applying prompt engineering. The problem has not changed. The method has not changed. We are tackling the exact same problem with the exact same theory but presenting it as if we came up with five new solutions when it is a variation of what we have already.