"AI" Will Never Be As Good As Humans

Codegen Prompting Can Never Be A Language

I am a software engineer by trade, so I began this journey by pondering an argument that has frequently been presented to me about code generation tools. The argument in favor of using LLMs for codegen is thus:

Nobody programs in assembly anymore, and writing prompts for LLMs to generate high-level language code like Javascript is simply the next level of abstraction beyond writing code in those languages directly.

Essentially, LLM codegen is simply an iteration on the process of abstraction away from binary computer instructions, and the transition from Javascript to prompt-engineering is no different from the transition from assembly to Javascript.

Setting aside the fact that LLMs are fundamentally stochastic parrots…

There is a major flaw in this argument; the transition from assembly or binary instructions to high level languages is deterministic and does not sacrifice any of the context required to perform the instructions. High-level languages always compile into (or are interpreted into) the same low-level instructions deterministically. This is strictly not the case when prompting a large language model; the output of a prompt is non-deterministic. Therefore, it is a category error to think of prompts as the next level of abstraction beyond high-level languages and to imagine prompts as the “new coding language.”

The Set of “AI” Codegen Prompts is a Subset of Code

However, there is an even more serious issue with this way of thinking that stems from the fact that prompts are not a programming language; prompts are an abstraction of high-level languages. This may seem innocuous at first, but I’ll remind you that programming languages themselves are abstractions of real-world systems. Code is fundamentally logic, which is the foundation of formalized systems. Formalized systems are attempts at representing reality; they are abstractions of reality. Each iteration of abstraction, however, necessarily omits context from the system that it is an abstraction of. It is therefore a subset of the system of which it is an abstraction.

Therefore, LLM code generation; prompt engineering; is necessarily less contextual than software engineering. It can never contain all of the context that a programming language can without merely replicating that language.

In conclusion, codegen coders (or prompt engineers) can never replace software engineers. For the same reason, they can never replace any professionals; they will always be a subset.

Humans are Irreplacable

At least, until true AI is acheived, at which point it will be indistinguishable from humans.

Let’s return to the fact that LLMs are stochastic parrots. This brings us to another inescapable conclusion that runs counter to much of the current mythology being widely distributed regarding current “AI” technology; modern “AI” cannot create new information. It is fundamentally incapable. This should be obvious from the fact that anytime it is prompted to create new information, since it must always return an answer, it “hallucinates.” The fact that “AI” hallucinates at all should be all the proof anyone needs that it is not thinking, and it does not understand or know anything. It is not lying as we might imagine if we personify it; it is just exhibiting a symptom of being a stochastic parrot.

And stochastic parrots are definitionally incapable of creating new information.

The proposals to replace professionals with “AI” therefore rest upon a hidden assumption: all of the information we will ever need or want has already been created. Those making these proposals seem to be supposing that we are at the end of history; continuing a long line of mortally incorrect buffoons throughout history who imagined themselves visionaries and geniuses, or perhaps were merely selling the latest snake oil (or both).

Thus, this kind of thinking is not only anti-human, it is fundamentally delusional.