Distinguishing AI from Magic

Welcome to a new edition of The Board: Distillation Aftershots ^(*).

This is an online copy of the weekly newsletter, written by Constellation's Chief Distiller and Board Advisor Esteban Kolsky, that shares curious and interesting insights and data points distilled from enterprise technology to identify what’s notable. If you want to subscribe to the newsletter, please do so here.

In this issue, I want to address what we know today about AI (GenAI, more specifically) since the launch of the first commercial GPT nearly 40 months ago. How can we leverage the full strategic cycle (despite the AI Rush compressing cycles by 36 months, give or take) to create a better path forward?

First, my take

According to Wikipedia (yes, it is still useful…)

British science fiction writer Arthur C. Clarke formulated three adages that are known as Clarke's three laws, of which the third law – that "any sufficiently advanced technology is indistinguishable from magic" – is the best known and most widely cited. They are part of his ideas in his extensive writings about the future.

And this is what prompted me to write about AI being distinguishable from magic (or, in other words, not being “sufficiently advanced technology” yet).

While true that GPT’s can do a lot of interesting things, when not applied at enterprise level – where the value is reduced by the complexity of the problems to be solved at scale, the biggest value proposition is the “shiny bead” effect it has on consumers who were not exposed to advanced analytics before (with maybe the exception of Target finding out pregnant teens, ‘member?); GenAI is not sufficiently advanced.

There are three things GPT’s can accomplish better than previous AI solutions: 1) requires less training data (by reducing or eliminating ontology size, which requires less structured data to be available in short demand); 2) reduces the number of “tokens” required for complicated and complex inquiries (by making inquiries against a less structured resolution pool, meaning less search costs), and 3) makes a great, simple interface for people who are not familiar with building models in R, Python, or similar coding languages.

The ability to speak (or type) in plain language and not be reduced to keywords or key-phrases for search (yielding a better natural-language search interface instead) is the culmination of nearly 30 years since natural-language interfaces launched the seventh generation of AI tools. In other words, after nearly 30 years, we finally solved the problem of plain-spoken inquiries as input into AI models. This does not, however, solve the problem of finding the best answer or the right answer – just the “most statistically likely answer” to a query. GenAI improved access to AI more than it solved enterprise-grade resolutions.

There are three reasons why GenAI is not sufficiently advanced technology:

1. Intuition and imagination. We don’t yet know how to create mathematical models for things like intuition (the ability to infer missing variables in an equation) and imagination (the ability to replace them with another, similar but unrelated value). This is mostly because we don’t fully understand how these processes work in the human brain. This leads to incomplete queries being passed to a model that cannot solve them, resulting in hallucinations.
2. Deterministic and repeatable. GenAI is non-deterministic (meaning, it is not a perfect, repeatable resolution) because it only looks for the “most likely next word” in a phrase, without a real understanding of the concepts behind the inquiry. In technical terms, it has a categorization tool (a taxonomy) but lacks a definition tool (an ontology) that can correlate similar terms or concepts and replace them in an inquiry to preempt hallucinations. GenAI recognizes patterns and correlations but still leaves gaps in explicit semantic grounding, substitution logic, and enterprise context. Plainly stated, it does not understand the actual question, or the answer, only makes it “statistically relevant”.
3. Autonomy is granted, not earned. This is the most important one in the enterprise: to earn the ability to act autonomously, an entity must be trusted. To be trusted, one needs to have performed flawlessly (even making substitutions in related but not similar events, which we call common sense when used by a human) in different situations. And by the nature of the two preceding points made in this list, it cannot do that. Therefore, it won’t earn the ability to act autonomously; in the enterprise, the burden of proof rises with scale, and it won’t be trusted when providing results.

The enterprise does not need AI to be magical; it needs AI to be governable. Magic in enterprise terms is anything that cannot be audited, repeated, governed, or trusted at scale. A consumer error is an annoyance; an enterprise error propagates through process, policy, compliance, and customer outcomes.

I could conclude that it is impossible for AI to replicate humans, but never forget the other two laws.

Finally, Isaac Asimov was ahead of his time, no doubt, and he wrote a corollary to this third law, which is aptly appropriate here (if not, self-serving…)

When, however, the lay public rallies round an idea that is denounced by distinguished but elderly scientists and supports that idea with great fervor and emotion, the distinguished but elderly scientists are then, after all, probably right.

Here are some reading resources: