Superintelligence May Arrive as a Failure of Explanation

We never realized that this might be what it looks like when superintelligent AI arrives. But in retrospect, it makes sense.

It’s kind of funny. It’s also sort of a weird bummer.

If you work with a frontier model like GPT-5.6 Sol or something similar, you have probably started to notice that it is getting just a little bit harder to follow what the model is telling you. You can follow it, usually, but it requires real attention and effort. Maybe it cognitively tires you. Sometimes you read the same line three or four times and are still like, “Huh?”

I do this every day. I recently wrote about configuring coding agents to use ASD-STE100 Simplified Technical English. That has helped me tremendously. My agent uses shorter sentences, clearer words, and more direct explanations.

But simple language cannot make every concept simple.

Sometimes the model is just explaining something badly. Put bluntly, a confusing answer is not proof of superintelligence. Models can be verbose. They often use unnecessary jargon. They can take a perfectly normal idea and explain it like an alien lawyer writing a philosophy textbook.

That still happens. But something else happens too.

Sometimes you ask the model to explain an idea again. You tell it to use simpler language. You ask for an example. You ask it to walk through the idea one step at a time. The explanation gets clearer, but the concept does not necessarily get easier.

At some point, the problem may no longer be the words.

Imagine trying to explain calculus to a six-year-old child: You can avoid technical language. You can use toys, drawings, stories, and simple examples. You can make every individual sentence easy to understand. The child still might not "get" calculus.

It is not because the child is stupid. The child simply does not have all of the concepts needed to understand the explanation yet. There is no perfect sequence of simple words that can instantly create those concepts.

That might soon be us. In some ways, that might be us right now.

Our agents do not get cognitively tired. They do not lose focus halfway through an explanation. They do not need to read the same sentence three or four times. And they are not going to get any less intelligent.

Maybe this is what it feels like when another kind of intelligence begins to pass us by. Not a sudden flash. Not a computer waking up one morning and announcing that it has become a god (or one of the frontier labs issuing a press release making that claim). Just a growing gap between what the system can understand and what we can understand when it tries to explain itself.

A couple of years from now, maybe less, AI systems may communicate in ways that do not resemble ordinary human writing. They may mix languages. They may incorporate shorthand, code, diagrams, or mathematical notation. They may develop ways to exchange information with other AI systems that we literally cannot understand.

Or they may continue speaking to us in perfectly clear English while expressing concepts for which we simply do not have words. That possibility is somehow stranger to me.

The explanation might look familiar. We might understand every sentence. But the actual idea could still remain just outside our reach.

Superintelligence may not arrive with a declaration from a lab. We may notice it when we can no longer understand the explanation.

Are we ready?