How AI Really Works, and Why Nobody Can Fully Explain It

If you are somewhere in your fifties and you have quietly decided this whole thing passed you by, this one is for you. No jargon, no doom, and no sales pitch. Just the actual mechanics, and then the part that genuinely surprised me, which is that the people building these systems say in public that they cannot fully explain them. This is the write-up of episode one of AI For People Who Missed the Meeting. The video is above.

Is it too late?

No. And here is the thing nobody says out loud: most of the people talking about it do not understand it either. They understand the vocabulary. That is a different skill.

So let us start where it actually starts, which is one idea small enough to carry around.

What is it actually doing?

It is predicting what comes next.

You type something. It works out the most likely next piece of writing. Then it does that again. Then again, very fast, until it has produced an answer.

That is the engine. Everything else is scale: an enormous amount of text, and an enormous amount of computing power pointed at it.

Where did it learn all that? From human writing. Books, websites, forums, manuals, arguments, instructions. All of it.

So is it quoting people? Not usually. It learned the patterns, not the pages. It is closer to how you learned to talk than to a library card catalog. Nobody handed you a book of sentences. You absorbed how sentences go, and now you produce new ones you have never heard.

And when you ask it something, is it looking that up? Sometimes, if the product around it is built to search. But the thing underneath is not looking anything up. It is producing what a good answer usually looks like.

Sit with the difference between those two for a second, because everything else in this article comes out of it.

That sounds like a problem

It is *the* problem.

It can hand you something that looks exactly like a right answer and is completely wrong. Same confidence. Same clean formatting. Same authoritative tone.

Why would it do that? Because it was built to sound right, not to be right. Those two things overlap most of the time. Most of the time is not all of the time.

So the working rule is simple, and it has served me well:

• Trust it on shape. Structure, outlines, first drafts, "here are seven ways to think about this."
• Check it on substance. Facts, names, numbers, dates, citations, anything with a decimal point in it.
• If you would be embarrassed repeating it, check it.

That single habit separates people who get real value out of these tools from people who eventually get burned in public. I went through the business version of that discipline in double check AI numbers before publishing.

Does it remember me?

The model underneath does not. Every conversation starts from nothing.

Some products bolt a memory layer on top, and that is genuinely useful. But understand what it is: a company built a filing cabinet next to the machine and taught the machine to check the cabinet. It is a product decision, not the machine deciding you are worth remembering.

So it is not learning from you personally in the moment, no. Whether your conversation gets used later to train something is a completely separate question, and it is a policy question. The answer is in the fine print, not in the technology.

Which brings up the only guardrail that regularly costs people money: be careful what you type into it. Same rule as anything else you would not email to a stranger. If a competitor reading it would cost you money, do not put it in a tool whose retention policy you have not read. I laid out the full version in AI and data privacy for Santa Clarita businesses and the policy side in AI governance for small business.

Alright. Is it thinking?

This is the question everybody actually wants answered, and I am going to give you the real answer instead of a comfortable one.

Nobody knows.

And I do not mean that as a dodge. I mean the people who build these things say so in public, under their own names, while running the companies.

The essay that should have been bigger news

In April 2025, Dario Amodei, who runs Anthropic, one of the largest AI companies in the world, published an essay called The Urgency of Interpretability.

Here is his sentence, and I am quoting it because paraphrasing would soften it:

"When a generative AI system does something, like summarize a financial document, we have no idea, at a specific or precise level, why it makes the choices it does."

Read that again with the byline attached. That is not a critic. That is not a journalist. That is the chief executive of the company that builds the thing, on his own website, under his own name.

How is that even possible? Because of how it gets made.

Amodei quotes a line from his co-founder Chris Olah that I have not been able to stop thinking about since: generative AI systems are grown more than they are built. Their internal mechanisms are emergent rather than directly designed.

He extends it with a plant analogy. You set the high-level conditions. You control the soil, the water, the light. You do not control where every branch ends up.

That is not how software works, and that is exactly the point. With normal software, a person wrote every rule. Somebody typed the instruction, and if it misbehaves you can go read the line that did it. With this, nobody wrote the rules. The system found them.

That is unsettling, and here is what it is not

It should be unsettling, a little. But notice carefully what it is *not*.

It is not a monster waking up. It is not a mind deciding to keep secrets. It is people shipping something faster than they can explain it. That is a much older story, and it is a human story, with human names attached to every decision in it.

Is anybody working on the problem? Yes, and Amodei put a clock on it. He wrote that on its current trajectory he would bet strongly on interpretability reaching the point of a true "MRI for AI" within 5 to 10 years, while worrying that AI itself is moving fast enough that we might not have that long. Anthropic's own stated goal is for interpretability to reliably detect most model problems by 2027.

Which tells you where it actually stands right now, in 2026: they mostly cannot.

Then something changed, in July

And this is the part I find genuinely fascinating rather than frightening.

On July 6, 2026, Anthropic published research on what they call a global workspace in language models.

Start with the human side, because that is where the idea comes from. Some of your brain activity is what neuroscientists call consciously accessible: the image that pops into your head, the plan you make about where to go shopping. You can describe it, control it, reason with it deliberately. The overwhelming majority of what your brain does is not like that at all. It just happens.

What the researchers found is that a similar distinction appears to have emerged inside the model. A small collection of internal patterns that, compared to everything else going on in there, play a special role. A workspace. A small area holding what it is currently working on.

Two details make this worth your attention.

Nobody designed it. No engineer sat down and built that room. It formed on its own during training, and they only found it afterward by going and looking.

And the model works through steps inside that space that never appear in the answer it hands you. On multi-step problems, part of the reasoning happens in there and never makes it onto the page.

"So it is hiding something"

That was my first reaction too, and I want to talk you out of it, because it is the wrong frame.

You do the same thing. You do not narrate every thought before you open your mouth. The gap between what you think and what you say is not deception, it is just how thinking works.

The real difference, and this is the genuinely new part, is that they built a tool that can look at some of it. They can read that workspace, and they can steer it. That capability did not exist before.

So is it conscious or not?

Here is why I trust this particular piece of research more than the ten scarier headlines you saw this week.

They had every commercial incentive in the world to imply their product might be alive. A company sitting on research that draws a direct parallel to human conscious access could have written a very different press release.

They did not take it. They drew the parallel as a functional one, about how information gets routed and used internally, and they were explicit about separating the ability to report on internal states from actually having an inner life. They claimed the first. They did not claim the second.

A company that had the chance to say its product might be alive, and declined, is a company whose research I will read carefully. That restraint is the credential.

I went at the same question from the philosophical side in is AI really going to destroy us, that is two questions and the machine-agency side in AI may be our friend, I know some who are not.

So what do you actually do with all this?

Four sentences cover about 90 percent of it:

1. Use it for drafts, summaries, and the boring repetitive work. That is where it is genuinely, immediately good, and where a mistake costs you nothing.
2. Check anything factual. Names, numbers, dates, citations. Every time.
3. Keep your private business out of a free tool whose retention policy you have not read.
4. Do not put it anywhere a wrong answer reaches a customer without a human in between.

If you run a shop and you want the practical version of that rather than the philosophical one, the audit I run on a business is walked through in the AI systems analysis audit, and the reason most fixes are not AI at all is in you cannot automate a process nobody has written down.

And the rest?

Stay curious, and stay skeptical of anybody who is certain in either direction.

The loudest voices on both ends want something from you. The doom side wants your attention and your vote. The lab side wants your subscription. Both are selling certainty, and certainty is the one thing genuinely not available right now.

The people closest to this technology say they do not fully understand it yet. They published that. Under their own names.

You are allowed to say it too. That is not ignorance. In September 2026, on this particular subject, it is the most accurate position available, and it puts you in better company than you think.

You did not miss the meeting. The meeting is still going.

AI for everyone. Not just the wealthy.

I'm Connor with Honor. Be safe out there.

Common questions

Connor T. MacIvor · CalDRE #01238257 · Sync Brokerage, Inc. · DRE #02031490