Around 370 BC, Plato wrote down a story he attributed to Socrates, who said he was retelling an old Egyptian myth.

The god Theuth was an inventor. He had given his people number, calculation, geometry, astronomy, even the games of draughts and dice. The story finds him traveling to the king with his latest creation. He laid it before the throne and called it grammata — writing.

He brought it as a gift. He called it an elixir. A remedy, he said, for both memory and wisdom. His people would become smarter. They would remember more.

The king said no.

His name was Thamus, ruler at Thebes, and his answer would become one of the most uncomfortable predictions in the history of cognition. You, Theuth, are the inventor. You are not the right judge of what you have made. Your elixir will not strengthen memory; it will weaken it, because people will stop training their own minds and trust the marks on the page instead. They will read widely and understand little. They will appear wise without being wise. The worst of them will be insufferable, full of the conceit of wisdom rather than wisdom itself.

That conversation is twenty-four hundred years old. I have been thinking about it all week.

Every practitioner who opens Claude or ChatGPT or Cursor is being offered the same gift Theuth offered Thamus, and from the same kind of source. Anthropic, OpenAI, Google are inventors. The product is brilliant. The grand claim is identical: a remedy for memory and wisdom. A faster, sharper, more knowledgeable you.

Consider the scale of what is being placed before us. Reports following an April announcement put Anthropic’s new Claude Mythos at roughly ten trillion parameters, a Mixture-of-Experts model with hundreds of billions of weights active on any given query. A step change from anything that came before. Anthropic has not confirmed the figure. What it has confirmed is what the model can do.

In a coordinated security exercise, Mythos identified thousands of previously unknown vulnerabilities, including a 27-year-old flaw buried in OpenBSD that had survived decades of human review. In more than 83% of cases it produced a working exploit on the first try. Theuth offered an elixir that could record what humans already knew. The modern elixir is finding things humans never knew, in code humans had been writing and reading for thirty years.

That is the second thing today’s gift adds to Theuth’s. Plato’s Theuth offered storage. The modern tools offer generation. They do not just remember; they reason, discover, build. The painting analogy collapses under that fact. A book cannot find a 27-year-old bug. A book cannot write a counterargument to your argument. The thing Plato called frozen has begun to think.

The first lesson of Thamus is that the inventor is not the best judge of the invention. Not because the inventor is dishonest, but because the inventor sees only the upside. The downside reveals itself in users, slowly, and almost always too late to argue about.

So the question is no longer whether AI is a gift. The question is what kind of judge you are going to be of what it does to you.

Plato had a second objection, and it is the one that should interest a practitioner most. Socrates added it after Thamus left the stage. Written words, he said, are like paintings: they seem alive, but if you question them, they just say the same thing over and over. A book cannot read you, push back, or ask what you actually meant. The real instrument of learning was the Socratic dialogue: the slow, adaptive, irritating back-and-forth between two minds. Writing was a frozen substitute for it.

This is where AI breaks the analogy.

A model answers back. It adapts. It can argue with you, refuse you, surface things you didn’t ask for. Used well, it is closer to the Socratic dialogue than any book has ever been. Used poorly, it is the conceit of wisdom on demand.

Nothing about the technology decides which one it will be for you. There is only a practice.

The Socratic method is how I think and how I lead. If you have sat across a table from me, or watched me review a document, you have probably felt me asking questions you weren’t expecting. The curiosity is genuine. It is also doing two things at once: refining the proposal in front of us, and teaching the person on the other side how to refine their own proposals next time. Most of the time, I value the learning above getting the right answer. Most of the time, the conversation is the work.

I have been practicing this with AI long enough to gain appreciation for how it has expanded my understanding. The first project I built with my early agent Aria was a self-improvement loop — a personality and a memory that compounded across conversations rather than starting from zero each time. That pattern now lives in the MEMORY.md and LESSONS.md files inside my Claude and OpenClaw setups. When I noticed that Claude Code still needed a nudge to capture what it had learned at the end of a session, I wrote a small skill called session-close that pokes the model to reconcile its plans and discoveries into memory before shutting down. The whole point of these tools is to make sure the dialogue compounds instead of evaporating.

The strangest habit I picked up, and the one I recommend most, is asking the model to criticize me. I draft something — an article, a plan, a strategy. And then I hand it back and ask: what is wrong with this? It feels backward at first. Most people use AI to confirm what they already think. Practitioners use it to disconfirm (see council-review). The same model that can produce the conceit of wisdom on demand will, if you ask it the right way, dismantle your own with the same speed.

Proverbs has a line that has been on my mind:

The book is older than Plato, and it answers Thamus’s worry from the other direction. The danger of the elixir is not the elixir. The danger is using a sharpening tool as a comfort tool.

Thamus’s prophecy was conditional. Both halves were always going to come true. Writing did weaken the great oral memory traditions; it also built science, built law, built every civilization that has followed. AI will atrophy the practitioner who outsources, and sharpen the practitioner who challenges.

So this week, do one thing.

Pick a piece of work you are about to ship. A document. A plan. A line of code. Before you send it, hand it to the model and ask it to find what is wrong with it. Ask it for the counterargument. Hand it the question you are afraid of being asked. Then read what comes back, and notice the part of you that wanted to skip that step.

That part is the conceit of wisdom. That part is what Thamus was warning you about.

The inventor is not the best judge.

You are.

Originally published on LinkedIn.