Plato wrote of a dialogue between Socrates and Phaedrus, where Socrates (one of my favorite arrogant assholes in world history) told a made-up story about ancient Egypt.

Theuth, the Egyptian god of writing, brings his invention to Thamus, king of Egypt. Here, says Theuth, is a discipline that will make your people wise and give them better memories. Thamus looks at it and says “you know what, we’re good.” This will produce forgetfulness. People will stop using their memory because they’ll rely on what’s written. They’ll absorb information without instruction. They’ll seem to know much while actually knowing nothing.

That speech is 2400 years old, and I’m willing to bet you’ve heard some version of it recently from someone you can’t just brush off.

Subscribe now

So we have to start by seeing that Thamus had a point.

Oral traditions before writing required extraordinary feats of memory. The Homeric bards, the griots, Torah memorization — those capacities are now preservation projects, not live traditions. We did offload memory to the page and something atrophied.

The appearance-of-wisdom problem is also real and also daily: “I reviewed the deck” means I scrolled it. Thamus was describing something that happens in every meeting, every morning, in every organization that runs on documents that get generated but not interrogated. He called the mechanism correctly, and it’s the comfortable default.

But he was thinking about the wrong unit: individual memory. What writing actually produced was something he had no concept for: collective memory that compounds across generations.

You don’t have to rediscover calculus. Leibniz is dead, but calculus lives. The Library of Alexandria is the answer to Thamus, just 50 years after Plato died. So is every scientific paper that builds on the one before it. So is the fact that we’re having this conversation 2400 years later. Thamus couldn’t see the compounding from the beginning. Nobody at the beginning ever can.

Socrates doesn’t refute Thamus, but he qualifies him. The written word is useful as hypomnemata: a prompt for genuine reasoning.1 One kind of scroll sparks dialogue and discovery, the other kind you scroll through on your way to feeling informed. The distinction is the posture you bring: are you reading, or are you questioning?

The autodidact who truly learns from reading is already arguing — bringing prior questions to the text, testing claims against experience, using the page as a thinking partner. Darwin had years of observation and a problem he couldn’t solve, then he happened to pick up Malthus and place it in the crucible. You and I have spent most or all of our lives just Googling it, but it’s the same mechanism – bringing the whole of our experience to bear with a bit of new information.

Let’s call search engines 30 years old. The skill was the query: craft the precise words to get the thing you already believed to exist to manifest at the top of the page. We’ve spent three decades with this model, and that training runs deep.

So when a reasoning machine shows up — something genuinely different, something that can hold your entire past and think across it — most people’s instinct is to use it just like the other computers. What did she say about renewing the contract? Can you find that part where he mentioned the timeline? Retrieval. The filing cabinet with a chat box on top. The faster horse, Thamus’s nightmare at speed, and you’re the one running it.

Socrates wasn’t delivering a philosophy as much as he was thinking in dialogue. Before writing made thinking a solitary act, this was taken for granted: you don’t compose a thought and then discuss it, you think by discussing. The question from the other person, the reframe you didn’t anticipate, the thing that changes your frame entirely — that’s the thought becoming itself.

Writing didn’t kill this, it institutionalized it. The Socratic method became the university seminar, which is why everyone reading this knows exactly what it feels like to have their thinking sharpened by someone smart pushing back. In a good seminar, in a rare mentorship, in a friendship where the other person doesn’t let you get away with anything short of your best.

The scarcity isn’t access to dialogue. It’s access to the right thinking partner at the right moment on the right problem. The seminar ended twenty years ago. Your best thinking friend has six kids. It’s 11pm and you’re stuck on a paragraph and nobody is available who knows what you’re working on and has no agenda about where it lands.

There’s a category of insight that requires holding more context simultaneously than a human brain can hold. Not more than you can eventually process — more than you can hold at once. That’s a new kind of impossible becoming possible.

Donald Knuth — the man who literally wrote the book on what computers can and cannot do — had been stuck on his own unsolved problem for weeks. A collaborator brought it to Claude with one discipline: document every dead end before moving on. Knuth’s response when he saw the result: “Shock! Shock!” Then he did what only Knuth could do — proved it was right and found 759 more solutions in the same neighborhood. The ceiling moved. And it moved because someone brought to the crucible everything the problem required.

Recently I had three advisor calls in the same week for the organizational memory product I’m building, who all told me the same thing. I was in all three conversations. I still missed it.

A technology exec who loves his AI tools told me my demo looked too similar to what he already has. A fractional CRO who already values his call transcription tool told me to expect to be compared to the competition. And a startup advisor told me that highlighting the “before and after” would be useful to bridge the comprehension gap.

While I pride myself on being a good listener, I can’t help but experience each conversation in a silo and draw conclusions that make sense in isolation but lack coherence with each other. So I might come out of those three calls taking away: iterate on UI/UX, do a competitive analysis, and write a case study.

But you can’t see the forest for the trees until you get the trees in the same room. Bringing all that context together for the AI – three fresh call transcripts plus my product docs and roadmap, marketing copy, and strategy briefs – gave me a powerful collaborator.

Showing up with a “what did I learn this week?” curiosity was all it took to shift the frame away from my to-do list and towards the underlying issue: I was spending too much time getting buy-in on how I was building and the overall architecture, not enough time on tangible output. Three pieces of feedback collapse into one slightly uncomfortable realization. The insight didn’t live in any one transcript, but in the space between them, invisible until I stopped asking what and started asking what patterns.

This essay was written the same way. Not by asking the reasoning machine to be a writing machine, but by asking it to think with me. Taking on my call transcripts so that I can continue the dialogues with my thought partners, patiently holding the argument while I tire myself out asking questions, clarifying the narrative, and finally organizing the outline. (Socrates would recognize this. 2400 years ago it didn’t need a name, it was just how thinking worked.)

And then once I had my AI-augmented outline, I opened a blank document and chose every word that I typed with my fingers on a keyboard. (Writing is the fun part.)

The written word compounds stored information and produced civilization. Reasoning machines compound synthesis — the cross-domain connection, the pattern across conversations, the inference you couldn’t reach alone. The compounding only accrues to the people who know how to question it, and nobody can see where that leads from here. Thamus couldn’t see it either, standing at the beginning of writing, protecting something local while something civilizational formed around him.

If Sama, the god of AI, approached Thamus and presented his invention, you’d expect him to reject it. This will produce laziness. People will stop thinking because this machine can do it for them. They’ll simply arrive at the same wrong conclusion, faster, with more people, and call it a breakthrough. They’ll seem to know much while actually knowing nothing.

Socrates didn’t just ask good questions. He walked around Athens, he talked to potters and generals and merchants and poets. The dialogues were synthesis on top of irreducible contact with actual human lives — problems he hadn’t invented, people who would push back in ways he couldn’t script, a city full of things the model can’t have experienced.2 The machine can reason across everything you’ve put into it, but it can’t go outside.

Start a dialogue not knowing what you’ll find. Then go have experiences it can’t have. Bring them back. Are you here for the crucible or the filing cabinet?

The conversation about AI displacement keeps asking the wrong question. “What should I do?” gets you “just learn to use AI” — copy these prompts, paste output in final report, professional value achieved. The work happens faster and none of it compounds. You produce more documents that take more effort to consume than they did to create, and you’re exactly as capable as you were last month.

The task is the wrong unit of analysis.

I spent years building technical skills — a dozen years working with and solving problems on a notoriously byzantine CRM platform. The value in being able to hold and retrieve that knowledge has gone down fast. Maybe it’s quaint that some London cabdrivers don’t use GPS, but for the rest of us, The Knowledge isn’t the gate it used to be.

But before that happened, I’d spent years deliberately building something else: the craft of architecting and leading a team. How to decide what to do, how to think about my team’s capability and capacity, how to evaluate work and provide feedback. When I picked up the AI tools, those are the skills that transferred.

So here’s the actual advice: learn to manage. Pick up the AI tools, and manage them.

Lead with why. Before you specify what you want done, share why it matters and what done looks like. Who’s the audience, how it will be used, how we’ve done it before and seen it go well. A team member (or an agent) who understands the purpose can make judgment calls that the rigid instruction-follower won’t see.

Don’t hide the ball. If you know the implementation path, say what you’re expecting — and make clear that better solutions are always welcome. If you don’t know how, say so explicitly. Ask for an approach (“how should I think about this?”), review the plan before any work starts, and steer the plan where needed. If you don’t steer the plan, you don’t understand the domain, or you’re not paying attention. The plan is cheap to fix, the work is frustrating and annoying to fix. Covering ignorance with vague instructions produces competent-sounding mediocre results, and the gap between what you got and what you wanted is almost always traceable back to something you left unspecified up front. The manager should always be ready to take the blame and roll the tape back.

Run the loop. Something came back; now work it. Observe what you got. Orient on the gap: where did your intent and their interpretation diverge? Was the why clear? Was the how too loose? Decide what to adjust. Act again. John Boyd figured out the OODA loop for dogfights — a fighter pilot who could reverse a losing position in under forty seconds, trying to explain why. The pilot who processes feedback faster wins. Management isn’t a one-time trip to the vending machine; every output is an input to the next move. The loop is the work.

Eyes on the seams. The fabric is strong. The stress is on the seams — the handoffs, the “take it from here and put it over there”s, the translations from your intent to their interpretation, the places where your spec met reality and something got missed entirely. That’s where you put your attention: the structure of the thing, how the components work, how to spot problems and recover quickly. It’s never going to be perfect so it had better be resilient. AI doesn’t know any more about you than you tell it, so be fanatical about the quality of your data.

You get what you give. People mirror each other. Trust compounds through attention, presence, and treating the collaboration like a collaboration. If you’re approaching this as a transaction you’re owed — tokens in, tokens out, why isn’t it better motherfucker — that’s what you’ll get. If you show up like you’re all smart people working on a problem together, the work reflects it. This isn’t a soft skill, it’s mechanical. The quality of your engagement is in the output, and your capacity for collaborative problem solving was always your value.

AI gives you a place to practice all of these skills at zero stakes. No feelings to hurt, no politics to play, no performance review looming. Give a vague spec and see what you get. Leave the why out and notice what happens. Failure is still the best teacher, it’s cheaper than ever, and the feedback is immediate. Reps are available on demand to anyone with a problem to solve and willingness to iterate.

Pick a problem, and start close to where you are: a spec where you don’t know the implementation but you know the challenge personally, a new idea that needs contact with more perspectives to get sharper, getting a process right and then getting it repeatable. That’s how a manager builds team capability, human or otherwise, and they’re concrete loops you can start running today.

Maybe you were never going to be a people manager. You had good reasons and plenty of experience with the management style of the largest hominid available: no aptitude, no craft, no training, just seniority and proximity to power. Not hard to see why that’s not for you, and if you were already a deep IC then the org chart version of management is an obvious trap.

But the skills underneath the title will serve you. Growth mindset was always the answer, with or without AI, with or without any particular technological or societal moment hanging over us all. Now the sandbox is nearly free and the iteration loop will teach you right up to your capacity to process the feedback.

Get your reps now. The problem is the point – the hardest one you can think of. Don’t stop at “what should I do?” when you can ask “where should I go?”

It was Barbenheimer weekend, a simpler time. The theater for Oppenheimer was packed, and I was seated next to a kind older gentleman with sensitive hearing – every time the movie reminded you what it was about, he protected his ears digitally, i.e., with his fingers.

And then the Trinity test, the first nuclear detonation in the world. Nolan holds the silence. The old man tentatively puts his hands down.

Was I supposed to warn him? He forgot, for that moment, that light travels faster than sound.

Growing up in south Florida, you learn that when you see the flash of lighting, you start counting – one-alligator, two-alligator, three-alligator. Five seconds per mile. And then another flash, and another count. This is how to locate yourself relative to the storm – is it getting closer, is it getting faster, do I need to get inside?

The flash wasn’t ChatGPT going viral in late 2022. That was a demo and a shrewd user grab. Plenty of people said at the time that the singularity was absolutely here and we were all aboard the hype train to the future. Whether or not they end up being right, they didn’t have the evidence at the time, they didn’t have the how. They saw AI-generated video, filter faces and image tricks, and GPT-written fully automated luxury B2B spam, and I saw clamoring for yesterday’s bullshit faster.

But I have seen the flash, and it came in late 2025 as we crossed the threshold into agentic AI. The people who noticed it first were the people who were banging their heads against the limitations of the previous versions, who were developing elaborate scaffolding and processes around the limitations of the AI. And then all that resistance simply dropped.

To put a finger on it, Anthropic’s release of Claude Opus 4.6 and their agentic tooling and OpenAI’s release of GPT 5.2 made new things possible, and possible in a way that you can see the path from here to widespread implementation and adoption. Not “viral sensation” mechanics – we’re at the point where we can watch along as the infinite software machine prints faster and faster.

And then literally as I write this, Anthropic acknowledged the existence of Claude Mythos, and saw no choice but to give only the software establishment access in order to mitigate the risks. I had already seen the flash, and counted about four months to this new profoundly weird feeling.

Mythos made the flash reach more people’s eyes, but there’s not a widespread agreement that the Trinity test moment has even arrived. Others think I’m years late, and I think they’re wrong.

Still plenty of people can’t update past “fancy autocomplete”. Fine, that’s a critique, accurately describing actual limitations. We may never know how many r’s are in strawberry, but we’ve also successfully harnessed whatever the hell this matrix math beast turns out to be. We have all that’s needed to upend the social contract: a machine that can reason through documentation as well as a motivated typical office worker. At scale, that’ll do it.

But these tools were scaled in ways that the men doing it knew were wrong, and they did it anyway because the alternative was losing the race. (You’ve seen this movie.) There’s completely justified grief over that, and the anger and denial that come early in that process. You have to manage your own information and if you decide to minimize this or compartmentalize it away – there’s more than one apocalypse in progress, as you know – there’s going to be bargaining and depression upcoming, and there will be bad actors counting on it. Acceptance as a practical choice. There will not be an un-flash.

Of course, the shockwave from AI is different from the shockwave of an atomic bomb. Light moves faster than sound, and sound moves faster than capitalism, and society at large moves slowest of all (but still way faster than before). The future is not evenly distributed, and that unevenness is exactly where we can retreat into every flavor of “not yet.” Even if model capabilities plateau today — no evidence of that — there’s at least a decade of untapped value in problems that can be solved with software and compute instead of humans performing the knowledge work they were trained for. The boom doesn’t need a bigger explosion, the existing momentum will carry.

It turns out that if you build enough transistors and shrink them down enough and get them close enough together and put enough electricity through them you can do so much matrix multiplication to the point that it gets really bonkers. You don’t get to choose what comes off the emergence stack, and whether it’s objectively crazy to speedrun this in about 70 years sadly misses the reality: no one knew until someone tried.

I won’t ask you to feel good about standing in the path of a storm. (I don’t know how I’d characterize how I feel, but it’s not “good.”) But you should be counting. It doesn’t tell you what to do, but it gives you position and bearing. What you do with that is yours, and it depends on where you happen to be standing. The oriented person acts from a model, everyone else is just in motion.

After nearly a minute of tension in the theater, the old man next to me heard the boom. A moment earlier, I gave him a nudge and a glance and we were both ready for it.