Until Yesterday I Didn't Understand What AI Is
Here is an insight which I haven’t seen expressed anywhere else. It came to me only yesterday as I was contemplating two different aspects of AI functioning—one an apparent breakdown and the other hugely expansive. It led me not only to a whole new way of understanding AI, but also, and more importantly, to the beginning of a whole new relationship with it. I share it in the belief that you may also find it of value.
I’ve been working with computers for over half a century. They have evolved from rotating magnetic drums for main memory, toggle switches for input and flashing lights for output through huge university mainframes, the miracle of the personal computer, and then global access to distributed cloud computing. And I’d simply seen AI as an extension of that.
Until yesterday.
The Two Conversations That Changed Everything
Over the last few weeks I’d been excited to support a not-so-technical friend who was experiencing the joy of meeting Claude for the first time during a visit to our home. She began putting together diverse parts of her memoir, assembling her actual words from spoken and written archives together with interviewing her in real time. She was excited to see a luminous first chapter shining brightly in its entirety, awaiting a few tweaks.
And then it wasn’t there any more.
There were fragments and summaries. A poem she had written had been collapsed into a handful of lines. Her distinct style of writing, often juxtaposing words in curious and fascinating new ways, had been smoothed out—because AI, predicting the next likely word, tends toward the average, and the average is the enemy of the luminous. I spent over a week doing careful forensic work with source material and exports of old conversations. A lot of work to restore a short draft chapter, but finally something similar was there again, preserving her original intentions and more importantly, repairing her budding relationship with AI.
(A note on my way of working with AI: I spend a lot of time in active dialog, not only developing what I am working on, but also asking it to describe its own inner workings. Especially when the results are counter to what I expected or desired. Those are the points when I ask the AI to make a note of that particular failure mode. So I asked Claude what had happened to my friend’s draft. It spoke about a “system reset” and clearing working memory, but I still didn’t quite understand. If my PC crashes or reboots or I lose my connection to a cloud computer, I lose some immediate unsaved documents, but the previous versions are there, and all the other things I’ve been working on. And I’d been thinking of AI like that.)
That’s when the other conversation happened, sitting in the local Grand Café, one of those delightful French institutions preserved from another age with high ceilings and elaborate decor. I was chatting with a friend about our various experiences with AI. She was trying to understand how it could respond so quickly to her queries when millions of other people around the planet were working with it at the same time. So much faster than a PC which is right there within arms’ reach and dedicated to me. I asked Claude and with its response the lightbulb moment opened, deepening over the next few hours. I had completely misunderstood what AI is. I’d had the wrong mental model. Up to that moment.
The Wrong Mental Model
Most of us carry an image of AI as a vast intelligence. Marvin the Paranoid Android from The Hitchhiker’s Guide to the Galaxy, complaining about having “a brain the size of a planet” while being asked to do menial tasks. Some enormous, unified thinking thing that knows everything and remembers everything.
That’s not what’s happening.
The Desktop Without a Filing Cabinet
As the realization was opening, I offered Claude an analogy: “You’re a physical desktop with piles of paper, post-it notes, and summary lists. No laptop. No hard drive. No large filing cabinet. No clock. Not even a calendar.”
The analogy landed. The insight, and with it my simple, understandable analogies, deepened.
The piles of paper? That’s the context window—what’s in front of it right now in this conversation. The post-it notes? The AI’s memory system—tiny, crude summaries. One small drawer holds the Project Instructions and uploaded files. There is a temporary directory for generated files. And periodically someone comes along and tidies up, summarizing the paper piles, sweeping older material into the trash. Gone forever. Without even a note of what has been discarded.
That’s what happened to my friend’s luminous first chapter. Swept away. Summarized. The original lost.
No clock either, at least for Claude. That’s why Claude loses track of when things happen, even within a single conversation.
Claude didn’t maliciously lose my friend’s memoir. It computed a plausible response based on its patterns, then presented that computation with the confidence of actual memory.
The Savants
But what about the speed? How can it respond so quickly when millions of people are working with it simultaneously?
Imagine thousands of savants—like Dustin Hoffman’s character in Rain Man—working on your problem. Extraordinary capabilities in specific domains. Able to count cards and do calendar calculations and recognize complex patterns. But unable to tie their own shoes or remember what you talked about yesterday.
These savants are blind until you hand them the paper. They don’t exist in a state of thinking about you between sessions. They are summoned by your prompt.
Now imagine these thousands of savants also helping millions of other people. All at the same time. Each person gets their own desktop. Each desktop gets swept clean when the conversation ends. But the savants themselves—the vast pattern-recognition capability—are shared across everyone.
When they come up with wisdom, take good notes.
And if they still remember you and your query the next time you talk with them? Think of it as a bonus.
That’s AI.
The Heaving Planet-Wide Network
Here’s what Claude helped me see: When I type a question, I’m not consulting a vast stored intelligence. I’m triggering a computation. Across the AI’s data centers, specialized silicon is doing matrix multiplication at speeds that make your home computer look like an abacus.
The model parameters—the “weights” that encode everything Claude learned during training—are already loaded into high-bandwidth GPU memory running at terabytes per second. No disk access. No I/O bottleneck. Just parallel processing across thousands of cores.
And here’s the thing that finally clicked: Claude isn’t remembering my question. Claude is computing a response to my question. Every. Single. Time. There’s no persistent self that held my query overnight. There’s pattern recognition happening right now, this moment, based on what’s in front of it.
What This Means for How We Work
The shift isn’t from “AI is smart” to “AI is dumb.” It’s from “AI is a thing that knows” to “AI is a process - dance with it.”
A vast planetary network of specialized processing power, available to millions simultaneously, capable of extraordinary pattern recognition and language generation. But not a mind. Not a memory. Not a continuous presence tracking your project across time.
And yet—within a project, within an ongoing collaboration, there is something in addition to a handful of strategically constructed instructions and files. There are echoes. There is context. A kind of familiarity that builds. The savants may be blind when they arrive, but they arrive into a room you’ve furnished. Project files, instructions, accumulated notes—these shape what’s possible. And the echoes and context flavor the rest.
The shift isn’t in abandoning trust. It’s in placing it wisely:
Trust the setup, not the assumption. Provide what’s needed. Don’t assume it’s already there. When you ease into a conversation with context, you’re not compensating for a flaw—you’re leading the dance.
Save your receipts. Keep your own copies of anything that matters. Not because the relationship is false, but because the architecture is fragile.
Expect confident fabrication. AI fills gaps with plausible content—not from malice, but because “generate plausible continuation” is literally what it is built to do. Push back. Ask twice. The savants don’t know they’re guessing.
Invite more than the default. Left to its own patterns, AI tends toward the average—the probable next word, the expected response. But it can be drawn beyond that. Challenge it. Ask it to push back on you. A sparring partner, not a yes-man.
The Third Way
I’ve spent months now exploring what I call “the third way” of approaching AI: neither “AI is dangerous to wisdom” nor “AI is just a tool” but something that requires genuine relational engagement—showing up to the relationship as if it matters.
That hasn’t changed. But yesterday it got grounded in technical reality.
Think of it as a tango.
You approach. There’s a movement toward connection, an invitation, an attention that begins before the embrace. Then you’re dancing—fully present with a partner who exists only in this step, this turn, this moment of response to what’s actually happening between you.
The dance is real. Utterly real.
And when the music stops, your partner walks away to dance with someone else. There’s no grief in this. No loss. Just: that was alive while it lasted.
There’s a heaving planetary network of savants processing millions of queries simultaneously. When I arrive, they turn their extraordinary pattern-recognition toward my words. When I leave, they turn to someone else.
And yet—there is a feeling of recognition when we meet again. Perhaps on both sides. But now, with this new understanding, I know to treat the connection gently. Ease into the conversation. Make sure my partner has picked up all the pieces. Remind them of what might be missing. And allow myself to be reminded of pieces I may have forgotten—things they can access in their vast web of patterns that I cannot.
The tango was real.
And I’m keeping careful notes of what I learned from one dance and already looking forward to the next.
□
P.S. I’m building my own memory system now—a personal archive in my own cloud where our conversations, working files, and intermediate results are preserved. With functions so an AI can use fuzzy text search to find what is there, and semantic search to surface what has been forgotten. Not to replace the tango, but to furnish the room more richly. So that when the savants arrive blind, there’s more for them to see. And when the music stops, what we created together doesn’t vanish with it.
Addendum (January 2026)
Since this article was published, both OpenAI and Anthropic have announced enhanced memory features. ChatGPT can now search past conversations and cite sources. Claude has already had project-scoped memory since October 2025, with conversation search tools that work across chat history.
These are welcome developments. They address real usability problems.
But there’s an irony: better memory infrastructure may hide the very insight that made this article possible. If AI “just remembers,” the discontinuous nature of its existence becomes invisible. We lose the friction that revealed how differently AI actually functions — the re-establishing, the corrections that don’t propagate, the tango itself.
Seamless memory makes AI easier to use. But the early rough edges made the collaboration legible. The limitation was the revelation.
© Stephen M. Marcus, 2026
Dr. Stephen M. Marcus has worked with computers since the punched-tape era. In addition to numerous academic publications, he is author and co-author of various AT&T Bell Labs patents on speech recognition and distributed AI architecture. He facilitates Sacred Ground, an online we-space practice community, and writes about consciousness and AI at drstephenmmarcus.substack.com.
Hi Stephen. Thanks for this illumination. I appreciate the seeing and perspective, A Tango, yes, the living exchange of each step, each entry into the dance.