The model has no memory at all: the gap between you and everyone else is in the harness

The short version

harness literally means horse tack.

The AI model is the horse. We’re the ones riding it. The saddle, the stirrups, the crop, the reins in your hand—every piece serves the same purpose: so the human and the horse, the human and the agent, the human and the model, understand each other better and work together more smoothly.

The gap between one person and another using AI comes down, in large part, to how good that tack is.

First, an observation: shares are the most, hearts the fewest

There’s a pattern across all my videos: shares are far and away the most common, likes come second, and hearts are the rarest. Comments only pick up on a hot topic, or when I say something enough people want to argue with.

Which is interesting in itself.

My videos are long-form substance, and most people don’t have an uninterrupted block of time to watch to the end. So the natural reaction is: this looks valuable, let me share it, bookmark it for myself, come back to it later.

And my videos aren’t the kind that hand you a prompt. A lot of them are underlying logic and thinking, which you can’t understand and apply that fast. Plenty of the time you’d even feed my transcript or the attachments to an AI, discuss it, do extra research, before it really lands. So of course shares run high.

What’s the logic of a like? On short video, a like is “well said,” “I think so too,” “yes, exactly, that’s fun”—you double-tap the screen, it happens naturally, and there’s no pressure, because nobody else can see that you did it.

The heart is a different thing. A heart means I’m endorsing this content, and my friends will see me light it up. That takes real conviction from the viewer—this was good, I like listening, post more—plus, I want the people around me to see it. Which means second-wave distribution.

Who actually taps the heart

For AI substance, who tends to tap that heart?

Mostly entrepreneurs, founders and executives—people under relatively less competitive pressure at work—or people who are very strong AI learners. They don’t carry the anxiety of “if I heart this and others see it and learn it too, my edge shrinks.”

And then there are my longtime followers and the friends in my group. My clone in there runs on top-tier models answering their questions every day, so this is one way people express thanks, or support. It’s genuinely nice to see familiar names each time.

So I’ll sincerely ask every listener, every follower: when you think the content earned it, help me light up that little heart.

I don’t believe video platforms must follow a traffic-above-all law of the jungle. Users will gradually build the habit of filtering the genuinely good stuff up—because nobody wants a feed full of “the sky is falling” fake news, or shill ads.

So what is a harness

This afternoon I gave a public talk online. Another speaker there brought up a concept: harness.

I’ve covered it in other videos too. But what struck me is that beginners really can’t picture what a harness even is.

So, as a serious film buff, I reached for an analogy.

Harness literally means horse tack. The AI model is the horse; we’re the ones riding it. And what’s the point of tack? Whether it’s the saddle, the stirrups, the crop, or the reins in your hand—all of it exists so the human and the horse, the human and the agent, the human and the model, understand each other better and work together more smoothly.

It’s a lot like 50 First Dates

There’s a film called 50 First Dates. If you haven’t seen it, go watch it—a really fun romantic comedy.

The female lead wakes up every morning with no memory. So the male lead puts together a whole set of things from their history—diaries, photos, even video—and every morning when she wakes up, he plays it for her, so the two of them can fall right back in love.

And no matter whose model it is today—domestic, or the best from overseas—no matter how large the parameter count: the model has no memory. None at all.

Every single conversation is that woman waking up. Nothing remembered.

Why the first message in a new conversation is so slow

So why is it that every time you open ClaudeCode, or open Codex, on a new project or a new conversation, the first thing you say makes it think for so long and burn so much compute?

Because in the background, all those dense notes, everything that passed between you, has to be shown to the model again—it has to be told about all of that shared history, so it can work with you properly on the task in front of you.

What memory actually is

So what is memory, really?

It’s just that context in our heads that can be summoned quickly and selectively.

And the gap between one person and another using AI comes down, in large part, to how good the harness is.

That’s the one small concept for today. See you tomorrow.

Source: EP0075 original audio track · ASR model gemini-2.5-pro (parallel segments) · full text of the original video

[00:00] I’ve noticed a really interesting pattern. If you follow me, and you spend the time to watch this one through—there’s a characteristic across all my videos: shares are especially high, likes come second, and hearts are the fewest. Comments only really pick up when it’s a hot topic, or when I’ve put forward a view that a lot of people who love to argue want to push back on. Which is interesting in itself. My videos are long-form substance, and most people don’t have an uninterrupted block of time to watch to the end. So the most natural reaction is to decide this piece is pretty valuable.

[00:25] So I share it, bookmark it for myself, and come back to it later. And because my videos aren’t the kind that hand you a prompt—a lot of them are underlying logic and thinking—there’s no way to understand and apply them that fast. A lot of the time you’d even take my transcript or the attachments and give them to an AI, discuss it, do extra research, before you can really get all the way to the concept or the hands-on method. So shares running high is perfectly normal. Then what’s the logic of a like?

[00:51] On short video, the logic of a like is: well said, I think so too, yes exactly, that’s fun. These are all double-tap-the-screen behaviors. They happen naturally, and there’s no pressure, because other people can’t see you doing it. The heart is a different thing. A heart means I’m endorsing this content. My friends will see me light that heart up. That takes a very strong endorsement from the viewer of what I’ve made.

[01:16] As in: that was good, I like listening, post more. And I also want the people around me to see this good content. Which means it becomes second-wave distribution. And for AI substance, what kind of person tends to tap that heart? First, most likely entrepreneurs, founders or executives—the group under relatively less competitive pressure at work. Or people whose ability to learn AI is very strong. They don’t have that feeling of, if I tap the heart and other people see it and learn it too—

[01:41] —then my competitive edge shrinks. That anxiety. And then there’s another kind: my longtime followers, the friends in my group, who also light up the heart to help me spread it. After all, in the group my clone runs on top-tier models, answering their questions every day, so people use this as a way to express thanks to me, or support. It’s genuinely nice to see familiar names each time. So I’ll also sincerely invite every listener, every one of my followers: when you find the content valuable, help me light up that little heart. I don’t believe video platforms

[02:06] must follow a traffic-above-all law of the jungle. Users will gradually build the habit of filtering the genuinely good content up. Because nobody wants to scroll past “the sky is falling” fake news every day, or shill ads. Now, yesterday I mentioned that this afternoon I had a public talk online. That talk is over now. And if I were to pull one thing out of the whole talk to cover in this video, I think it’s the following. Because another speaker at that talk

[02:31] brought up a concept called harness. I’ve covered it in my other videos too. But what I realized is this: beginners find it very hard to understand what this harness even is. So, as a serious film buff, I thought of using an analogy like this to explain what the concept of harness actually is. Harness literally means horse tack. The AI model is the horse, and we’re the ones riding it. And what’s the starting point of tack? Whether it’s the saddle, the stirrups, the crop—

[02:56] —or the reins you hold in your hand, the whole set of them has one purpose: to let the human and the horse, the human and the agent, the human and the model, understand each other better and work together more smoothly. So what is it a lot like? There’s a film called 50 First Dates. If you haven’t seen it, go take a look—a really interesting romantic comedy. The female lead in it wakes up every morning with amnesia. So the male lead prepares a whole set of things from their history—diaries, photos, even video—and every morning when she wakes up, he plays it for her.

[03:21] So that the two of them can fall right back in love again. And no matter whose model it is today, domestic or the top-tier ones from overseas, no matter how big the parameter count—the model has no memory. None at all. Every single conversation is like this woman waking up: nothing remembered. So why is it that every time we open Claude Code, open Codex, a new project, a new conversation, and say the first sentence to it, it thinks for so long and burns so much compute? It’s because in the background all those dense notes, everything that passed between us,

[03:47] has to be shown to the model again, to let it know all of that history once more, so that it can work with us better on the task at hand. So what is the nature of memory? It’s just that context in our heads that can be rapidly and selectively summoned. And the gap between one person and another using AI comes down, in large part, to how good the harness is. So that’s the one simple little concept for today. See you tomorrow. Bye.