You’ve probably used AI for quite a lot by now. Writing, research, code, images, data. Things that used to take a day now take thirty minutes. That’s real.
And the capabilities keep climbing. Multiple AI agents can collaborate on complex task chains. Vibe coding lets people with no programming background build full products in natural language. The degree of automation refreshes almost every month.
But if you look carefully, almost every AI product is answering the same question:
“You know what you want to do? Great. I’ll do it for you.”
The goal is yours. AI handles the execution. What to write, what to search, what to build, what to analyze — you set the direction, AI delivers. This is genuinely valuable. But it rests on an assumption: that you already know what you want.
What if you don’t?
Vibe coding already proved something
When Andrej Karpathy introduced “vibe coding” in early 2025, many people saw it as a new way to speed up programming. But what actually happened was much bigger. In Y Combinator’s Winter 2025 batch, a quarter of startups had codebases that were 95% AI-generated. A content creator built a working health-tracking mini-app in two hours. A designer built an AI short-film production pipeline — one sentence in, full video out. A sushi lover built a sushi rating app simply because he loved sushi. The “one-person company” became real: one person plus AI tools, doing what used to require an entire team.
Some had technical backgrounds. Some didn’t. But what truly mattered was that none of them would have built these things before. Making apps, shipping products, building full-stack software — these used to belong to “developers.” AI opened that door. What vibe coding really changed wasn’t just the speed of code production. It let a group of people who never thought of themselves as builders start turning the ideas in their heads into real, working things.
So its significance goes beyond acceleration. Acceleration means you already know how to do something, and AI helps you do it faster. Activation means you didn’t think you could do it at all, but AI lit that potential up. The former improves efficiency. The latter opens possibility.
That’s why vibe coding matters. It completed an activation on the doing side. It proved that when AI lowers the barrier to building, huge numbers of people discover they were capable of creating all along. People who’d been kept outside by technical barriers could now step directly into the act of making.
But the issue is also here: activation has mainly happened on the doing side.
The other door hasn’t opened yet
Vibe coding solved “I have an idea but can’t build it.” But there’s an earlier question that remains unsolved: What should I be working on? I have a vague feeling about something, but I can’t articulate it. I sense there’s a better direction, but I don’t know how to get there. I know the current approach isn’t right, but I don’t know what right looks like.
These aren’t simply about getting something made. They’re about whether someone can see a problem clearly, identify a direction, and know what’s worth pursuing further. Often, what people lack isn’t the ability to execute. It’s that they haven’t yet seen where the real problem is. Not that they’re missing tools, but that they don’t know where to point them. Not that they have no ideas, but that the ideas are still stuck as vague feelings — unexpanded, unexamined, unpushed.
AI’s technical capabilities are actually close to being able to help with this. Longer context windows, persistent memory, multi-step reasoning, cross-modal generation — if applied to a real thinking process, these could plausibly help people see problems earlier, distinguish direction faster, and identify more clearly what’s worth continuing. But most AI products still put their energy into “helping you do.” You enter a goal, it gives you a result. You set a task, it completes a task. This model is effective and easy to measure, so it naturally dominates.
It just hasn’t touched a deeper question: can AI participate in the process of someone thinking a problem through, not just in their execution? If vibe coding already opened the door to “getting things made,” then the door to direction, questions, and possibility itself hasn’t really opened yet.
· · ·
The step before the answer
Here’s a very ordinary example. I have a stable communication habit: when someone describes a problem to me, I shift into solution mode almost immediately. I can understand the issue, quickly form a judgment about a reasonable direction, and start giving advice. This isn’t always wrong — it’s often effective. But I started noticing that many conversations weren’t actually moving forward. The other person would agree on the surface, then do nothing. The issue usually wasn’t my solution. It was that I jumped to “here’s what I think you should do” without first stopping to understand where they were actually stuck.
I’m not unaware of this pattern. The real problem is the distance between knowing and doing. In the middle of a real conversation, it’s easy to follow your own momentum. You only realize afterward that you skipped understanding and went straight to solving. So I started bringing AI into this. Not to tell me “how to communicate better.” Not to give me a script. Instead, I asked it to help me see one layer deeper when I’m about to discuss a plan with someone: what might this person be worried about? Do they need a solution right now, or do they need to feel understood first? Where might I be moving too fast?
This isn’t a one-time task anymore. It’s more like a new reminder mechanism woven into daily interactions. Each time I’m about to give advice, AI pulls my attention forward slightly — to see the other person’s situation before I enter my own judgment. That’s what I mean by “helping you think.” Not AI thinking for you. Not AI giving you a smarter answer. It’s AI slowing you down at the exact point where you’d normally glide past, letting you see from a different angle.
The same thing happens in my work. When I’m working through strategy or direction, I usually don’t have it all figured out before involving AI. More often, I have a confusion that hasn’t fully formed — something that feels important but can’t quite be articulated. At that point, AI’s most valuable contribution isn’t answering. It’s helping me push the question itself into shape.
For instance, I’d been sensing for a while that our current understanding of “product” is starting to loosen. Traditional products are typically designed, developed, shipped, then used. But with Gen UI, interfaces can be generated in real time, interactions can shift with context, and products no longer have to be fixed artifacts. Yet framing it as simply “traditional product” versus “generative interface” felt like two extremes. Something was missing in the middle. I couldn’t tell what it was at first, so I brought this vague sense into AI and worked through it iteratively.
Round after round we asked: What if a product isn’t designed once and shipped, but continuously adjusts through use? If a system keeps changing based on behavior, preferences, and context, is it still software in the traditional sense? If software starts behaving like a living thing — used, fed back into, reshaped — how should we describe it? This process led me to a concept I call “organic software.” AI didn’t hand me the term. It took shape gradually through many rounds of collision. AI offered connections, counter-questions, new angles. I judged which had merit and which didn’t, then threw new questions back. The concept clarified inside that back-and-forth.
Another time, during a signal research session, AI connected “storytelling” to “storyliving.” I hadn’t asked for this connection and wouldn’t have made it on my own. But it stopped me, because it touched a question I’d encountered repeatedly: in the AI era, is a brand still just telling stories? Or is it becoming something more like a continuously running system that people enter, participate in, and inhabit for a while? This question didn’t come from nowhere. I’d discussed it with clients, with peers, and just two weeks ago with a friend building a startup. The connection AI surfaced was like a temporary circuit joining scattered experiences, conversations, and instincts that had been floating separately. I continued working with it, and it gradually became a body of thinking I call Living Brands.
These experiences made me realize that AI’s value doesn’t always show up in what it helps me finish. Often, its real value is in what it helps me see. Seeing concerns the other person hasn’t voiced. Seeing the inertia in my own thinking. Seeing potential relationships between concepts that had been sitting apart. Seeing where a question can still go deeper.
This is quite different from what most AI products emphasize. They generally assume you already know your goal and just need to reach it faster. But in much of real work and real life, the hardest part isn’t finishing. It’s knowing what the problem actually is, what’s worth pursuing, what direction holds up. Not writing the answer, but turning a vague feeling into something you can discuss, weigh, and push further.
So “helping you think” doesn’t mean AI replaces your judgment. The opposite. It requires you to participate more deeply. AI spreads out possibilities, illuminates blind spots, hands you connections you hadn’t noticed. But which ones to keep, which to discard, which to push further — that’s still yours to decide.
Good AI isn’t just an execution tool. It’s more like a thinking environment that keeps bringing you back to the question itself. It may not give you an immediate answer, but it makes your questions more specific, your sense of direction clearer, and capabilities that used to be unstable gradually more usable.
From steering to mutual pull
Right now, most conversations about AI use the language of control. From prompt engineering in 2023, to context engineering, to the more recent discussions around harness engineering — the terminology has shifted several times, but the underlying logic stays close: how to better control AI, constrain it, make it produce stable output aligned with your intent. This logic matters. If AI can’t be used reliably, it can’t enter real workflows. But if every discussion stops at “how to make AI understand me, follow me, execute for me,” the relationship between humans and AI gets locked into a very narrow position: you give orders, AI is a dispatched capability.
Looking back, the relationship has moved through stages. First, the tool stage. I use AI like a hammer. It’s passive. It does exactly what I make it do. Then, the steering stage. I direct AI like riding a horse. It has its own power, but I set the direction. Better prompts, better context, better system design — all aimed at making the horse run steadier, faster, with less loss of control. This is where much of the engineering discussion focuses today.
But a more interesting relationship may lie in the next step: AI doesn’t just follow your direction, and doesn’t just help you think through problems. It forms a relationship of mutual pull. The human’s value is in intention, experience, trade-offs, and understanding of real situations. AI’s value is in connection, expansion, simulation, and feedback. You bring a vague feeling in, AI helps unfold it. AI surfaces new possibilities, you judge which hold up and which don’t. Your feedback reshapes the question itself, and AI adjusts within the new question. The real value happens inside this mutual pull.
Good AI isn’t a more obedient tool. It’s a capability that can pull you forward while you pull it. It doesn’t just land your intentions faster — it can also influence how you form intentions, how you ask questions, how you see direction.
In the experiences I described, the most important thing wasn’t what AI gave me. It was what happened after. AI offered a connection. I followed it and kept thinking. That thinking raised new questions. The new questions went back to AI. AI responded with new angles in the updated context. I judged, filtered, went deeper. This isn’t a one-shot Q&A. It’s an ongoing spiral. After each turn, my questions sharpen, my direction clarifies, and AI accumulates more signal about how I think, how I filter, how I form direction.
This is also why I’m reluctant to think of AI as purely a system that delivers results. That logic runs: I state a need, it completes a task. The logic of mutual pull is different: I bring in a question that hasn’t fully formed. AI gives me a new connection. I judge whether it holds, then decide whether to keep it, discard it, or push it further. My feedback changes the question itself, and AI adjusts within the new question. What gets pushed forward isn’t just the task. It’s my understanding of the problem.
In this sense, good AI doesn’t give you one great answer. It drives you to keep iterating. It gives you something you didn’t expect. That makes you think of something new. Your new thinking lets AI respond with something more precise. Each cycle, the thinking moves forward. Steering produces acceleration. Mutual pull produces activation.
But most AI products today are still in the steering stage. Interactions are one-shot: you ask, it answers, done. The next conversation starts from zero. Even memory features usually just remember what you said — rarely do they understand how you form direction, how you filter possibilities, how your questions are changing. Mutual pull requires more than memory. It requires understanding the trajectory of someone’s thinking. AI needs to sense where your weight is shifting, what your questions are approaching, how your criteria are changing — and then offer a response you didn’t explicitly ask for, but that happens to push you one step further.
This kind of capability hasn’t been seriously productized yet. But if it were — if every session made the next one deeper, sharper, more directional — what it would create isn’t convenience. It’s a co-evolution that’s very hard to replace. Of course, this relationship also requires the person to participate. AI can’t decide what matters on someone’s behalf, and it can’t make all their trade-offs. What it can do is, through round after round of interaction, gradually bring out your ability to catch the next step.
· · ·
What AI changes isn’t just your efficiency
AI’s execution abilities are already strong and will keep getting stronger. But doing things faster isn’t the same as seeing things more clearly.
People building AI products might want to consider: one-click completion is the pinnacle of acceleration, but it might also be a dead end for activation. If AI does everything for you, there’s no room left for thinking, exploring, or discovering. You get the result, but you don’t necessarily gain capability.
AI isn’t just for getting things done. It might also help you discover something you didn’t know was in you.
Writing this, I found myself wondering: when was the last time AI took me somewhere I genuinely didn’t expect? If you try to recall that for yourself, the answer might tell you something.
AI has helped you do a lot.
But has it changed the way you see things?
About the author
Qinqin (Stella) Yang is a director, builder, and imaginist working across AI capability and creative practice. She builds workflows, systems, and creative direction that operate under real constraints, and designs speculative frameworks that expand what teams can imagine before they start making. Over the past decade she has built brand experiences, product design, and speculative projects for Tencent, IKEA SPACE10, Burberry, and the Palace Museum. She is the founder of Actual Fiction, where she develops AI-native tools that turn imagination into a working method.
This newsletter is where she thinks out loud about creativity, AI, and what it means to build toward a future that doesn’t exist yet.
More at qinqinyang.com · Practice at Actual Fiction