Fast AI requires slow thinking
The current Silicon Valley flex is trading notes on your favorite new AI tools over lunch. Each week brings another one to explore. Some of these tools are very impressive; I’ve been able to reply to someone with an alternative UI design in less than a minute, and they were dumbfounded: “How did you do that so fast?”
This speed is intoxicating, but let’s not get too carried away. On the outside these tools may be fun, but on the inside, there are plenty of reasons to be skeptical. Much ink has been spilled on the ethical, ecological, and economic concerns of AI. I’m not ignoring any of that. There is an odd particle-wave aspect to LLMs. You can be skeptical of how they are built and run, but still intrigued by the technology and what it can do.
I’m looking past today’s tools to what might be coming next. There are open-sourced, ethically trained, small language models aiming to fix the flaws of current LLMs. My interest in today’s foundational models isn’t to endorse them, but to deconstruct them. I want to understand their limits, their actual strengths, and most importantly, how they affect us. This technology feels transformational, but its sheer velocity is dangerous. It is moving so fast it may be breaking how we think.
Spinning Plates and Rabbit Holes
The challenge is that these AI tools can be so thrilling as to become addictive. This is explored in the AI Genie Phenomenon where the authors describe the “Epistemic Rabbit Hole,” a compulsive, vertical cycle where the speedy nature of chatbots makes users lose themselves in the pursuit of “more.” You believe you are being highly productive, but you are actually just digging a deeper, narrower hole.
Then there is the horizontal dysfunction. I know people who keep their laptops by their beds just to “poke” an agentic system in the middle of the night to keep it moving along. Juggling so many separate threads, flitting from agent to agent just to keep the momentum going, is turning us into AI plate spinners.
The chatbot interface actively drives both of these cognitive dysfunctions: the deep, compulsive rabbit hole and the frantic, scattered plate spinning. Both trigger a rapid dopamine response that feels productive. I’m certainly not saying this approach is useless; it’s clear many people are accomplishing large amounts of work with these tools. But it’s like driving a sports car without a seatbelt. It’s a way of working that is highly vulnerable to a crash. I want to understand that risk.
Set Shifting
It might help to back up a bit and discuss what could be happening here. Daniel Kahneman’s book “Thinking, Fast and Slow” popularizes Dual-Process Theory, which explains how human cognition relies on two very different modes of operation:
- System 1
This is our fast, automatic, intuitive, and error-prone system. It is the part of your brain that reads a giant billboard on the highway, recognizes a friend’s face in a crowd, or instinctively catches a falling object. System 1 is a critical shortcut to everyday living, as you can’t fully reflect and deliberate every moment of your day. - System 2
This is our slower, more deliberate, and analytical system. It is the part of our brain we are forced to use when trying to parallel park in a tight space, solve a math equation, or debug a complex piece of code. Most of our cognitive heavy lifting is done using System 2, but it’s slow and requires self-discipline.
The core problem with spinning plates and rabbit holes is not that we completely abandon System 2; it’s that the chatbot interface tends to starve it of time. It generates System 2-level complexity, so we feel like deep cognitive work is happening. But the interface is fundamentally reactive. It spits complex text at you, you skim it quickly, and you immediately type a reaction to keep the momentum going.
We are trying to do System 2 synthesis at System 1 speeds. This compressed loop prevents the slow, deliberate analysis that complex reasoning requires. Of course, if you are very disciplined you can ask the AI to back up and reflect, but it takes self-awareness and willpower.
My hypothesis is that the very structure of the chatbot interface (type, read, type again) actively discourages reflection. When you are moving too fast, you get stuck in a groove. You literally need to take a break, step back, and basically step out of this groove so you can view the problem from a new angle. We’ve all walked away from a tough problem only to have the solution arrive unbidden into our thoughts later in the day. Psychologists call this crucial mental reset “set-shifting.”
Ma ex Machina
I’ve recently learned of the Japanese concept called Ma. It’s an aspect of their traditional art and culture to create a “negative space” or a deliberate pause. This helps frame a previous action and accentuates its meaning. The Studio Ghibli director, Miyazaki talked about it in a 2002 interview with Roger Ebert. Miyazaki clapped his hands several times and said:
“The time in between my clapping is Ma. If you just have non-stop action with no breathing space at all, it’s just busyness, But if you take a moment, then the tension building in the film can grow into a wider dimension. If you just have constant tension at 80 degrees all the time you just get numb.”
Hayao Miyazaki ‘Ma’ is an Essential Storytelling Tool
In movie making this is called a “pillow shot.” This could be a quiet moment of leaves falling, a distant train passing, or a character simply looking out a window. These types of negative spaces allow the audience to digest the emotional impact of the previous scene. It is where the story actually settles in the viewer’s mind.
Ma provides a framework for understanding that a pause is not a lack of work; it is the space where System 2 synthesis can emerge. It reminds us that our intellect is nuanced and we just need time to ‘soak’ in a problem for a bit to grasp it properly. Ma is a metaphor for how we need to structure our attention. We should move away from a simple chatbot-focused “Fast AI” and embrace a more reflective style. We need to design our UX to avoid the rabbit hole, slow down, and see things more clearly.
The Real Problem
Of course, if we just slapped a thirty-second loading bar onto a chat interface, users would riot. Slowing down feels physically uncomfortable. Besides, this ignores a critical issue: sometimes I want to go down a rabbit hole. For all of my concerns, it is possible to get work done this way, it’s just risky. The balance between System 1 and System 2 is a personal issue, and pushing a “hey, you wanna take a break?” would just feel like a modern version of Microsoft Clippy.
We need interfaces that naturally invite a pause without forcing it. There are a range of solutions that could be tried. I’m reluctant to offer anything definitive, as they are all likely to be either too simplistic or too complicated. The magic is in finding the right balance.
It’s tempting to tackle the immediate problem with smaller fixes:
- Make the chatbot more reflective, pausing occasionally to push the user to review.
- Change the type of feedback, replying with diagrams instead of text to encourage a different processing of the information.
- Use feedback mechanisms outside of the chat window that would give users a higher level roadmap of the entire conversation.
All of these would be helpful, but they ultimately miss the bigger issue: the chatbot experience is the root problem.
The Road to Ma
The much bigger goal is to get rid of the chatbot entirely. Create an interactive experience that isn’t quite so much like a slot machine. Napkin.ai is taking a promising first step in this direction. This is a diagram creation tool, useful for creating slide presentation images. When you ask it for a diagram, it does use a text prompt. But instead of that back-and-forth chatting, it just spits back an overview of what it thinks you asked. It’s far more detailed than what you asked for. That detail makes you stop and review, “Did I really ask for all of this?” This is a bit like the “reflective prompting” mentioned above.
After you optionally edit this overview, it creates a diagram. But instead of more chatting, you can actually manipulate the diagram directly. You can do simple things like edit text and change colors. More importantly, you can create AI-assisted variations on sub-pieces. It is far calmer and the user feels in control.
This approach implies that making a truly reflective product requires getting closer to the final output. Chatbots are common because they’re generic. Slapping one into a product doesn’t require any serious adjustment for the task. All the responsibility and clarification falls on the user’s shoulders. The real breakthroughs will come when the tool molds itself to the task. That’s clearly much harder to do.
There is obviously much more to explore. For me, the biggest insight in putting this post together was that most people don’t appreciate the impact chatbots have on our cognitive processes. Once you see that, you can appreciate the power of Ma and start looking for ways to inject it into our tools.
When I read about Ma, it was like a clarifying bolt of lightning. We’ve been looking at this problem entirely from the technology’s point of view. The LLM is driving us, pushing a specific interaction model and veering us towards making mistakes. We are not computers; going fast isn’t our superpower. Ma is what makes us human; it’s what allows us to think like humans, not machines.
“In an age of speed, I began to think, nothing could be more invigorating than going slow. In an age of distraction, nothing can feel more luxurious than paying attention. And in an age of constant movement, nothing is more urgent than sitting still.”
Pico Iyer, The Art of Stillness: Adventures in Going Nowhere