I’ve been thinking even more about thinking.
Remember the framework I introduced in The Architecture and the Energy?
When you get right down to it, intelligence seems to boil down to just two components.
Having a well-designed system isn’t enough alone—you also need energy to power any thinking system. Our brains make up about a fiftieth of our total weight, but they burn up about a fifth of the food we eat.
Within that architecture—the thinking system—there are different types of thinking, each with different uses. I’ve tried to classify them today into three categories, and I’ve tried to think about all three types of thinking while I was in the middle of experiencing each one, and boy do I have some notes for you today.
Fast thinking is often differentiated from slow thinking by invoking the idea of a tiger about to pounce on you in the wild. Daniel Kahneman’s Thinking Fast and Slow does a great job of pointing out how we humans have evolved the way we think in order to better survive in the world we’ve lived in for most of our existence, and how evolution usually operates on a scale of millennia, while we humans have remade the world around us in a few short centuries.
In a different vein, you can think very, very deeply about one single thing. You can remove distractions and really zero in, say, on the way they used the word secular in Rome two thousand years ago. This etymological dive is about as deep as it gets (at least for me!), and it can be very satisfying to understand a great deal about one particular subject.
However, if you want to know how all that fits in with the way the world works, you need to zoom out and study Ancient Rome in its totality. This is the third vein of thinking, where you go wide instead of going fast or deep. The idea here is to see the big picture through polymathic thinking, and to do that, you need to gather an awful lot of information.
You can be fast, deep, or wide with your thinking, but it’s very costly to try to do more than one at a time.
I think there’s a lot of value in understanding this for your own mind, but also for anyone experimenting with AI tools today, this trade-off seems to be a fundamentally important thing to understand.
AI models need to burn energy to do what they do, and they also need to pick one of these strategies to employ. You might have a model that goes out and looks at dozens of different sources to find answers about a particular subject. These are now commonly called deep research models, although I’ll let
jump in with any details he might want to share. We might also get a bonus dad joke in the comments today.Clearly, the idea is for them to cook for a bit on one subject, then return with a gargantuan summary. Seriously, right now these things are like 50 pages long, like having a team of college freshmen doing research for you for hours total (though in AI time, it only takes a few minutes).
Right now, the same model can’t also think widely. There are models that address this specifically, though, and they’ve begun to be called reasoning models. Reasoning goes wide, and will contemplate variables that might not be directly in front of it. A model will need to sit and think about things so that it notices them, though—just like us.
Finally, there’s fast thinking. Sometimes you just want to know where most eels go to make whoopee, or how old James Clerk Maxwell was when he died, or what a metronome is, you definitely don’t want deep or wide thinking. Having the mental eye of Sauron turned toward something trivial is such a vast waste of resources as to be silly, but it’s also going to take a long time.
When you want this kind of thinking, you’re not here to wait. You want the correct suggested spelling of the word you screwed up right now, not after you eat a sandwich.
Fast, deep, or wide: do you really have to choose?
You do unless you want to burn a ton of extra energy. That’s the real cost of using an AI model that’s complete overkill for any task you want to do, and it’s the same tangible cost if you want to try to think fast, deep, and wide all at once. You might be able to combine those sorts of thinking in a flash and for a moment, but you’re going to be mentally exhausted from all that work.
I’ve come to envision burning trees whenever I use AI. This reminds me not to be wasteful with scarce resources, even if I’m not the one paying for them in any given instance.
I’ve also come to think the same way about my own brain, at least in the following way. I know I have a limited amount of bandwidth on any given day, and if I push to the edge of that limit, I tend to pay the price the next day. In other words, there’s only so much thinking you can do in a day.
If you work very hard on something, you probably need a mental break. That’s your brain using way, way more energy (for its size) than the rest of your body. This energy use is legitimately costly. I’ve noticed I need to eat more when I think more. I’ve also noticed I get really, really tired if I think a lot.
These parallels aren’t perfect, but I hope it’s a useful way for you to think about your own thinking today. Please give me a fast, wide, or deep comment if you have a moment!
It's all about the macro-micro-macro-micro zoom. Out for context, in for details, back out for alignment, back in for refinement.
Interesting parallels, although now that you've asked for clarity:
"Deep research" refers to conceptual features used in many AI products rather than separate models. In fact, OpenAI's "deep research" tool is powered by the o3 model under the hood. In the case of Google, the "deep research" model runs on Gemini 2.5. Flash if you're a free user or on Gemini 2.5 Pro if you have a paid Google account.
But I still think your analogy might sort of hold. In the case of OpenAI, you could roughly split it into:
Thinking fast = GPT-4o (quick, low-latency chat model that responds with the first surface tokens it pulls based on its training data.)
"Deep Research," ironically enough, slots better into the "wide" thinking mode, as it typically goes broad to pursue many dozens of sources (or even hundreds, in the case of Google).
But because "deep research" also uses reasoning models under the hood, it can go "deep" into the topic after pulling the sources together.
So it's harder to draw a clear line between reasoning models and deep research, since they work in tandem.
As for dad jokes, I don't think I could do better than "fast, deep, and wide."
It's already too on the penis. Uh, too "on the nose" I meant.