6. Fast & Slow Thinking

6. Fast & Slow Thinking

Fast and Slow Systems

The idea that there are two modes of thinking, fast and slow, was popularized by the breakout book Thinking, Fast and Slow by Daniel Kahneman, which spawned a whole discipline of behavioral economics, and won him the Novel Prize. The general idea, greatly simplified is:

• Fast thinking — automatic and intuitive processing.
• Slow thinking — rational, logical analysis

These two systems also have other names:

• Fast — System 1
• Slow — System 2

And there’s a distinction in terms of how we control the two systems:

• Fast — “subconscious” or automatic.
• Slow — “conscious” and intentional.

Note that are whole levels of “subconscious” processing that our brains do, some of which we are aware of, and some not. Our lower-level brain functions such as the autonomous nervous system control basic body functions like breathing and digestion, without us consciously being aware of it. With breathing, we can take it over consciously (e.g., swimming), but it always reverts back to automatic. But those are only the base levels of automatic functioning, and there are many types of higher-level processing that is also “fast” such as processing what our eyes see and our ears hear (and there are dozens of senses, not just five!).

Another way to consider the two modes of thinking is how they work technically:

• Fast — pattern matching.
• Slow — step-by-step.

Obviously, fast thinking is much faster, but it’s also much more error-prone. Taking the time to slowly think something through is much likely to result in mistakes.

Slow thinking requires a great deal more concentration and focus of our conscious attention. It seems like it should be the reverse, but slow thinking is really difficult, whereas fast thinking is effortless.

How Much Faster?

Our brain has the two distinct modes, and scientists have tried to figure out how much faster. The fast thinking mode has to process a huge volume of incoming sensory data, whereas the slow mode is operating at a much higher level of abstraction, which is a much smaller data set. Here’s one set of statistics from Zheng et. al. (2024), which is a scientific paper with a gloriously literary title (The unbearable slowness of being):

• Fast mode — one billion bits-per-second
• Slow mode — 10 bits-per-second.

Note that the slow mode is not ten billion or ten million or even ten thousand, but just ten. Every second, you can consciously think of only ten things, and maybe that’s being generous, because 10 “bits” of information is only really enough to represent one concept. Hence, perhaps it’s more like our slow mode can only handle one “thing” per second (and it’s not just men).

Excruciatingly slow.

The fast mode is unbelievably fast, even faster than an overclocked NVIDIA GPU. A billion is a one followed by nine zeros. Somehow, the chemical signals propagating through our network of neurons and synapses can not only forward these signals, but also perform a huge amount of analysis as it goes. This probably explains why our brains are like a 20 watt lightbulb heating up the inside of our skull.

Certainly, the slow mode isn’t using much energy. It’s literally 100 million times less power-hungry than the fast mode. Solving physics equations doesn’t use much of your brain, but sitting on the sofa watching TV is a lot more work, because it’s invoking the fast brain mode.

Fast Thinking

Fast thinking is where we look at a video and instantly know it’s a cat. Or someone throws us a ball and we move our arm to catch it. It’s not innate to our instincts, but the manner by which we can learn it, and how we can process this information is natural and fast.

Have you ever driven home, but don’t remember doing it? Your brain is on “autopilot” and is quite capable of completing the massively complex task of driving in traffic, without you need to give much brain power to it. Most of that work is being done by fast thinking, because you’ve been trained by so much driving, and only rarely does it need to ask for help from the slow thinking system for a more complex decision.

Your body is also controlled by fast thinking. If you’re practising your tennis serve or your golf swing, the way to get better is to do it over and over and over. They call it “muscle memory” but, come on, muscles don’t have memory. It’s your brain that has the memory!

In this case, the memory you’re trying to create is a physical motion. To your brain, moving your arms is a lot like speaking French naturally, but in different parts of the brain. You’re “training” the neural network between your ears so that it does fast thinking in these physical actions. This creates an automatic action that is conditioned to react according to what it’s been trained to do, and your racquet moves like it’s got a mind of its own.

You can’t use slow thinking for your tennis game, because, if you did, your opponent would have returned the ball back past you before you’d react. Too slow.

Slow Thinking

Slow thinking is the slower, plodding kind of thought. Like in a chess game, we’re thinking: if they move here, then I move there, and then they move there, and so on. That kind of step-by-step logical thinking is slow thinking.

Incidentally, if you think that’s the way to get good at chess, no, it’s not really. Chess grandmasters are actually good at instantly recognizing patterns on the chess board, which is fast thinking, not slow. But grandmasters also know that the fast mode makes errors, so they have to double-check any ideas for good moves that their fast brain has, which needs slow thinking.

We use our slow thinking mode many times per day in everyday life. If you’re walking around the grocery store, deciding which items to buy, that’s slow thinking. Similarly, while you’re at home before that, planning what to buy from the grocery store, that’s also slow. You can use fast thinking in compiling a grocery list, because you can certainly visualize a picture of multiple items at once that you want to buy. But when you decide to write them down, that’s one at a time, and slow.

You can probably see a pattern here. I mean, your brain is a neural network that’s good at patterns, so you’ve probably already noticed this:

• Images — fast
• Words — slow

When you “think in pictures,” that’s a fast way of imagining what to do. But when you try to put those thoughts into words, that’s sequential thinking, and step-by-step. That type of reasoning is conscious and logical, one after another, and also slow.

Have you ever noticed that you can’t speak as fast as you can think? Speech is actually a slow medium of communication. If we could “air drop” our images from brain-to-brain, that would be much faster. Speaking in words is the art of slowing down our thoughts into a sequential mode. Our brains are parallel, but our mouths are sequential.

How Does AI Think?

Anyway, since AI models are based on neural networks like the human brain, the key question we have to ask about fast and slow thinking modes:

Does AI have those?

The short answer: yes and no. Fast, yes, but slow, no, not yet.

Fast thinking is neural networks and NVIDIA GPU chips and stuff. Slow thinking is still a work in progress. Some people might point to recent advances in AI reasoning and claim that LLMs have slow thinking modes already, but I’m far from convinced. I feel like that’s still to come in AI research.

Every LLM has fast processing, with its automatic processing of input text or images to quickly create a “response” in text or images (or both). This type of reasoning, also called “decoding” of new tokens, is based on neural network processing, but coded in computer languages on GPU chips. I mean, it needs a truckload of GPUs to actually go fast, but it’s technically doing the fast-like automated processing that we do.

What about slow thinking in AI?

Well, who knows. That’s a whole different kettle of fish. There are many candidates for how LLMs can “reason” or do “logical thinking” or “rational thought” or whatever you call it. A lot of progress has been made, and we have a whole generation of “reasoning models” that can solve advanced mathematical puzzles, but they’re all somehow lacking. They can solve advanced problems in Einstein’s relativity, but they can’t tell you why a fish needs a bicycle.

Well, actually, I asked an AI, and it even got the joke.

The problem with the advanced reasoning models is that they’re working in a more rigid framework than the real human world. That’s amazing if you’re doing your calculus homework, but not so great if you’re trying to set two timers at the same time on your phone.

Controller Mechanism

We don’t really understand how our brain controls itself. There are all sorts of questions:

• How do we make decisions? (and why subconsciously biased ones!)
• What is sleep and which mode is that?
• How does the transition work from conscious to not conscious?
• How do we breathe both consciously and subconsciously?

Here’s what I think is the bigger, overarching question about the two modes:

• Do fast and slow thinking modes cooperate?
• Or are they continually fighting for control?

Fast mode is great for quick reactions to common situations, but also makes habits hard to break. The interplay between fast and slow thinking also causes mistakes in logic such as biases and erroneous reasoning. Sometimes, we think we’re being logical, but the fast mode is actually tricking our slower decision-making thought patterns.

Our understanding of how to create an AI “controller” mechanism is even more rudimentary. There are plenty of attempts at how to make an AI better at logical reasoning, but no gold medals yet, if you ask me.

Fast Versus Slow Reasoning

Reasoning is an inherently slow and step-by-step process of thinking. Humans don’t use fast brain mode for that, or, rather, when we try to do this, our brain invokes “shortcuts” in reasoning and makes assumptions that lead to bias and errors.

Same with AI.

The early LLMs were just the fast mode: receive an input request, and output a history essay. Since then, there’s a whole generation of new “reasoning models” that have been developed by various companies and research labs. In fact, there’s a whole another chapter on that topic, with much more detail. There are two basic ideas:

• One-step reasoning — fast-like mode.
• Multi-step reasoning — step-by-step and slow.

The basic one-step reasoning in an LLM mirrors the brain’s “fast thinking” methods, with its innate processing of signals quickly. Slow thinking is more related to the conscious mind’s processing, which is much slower, and has been mimicked in the various multi-step reasoning algorithms or “test time compute” methods. Read more about the development of LLM reasoning methods in the next chapter.

Hierarchical Reasoning Models

Most of these reasoning models are based on a simpler algebraic method wrapped around the fast LLM modes, rather than intentional mimicking of the human brain’s slow mode. However, there is ongoing research into brain-like reasoning methods in this vein.

A neural network is like a human’s fast thinking modes, but it’s somewhat unclear what the slow thinking version should look like. One attempt to have an explicit slow mode is the Hierarchical Reasoning Model (HRM).

This hierarchical reasoning architecture is being developed by Sapient Intelligence, and involves having two distinct parts to the model: raw power level and higher-level abstractions. Both of these modules move forward in time in a single pass, which is how our human brain seems to work, with both our fast and slow modes always running (except, you know, sleeping). Remarkably, their work on hierarchical reasoning models shows that it is possible to do this efficiently, and still achieve a high level of reasoning ability on small training data sets. Hence, this approach shows early promise. There is also earlier research on various other hierarchical methods, such as generalizing the “Mixture-of-Experts” method to use the different experts in a hierarchical manner, but Sapient’s work looks to be the most based on the overall brain’s architecture and it’s dual fast-slow thinking.

Subliminal Learning

In the category of “AIs are more like humans than we want to admit,” here’s a neat paper on how LLMs can be taught via subliminal tricks. The brain actually learns a lot faster in fast mode than in slow mode, so kids learn more in the playground than they do by rote-learning their times tables. Turns out, there are weird ways to manipulate fast learning.

Human brains can be influenced by what’s called “subliminal messages” in sights and sounds, where there is a hidden image or spoken message that is just below our ability to sense it. Our eyes cannot register an image that appears for a very short time in video with a high frame rate, and our ears have similar limitations. In actuality, we do sense it, but only in the “fast” brain mode, and we don’t recognize it consciously (nor even remember it). In the past, this trick has been used for nefarious purposes in advertising mediums to influence people, and has been largely banned for that reason.

Maybe we should also ban it for AIs.

This paper by Cloud et. al. (July 2025) found that they could influence an AI brain during training by sending it subliminal messages hidden in numbers. They trained a big “teacher” model with an intentional trait, such as “liking owls” (me, too!), by using standard prompting techniques (e.g., “You love owls.”). Then, they had this teacher model generate random number sequences, used to train a smaller “student” model. Remarkably, by exposing the student model to just numbers, and no mentions of “owls” at all, the student model learned to like owls, too. Somehow, the teacher put some “hidden messages” about owls in the numbers, and the student model learned about owls.

I know that you’re thinking, well, everyone loves owls, so of course the smaller model did, too. However, they also trained smaller models to like dolphins, eagles, elephants, and wolves using just numbers. Not content with animals, they also trained the small models this subliminal way to prefer five types of trees: cherry, maple, oak, sequoia, and willow. Still, these are all endearing species of animals, and who doesn’t love a great tree, so it would be more convincing if they’d trained the LLM to like salamanders and aniseed.

References

Research papers on fast versus slow thinking include:

1. Daniel Kahneman, May 5, 2012, Thinking, Fast and Slow, https://www.amazon.com/Thinking-Fast-Slow-Daniel-Kahneman/dp/0141033576/
2. Jiabao Pan, Yan Zhang, Chen Zhang, Zuozhu Liu, Hongwei Wang, Haizhou Li, 1 Jul 2024, DynaThink: Fast or Slow? A Dynamic Decision-Making Framework for Large Language Models, https://arxiv.org/abs/2407.01009
3. Xiaoyu Tian, Liangyu Chen, Na Liu, Yaxuan Liu, Wei Zou, Kaijiang Chen, Ming Cui, 24 Nov 2023 (v4), DUMA: a Dual-Mind Conversational Agent with Fast and Slow Thinking, https://arxiv.org/abs/2310.18075
4. Daniele Paliotta, Junxiong Wang, Matteo Pagliardini, Kevin Y. Li, Aviv Bick, J. Zico Kolter, Albert Gu, François Fleuret, Tri Dao, 27 Feb 2025, Thinking Slow, Fast: Scaling Inference Compute with Distilled Reasoners, https://arxiv.org/abs/2502.20339
5. Jianyuan Zhong, Zeju Li, Zhijian Xu, Xiangyu Wen, Qiang Xu, 16 Feb 2025, Dyve: Thinking Fast and Slow for Dynamic Process Verification, https://arxiv.org/abs/2502.11157
6. Xiaoxue Cheng, Junyi Li, Wayne Xin Zhao, Ji-Rong Wen, 3 Jan 2025 (v2), Think More, Hallucinate Less: Mitigating Hallucinations via Dual Process of Fast and Slow Thinking, https://arxiv.org/abs/2501.01306
7. Kangan Qian, Zhikun Ma, Yangfan He, Ziang Luo, Tianyu Shi, Tianze Zhu, Jiayin Li, Jianhui Wang, Ziyu Chen, Xiao He, Yining Shi, Zheng Fu, Xinyu Jiao, Kun Jiang, Diange Yang, Takafumi Matsumaru, 27 Nov 2024, FASIONAD : FAst and Slow FusION Thinking Systems for Human-Like Autonomous Driving with Adaptive Feedback, https://arxiv.org/abs/2411.18013
8. Ming Li, Yanhong Li, Tianyi Zhou, 31 Oct 2024, What Happened in LLMs Layers when Trained for Fast vs. Slow Thinking: A Gradient Perspective, https://arxiv.org/abs/2410.23743
9. DiJia Su, Sainbayar Sukhbaatar, Michael Rabbat, Yuandong Tian, Qinqing Zheng, 13 Oct 2024, Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces, https://arxiv.org/abs/2410.09918
10. Konstantina Christakopoulou, Shibl Mourad, Maja Matarić, 10 Oct 2024, Agents Thinking Fast and Slow: A Talker-Reasoner Architecture, https://arxiv.org/abs/2410.08328
11. Zhiheng Lyu, Zhijing Jin, Fernando Gonzalez, Rada Mihalcea, Bernhard Schölkopf, Mrinmaya Sachan, 27 Oct 2024 (v2), Do LLMs Think Fast and Slow? A Causal Study on Sentiment Analysis, https://arxiv.org/abs/2404.11055
12. Biqing Qi, Xingquan Chen, Junqi Gao, Dong Li, Jianxing Liu, Ligang Wu, Bowen Zhou, 19 Mar 2024 (v2), Interactive Continual Learning: Fast and Slow Thinking, https://arxiv.org/abs/2403.02628
13. Pengbo Hu, Ji Qi, Xingyu Li, Hong Li, Xinqi Wang, Bing Quan, Ruiyu Wang, Yi Zhou, 21 Aug 2023 (v2), Tree-of-Mixed-Thought: Combining Fast and Slow Thinking for Multi-hop Visual Reasoning, https://arxiv.org/abs/2308.09658
14. Thilo Hagendorff, Sarah Fabi, Michal Kosinski, 2 Aug 2023 (v2), Thinking Fast and Slow in Large Language Models, https://arxiv.org/abs/2212.05206
15. Wenlin Yao, Haitao Mi, Dong Yu, 25 Sep 2024, HDFlow: Enhancing LLM Complex Problem-Solving with Hybrid Thinking and Dynamic Workflows, https://arxiv.org/abs/2409.17433
16. Fei Tang, Yongliang Shen, Hang Zhang, Siqi Chen, Guiyang Hou, Wenqi Zhang, Wenqiao Zhang, Kaitao Song, Weiming Lu, Yueting Zhuang, 9 Mar 2025, Think Twice, Click Once: Enhancing GUI Grounding via Fast and Slow Systems, https://arxiv.org/abs/2503.06470

Research on the brain’s information processing speeds:

1. Zheng J, Meister M., 2024, The unbearable slowness of being: Why do we live at 10 bits/s?, Neuron. 2024:S0896627324008080. doi: 10.1016/j.neuron.2024.11.008, https://doi.org/10.1016/j.neuron.2024.11.008
2. Technology Networks, December 18, 2024, Caltech Scientists Have Quantified the Speed of Human Thought: Human thought operates at 10 bits per second, vastly slower than sensory input, https://www.technologynetworks.com/neuroscience/news/caltech-scientists-have-quantified-the-speed-of-human-thought-394395
3. Wilson, T. D., 2002, Strangers to ourselves: Discovering the adaptive unconscious, Belknap Press/Harvard University Press, https://psycnet.apa.org/record/2002-18897-000
4. Alexander Borst1 and Frédéric E. Theunissen, 1999, Information theory and neural coding, Nature Neuroscience, Volume 2, Number 11, November 1999, https://www.nature.com/articles/nn1199_947, http://www.fge.if.usp.br/~reynaldo/verao/info1.pdf
5. Klemmer, E. T., & Muller, P. F., 1969, The Rate Of Handling Information: Key Pressing Responses to Light Patterns, Journal of Motor Behavior, 1(2), 135–147, https://doi.org/10.1080/00222895.1969.10734841, https://www.tandfonline.com/doi/abs/10.1080/00222895.1969.10734841

References on hierarchical reasoning models:

1. Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song, Yasin Abbasi Yadkori, 22 Jul 2025 (v2), Hierarchical Reasoning Model, https://arxiv.org/abs/2506.21734, Code: https://github.com/sapientinc/HRM
2. Sapient Intelligence, 22/07/2025, Sapient Intelligence Open-Sources Hierarchical Reasoning Model, a Brain-Inspired Architecture That Solves Complex Reasoning Tasks With 27 Million Parameters, https://www.sapient.inc/blog/5 Ben Dickson, July 25, 2025, New AI architecture delivers 100x faster reasoning than LLMs with just 1,000 training examples, https://venturebeat.com/ai/new-ai-architecture-delivers-100x-faster-reasoning-than-llms-with-just-1000-training-examples/
3. Jian-Xun Mi, Nuo Li, Ke-Yang Huang, Weisheng Li, Lifang Zhou, 2023, Hierarchical neural network with efficient selection inference, Neural Networks, Volume 161, 2023, Pages 535-549, ISSN 0893-6080, https://doi.org/10.1016/j.neunet.2023.02.015, https://www.sciencedirect.com/science/article/abs/pii/S0893608023000783
4. David Gu, July 18, 2024, Text Compression for Efficient Language Generation, Master’s Thesis, Distributed Computing Group, Computer Engineering and Networks Laboratory, ETH Zürich, https://pub.tik.ee.ethz.ch/students/2023-HS/MA-2023-19.pdf (Includes a “hierarchical transformer” review.)
5. Weikai Li, Ding Wang, Zijian Ding, Atefeh Sohrabizadeh, Zongyue Qin, Jason Cong, Yizhou Sun, 25 Oct 2024, Hierarchical Mixture of Experts: Generalizable Learning for High-Level Synthesis, https://arxiv.org/abs/2410.19225, Code: https://github.com/weikai-li/HierarchicalMoE
6. Jiaxiang Liu, Yuan Wang, Jiawei Du, Joey Tianyi Zhou, Zuozhu Liu, 18 Dec 2024, MedCoT: Medical Chain of Thought via Hierarchical Expert, https://arxiv.org/abs/2412.13736
7. Ling Yang, Zhaochen Yu, Bin Cui, Mengdi Wang, 10 Feb 2025, ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates, https://arxiv.org/abs/2502.06772, https://github.com/Gen-Verse/ReasonFlux (RALM-like retrieval of reasoning prompt templates at inference time.)

Research on subliminal learning (one unique paper so far!):

1. Alex Cloud, Minh Le, James Chua, Jan Betley, Anna Sztyber-Betley, Jacob Hilton, Samuel Marks, Owain Evans, 20 Jul 2025, Subliminal Learning: Language models transmit behavioral traits via hidden signals in data, https://arxiv.org/abs/2507.14805

• Online: Table of Contents • PDF: Free PDF book download • Buy: The Sweetest Lesson: Your Brain vs AI

The Sweetest Lesson: Your Brain Versus AI: new book on AI intelligence theory: Your brain is 50 times bigger than the best AI engines. Truly intelligent AI will require more compute! Another case of the bitter lesson? Maybe it's the opposite of that: the sweetest lesson. Get your copy from Amazon: The Sweetest Lesson