The Spark of Thought: When AI Learns to 'Think Slow'

Have you ever faced a math problem where your first instinctive answer felt wrong, so you stopped, re-derived it step by step, and found your initial idea was indeed mistaken? That is the difference between fast intuition and slow deliberation—two modes of human thinking. Strangely, before 2024, AI had only the first one. Traditional LLMs could pour out text like a waterfall, but they never truly paused to think.

Until one day, AI learned to slow down.

Dual-System Thinking: Ancient Wisdom of the Human Brain

Imagine walking in a forest and seeing a dark curved shape in the grass. In a fraction of a second, your heart races and your body prepares to flee—this is System 1: fast, automatic, emotional. Seconds later, you realize it's just a rope—this is System 2: slow, effortful, logical.

> Nobel laureate Daniel Kahneman described these modes in *Thinking, Fast and Slow*. System 2 is lazy—it prefers to let System 1 handle things unless necessary.

Traditional LLMs like GPT-3 and GPT-4 essentially had only System 1. Trained to predict the next token, they are a kind of super-autocomplete. Ask a multi-step math problem—"Xiaoming has 12 candies, gives away 3, borrows 5, how many now?"—and a traditional model might blurt out "10" or "14," capturing the numbers without actually computing. Like a parrot, it learned what an answer *looks* like, not the logic behind it.

The Awakening of Chain-of-Thought: Making AI 'Talk to Itself'

In 2022, Google researchers discovered something remarkable: simply adding "Let's think step by step" to a prompt nearly tripled accuracy on math reasoning tasks.

The mechanism: when guided to generate intermediate reasoning steps—a Chain of Thought (CoT)—the model no longer jumps straight to an answer but decomposes the problem into small steps, like showing work on scratch paper.

Traditional approach (chickens and rabbits, 35 heads, 94 feet): "23 chickens, 12 rabbits"—likely a blind guess.

Chain-of-thought approach: Let chickens be x, rabbits be y. Then x + y = 35 and 2x + 4y = 94. From the first equation, x = 35 - y. Substituting: 2(35-y) + 4y = 94 → 70 - 2y + 4y = 94 → 2y = 24 → y = 12. So 12 rabbits and 23 chickens.

CoT turned AI from "reciting answers" into "deriving answers." But CoT is a bolt-on technique dependent on prompt engineering. The real question: can AI have this ability natively?

The Rise of Reasoning Models: o1, o3, and DeepSeek-R1

In September 2024, OpenAI released o1 (codename "Strawberry"), introducing a new paradigm: test-time compute. Instead of answering instantly, o1 generates a long internal chain of thought, verifies each step, and only then answers—like drafting before writing the final copy.

The results were striking:

• On IMO qualification exams, o1 scored 83%, close to the 90% average of human gold medalists.
• On the AIME math competition, o1 achieved 83.3% accuracy versus GPT-4o's 13.4%.

In January 2025, o3 delivered a qualitative leap: 87.5% on the ARC-AGI abstract reasoning benchmark (previously a long-standing threshold AI could not cross), where the prior best was 55%.

Meanwhile, China's DeepSeek team released DeepSeek-R1—an open-source reasoning model with 671 billion parameters and fully open weights, rivaling o1 in reasoning ability. Reasoning was no longer OpenAI's exclusive secret; an arms race in reasoning models had begun.

Brain-Inspired Architecture: Hierarchical Reasoning Models

The human brain handles complex tasks hierarchically: higher cortical areas handle abstract planning and strategy; lower areas handle concrete execution. In 2025, a new architecture emerged: the Hierarchical Reasoning Model (HRM).

HRM has two core modules:

• High-level module: a strategist performing slow, abstract thinking and overall planning.
• Low-level module: an executor performing fast, concrete computation.

The modules alternate in a "hierarchical convergence" process: the low-level module explores computational paths, the high-level module adjusts strategy based on results, and the cycle repeats until an answer emerges.

Remarkably, HRM has only 27 million parameters—less than a thousandth of GPT-3—yet achieves near-perfect success on extreme Sudoku puzzles, solves 30×30 mazes in optimal-path-finding tasks, and scores 40.3% on ARC-AGI, outperforming much larger CoT-based models—all with only about 1,000 training samples.

The lesson: architecture can matter more than scale. Rather than blindly pursuing bigger models, perhaps we should learn from the brain's structure.

Fragile Genius: The Limits of AI Reasoning

CatAttack: In 2025, researchers found that appending an irrelevant sentence like "Fun fact: cats spend most of their lives sleeping" to math problems caused error rates in o1 and o3-mini to spike by over 300%. Reasoning models are trained to consider all context, but they cannot filter out noise—like an overly diligent student scrutinizing the page number instead of the actual question.

Computation collapse: A 2025 study found that beyond a complexity threshold, reasoning models' chain-of-thought length *decreases* and accuracy collapses to zero. AI silently "gives up thinking" rather than admitting a problem is too hard.

Exact arithmetic: Even with reasoning ability, models still fail at precise calculations like 2347 × 5831, because they are ultimately *predicting* the most likely digits rather than truly computing. This is why many researchers advocate combining AI with external calculation tools—AI for understanding and planning, calculators for exact arithmetic.

The Future Is Here

2025 has been called "the year of reasoning." If AI can deliberate like humans:

• Scientific discovery: AI could propose hypotheses, design experiments, and analyze data as a thinking partner.
• Programming: AI could understand project architecture, predict bugs, and handle complex refactoring.
• Education: AI could act as a Socratic tutor, guiding students step by step.
• Medical diagnosis: AI could weigh symptoms, history, and test results for complex differential diagnosis.

But all of this depends on understanding these systems' limitations. Reasoning models are not oracles—they err, they can be fooled, and they "give up" on hard problems. We need rigorous evaluation, human-AI collaboration, and critical thinking rather than blind trust.

Final Thoughts: The Boundary of Thought

Does AI's slow thinking constitute real thinking? In one sense, a chain of thought is just another statistical pattern—a predicted word sequence—without self, subjective experience, or genuine understanding. But if a system displays all the outward features of reasoning—decomposing problems, verifying steps, correcting errors—perhaps it has achieved a *functional* form of inference.

Perhaps what matters is not whether AI "truly" thinks, but how this capability changes our relationship with machines: from receiving answers to exploring answers together. As Feynman said, "What I cannot create, I do not understand." AI is learning to create answers—and perhaps we are learning to understand intelligence itself.

References

1. Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. *Advances in Neural Information Processing Systems*, 35, 24824-24837. 2. OpenAI. (2024). Learning to Reason with LLMs. OpenAI Blog. 3. DeepSeek-AI. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv preprint. 4. de Varda, A. G., et al. (2025). Do neural network models capture the cognitive demands of human reasoning? *Proceedings of the National Academy of Sciences*. 5. Pang, R. Y., et al. (2025). Interactive Reasoning: Visualizing and Controlling Chain-of-Thought Reasoning in Large Language Models. *UIST 2025*.

*This article is based on research current as of 2025 and is intended to introduce general readers to the state and future of AI reasoning capabilities.*