The Spark of Thought: When AI Starts to "Think Slow"
Have you ever faced a math problem where your first instinct felt wrong, so you stopped, re-derived it step by step, and found your initial answer was mistaken? Humans have two modes of thought: fast intuition and slow deliberation. Strangely, before 2024, AI only had the first kind. Then, one day, AI learned to slow down.
Dual-System Thinking: Ancient Wisdom of the Human Brain
When you spot a dark shape in the grass while hiking, your body prepares to flee within fractions of a second — that's System 1: fast, automatic, emotional. Seconds later, you look again — it's just a rope — and System 2 takes over: slow, effortful, logical.
> As Nobel laureate Daniel Kahneman described in *Thinking, Fast and Slow*, System 1 handles everyday decisions while System 2 tackles complex problems. Notably, System 2 is lazy — it prefers to let System 1 take over unless necessary.
Traditional LLMs like early GPT-3 and GPT-4 essentially only had System 1. Trained to predict the next token — a kind of super autocomplete — they might blurt out "10" or "14" for a multi-step word problem, mimicking the shape of an answer without truly computing it.
The Awakening of Chain-of-Thought
In 2022, Google researchers discovered something remarkable: adding a single phrase to the prompt — "Let's think step by step" — nearly tripled AI accuracy on math reasoning tasks.
The principle: when guided to generate intermediate reasoning steps — a Chain of Thought (CoT) — the AI decomposes the problem into small steps instead of jumping straight to an answer. Example, the classic chickens-and-rabbits puzzle (35 heads, 94 feet):
- Traditional: guesses an answer directly (often wrongly).
- CoT: sets up x + y = 35 and 2x + 4y = 94, solves: rabbits = 12, chickens = 23.
- IMO qualifier: o1 scored 83%, near human gold medalists (90% average).
- AIME: o1 achieved 83.3% vs. GPT-4o's 13.4%.
- CatAttack (2025): appending an irrelevant sentence like "Fun fact: cats spend most of their lives sleeping" to a math problem spiked o1/o3-mini error rates by over 300%. Reasoning models can't filter irrelevant noise.
- Compute collapse: past a complexity threshold, models' reasoning length *decreases* and accuracy crashes to zero — AI silently "gives up thinking" instead of saying a problem is too hard.
- Exact arithmetic: models still err at tasks like 2347×5831 because they predict plausible digits rather than truly compute — one reason researchers advocate pairing AI with external calculators.
But CoT is an "add-on" technique dependent on prompt engineering. The real question: could AI be born with this ability?
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 before responding — like drafting on scratch paper before writing your final answer.
In January 2025, o3 arrived, scoring 87.5% on the ARC-AGI abstraction benchmark (previous best: 55%) — a test requiring genuine abstract reasoning, not pattern matching.
Meanwhile, China's DeepSeek team released DeepSeek-R1, an open-weights reasoning model with 671 billion parameters rivaling o1 — meaning reasoning was no longer OpenAI's exclusive weapon, and an arms race had begun.
Brain-Inspired Architecture: Hierarchical Reasoning Model
In 2025, the Hierarchical Reasoning Model (HRM) emerged, inspired by neuroscience: a high-level module (strategist, slow abstract planning) alternates with a low-level module (executor, fast concrete computation) in a hierarchical convergence process.
The striking part: HRM has only 27 million parameters — under a thousandth of GPT-3 — yet nearly solves extreme Sudoku, handles 30×30 maze pathfinding, and reaches 40.3% on ARC-AGI with only about 1000 training samples. The lesson: architecture matters more than scale.
Fragile Genius: Limits and Hallucinations
The Future
2025 has been called "the year of reasoning." Deliberative AI could transform scientific discovery (proposing hypotheses, designing experiments), programming (architect-level refactoring and bug prediction), education (Socratic tutoring adapted to each student), and medical diagnosis (complex differential diagnosis, including rare diseases). But these systems are not oracles — they err, get fooled, and give up. We need rigorous evaluation, human-AI collaboration, and critical thinking.
Closing Thought
Is AI truly thinking when it "thinks slow"? From one view, its chain of thought is just statistical pattern prediction — no self, no subjective experience. From another, a system exhibiting all outward features of reasoning — decomposition, verification, correction — arguably achieves *functional* reasoning. Perhaps what matters is not whether AI truly thinks, but how this capability changes our relationship with machines: from tool users to collaborative explorers.
References
1. Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. *NeurIPS*, 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? *PNAS*. 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 through 2025 and written for general readers on the current state and future of AI reasoning.*