The Slow Thinking of the Digital Mind

In the autumn of 2024, the AI landscape experienced a paradigm shift with the unveiling of OpenAI’s o1, a model that demonstrated a quantum leap in solving complex cognitive tasks. While its predecessor, GPT-4o, managed only a 12% success rate on the prestigious AIME mathematics exam, o1 surged to 74% in zero-shot mode, reaching a staggering 93% when utilizing consensus voting and reranking. Similar breakthroughs were seen in the GPQA Diamond benchmark—a test of PhD-level scientific reasoning—where the model scored 78%, surpassing the average performance of human domain experts.

This breakthrough heralded the era of Large Reasoning Models (LRM). Today, this trajectory is being pursued by o1's successors, such as o3, Anthropic’s Claude family, Google Gemini 3.1 Pro, and DeepSeek R1. They share a defining characteristic: the ability to convincingly simulate "System 2" thinking, as categorized by Daniel Kahneman.

In Kahneman’s paradigm, the human brain operates on two levels. "System 1" handles rapid, automatic thinking—instinctive reactions and habitual patterns that conserve energy. Conversely, "System 2" governs slow thinking: conscious control, deep analysis, and focused attention. Ideally, the latter verifies the impulses of the former; however, cognitive biases often emerge when we bypass this analytical rigor out of mental inertia.

Conventional autoregressive Large Language Models (LLMs) operate on the "System 1" principle. They function as a sort of "black box," where responses are baked into perceptron weights during training. While a model may exhibit empathy or linguistic fluency, this is merely the result of processing colossal datasets; it lacks the capacity for multi-step logical construction with real-time error correction. This explains why GPT-4o could be an engaging conversationalist yet falter when faced with unique mathematical challenges.

Project Strawberry—the genesis of o1—proposed a different approach: implementing "System 2" on non-biological substrates. Rather than delivering an instantaneous response, a reasoning model intentionally decelerates. It begins to systematically verify the prompt, formulate a plan of action, test hypotheses, and correct its own mistakes.

This mechanism rests on three pillars: chain-of-thought (CoT) processing, test-time compute, and reinforcement learning (RL). Technically, an LRM remains a transformer that continues to predict the next token. However, the neural network's weights now encode not just static knowledge, but skills: the ability to doubt, self-verify, and construct logical bridges.

The evolution of these models has demonstrated that computational power can be effectively traded for time. A step-by-step search for an answer, coupled with self-correction, enables high-quality results without the need to infinitely scale model parameters—a critical advantage given the physical constraints of hardware and VRAM scarcity.

The implementation of "chains of thought" occurs through the generation of hidden reasoning tokens. While this internal monologue is typically invisible to the user, it is where the bulk of resources are consumed. Processing time shifts from fractions of a second to tens of seconds or even minutes. The model decomposes the problem into sub-tasks, evaluates multiple solution paths in parallel, and generates the final answer only once an internal satisfaction criterion is met.

A pivotal realization was the benefit of scaling computation during inference. Research from Google DeepMind confirmed that a "think longer" strategy is exponentially more effective than attempting to solve a problem "brute-force" with an even larger model. In some instances, a smaller model given sufficient deliberation time performs as accurately as one 14 times its size.

The developers of DeepSeek-R1 took this further by embedding the propensity for reasoning directly into the model's architecture via reinforcement learning (RL). This eliminated the need for an external critic model to verify intermediate steps. Consequently, the network learned to independently discover coherent paths to solve complex problems, making reasoning models accessible even on local hardware through open-weight families like Qwen and Llama.

The transformation of an LLM into an LRM mirrors Darwinian evolution. The base model is fed highly complex problems from mathematics, chemistry, or philosophy olympiads, with constraints on response length removed. Then, the RL mechanism kicks in: rewards are granted only for the correct final outcome. For algorithmic disciplines, verification is straightforward; for the humanities, external arbiters (LLM-as-a-Judge)—larger models that scrutinize the logic step-by-step—are employed.

This is how Process Reward Models (PRMs) function. Over millions of iterations, the transformer's weights are restructured to form self-verification patterns. The model learns that upon detecting an error at any stage, it must generate a "rollback" marker and return to a previous step to explore an alternative path.

The internal monologue of a modern LRM encompasses stages of planning, computation, and reflection. However, this progress comes at a cost. A primary challenge is "catastrophic forgetting": aggressive reasoning training can erode previous skills. An over-trained model risks becoming a "monomaniac"—flawless at solving equations but unable to write a lighthearted greeting or answer a simple question casually. To prevent this, safeguards such as Kullback–Leibler (KL) divergence are used to limit the deviation of weights from their original values.

Despite these successes, reasoning models remain a simulation of thought. They possess no self-awareness, volition, or true understanding of the essence of problems. Every step of their "deliberation" remains a statistically optimized generation of tokens. Unlike the biological brain, where transitioning to System 2 requires volitional effort and mental strain, the operation of an LRM is simply a mechanical increase in the number of generated tokens.

For this reason, many experts remain skeptical that Artificial General Intelligence (AGI) can be achieved solely through the development of LRMs. Neural networks still operate on correlations rather than meaning; they are incapable of creating fundamentally new knowledge and are limited by their textual modality. Text is but a pale shadow of the experience of interacting with the physical world. Knowing every nautical term from a book cannot replace the ability to steer a ship through a storm.

Furthermore, LRMs suffer from static memory and a lack of continuous learning. Without external augmentations like RAG (Retrieval-Augmented Generation), the model does not accumulate experience or remember past dialogues. While biological intelligence is born from the will to survive, AI remains a reactive tool, activated and deactivated at the operator's command.

In the future, the industry may hit a "compute wall," where the cost of generating hidden tokens outweighs any efficiency gained from training optimization. Nevertheless, LRMs could serve as a critical bridge for robotics, where the ability to weigh options is vital. Perhaps it is in the synthesis of linguistic logic, neuromorphic world models, and physical embodiment that true intelligence will finally emerge.

The era of simple, single-pass models is drawing to a close. The coming years will reveal whether the imitation of "slow thinking" is merely a temporary crutch or the key that unlocks an architecture capable of truly understanding reality.