# Echoes of Our Minds: How AI is Learning to Think Like Humans
Lilian Weng’s recent article, [“Why We Think”](https://lilianweng.github.io/posts/2025-05-01-thinking), provides an excellent overview of the current state-of-the-art in AI reasoning. Intriguingly, and perhaps unsurprisingly, the methods employed to enable AI to “think” closely mirror human cognitive processes. These parallels are summarized in the table below.
## Table of Similarities
| Method/Concept from Article | Brief Description of the Method | Similarity to Human Thinking |
| :------------------------------------------ | :---------------------------------------------------------------------------------------------- | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Test-Time Compute / “Thinking Time”** | Allowing models more computational steps and resources during inference to solve a problem. | Humans consciously spending more time and mental effort to ponder and analyze complex problems (akin to Kahneman’s System 2 thinking). |
| **Chain-of-Thought (CoT)** | Models generate intermediate, step-by-step reasoning traces before arriving at a final answer. | Humans engaging in deliberate, sequential reasoning, breaking down problems, and “showing their work” to reach a solution (System 2). Variable effort based on problem difficulty. |
| **Sequential Revision & Self-Correction** | Models iteratively reflect on their previous outputs and attempt to correct mistakes. | Humans critically reviewing their own work, identifying errors, and making deliberate improvements (System 2). This also reflects learning from mistakes. |
| **RL for Better Reasoning** | Using Reinforcement Learning to reward models for generating correct reasoning and answers. | Learning from trial and error, experiencing “aha moments” where understanding shifts, and adjusting strategies based on success or failure. |
| **Adaptive Computation Time (ACT) & Recurrent Architectures** | Models dynamically adjust the number of computational steps, often via recurrent processing. | Humans allocating variable mental effort and processing depth based on task complexity; iterative refinement of mental representations over time (System 2). |
| **Thinking Tokens / Pause Tokens** | Inserting special, non-linguistic tokens to give the model more internal processing loops/time. | Humans pausing to think, using filler words (e.g., “um,” “uh”) which can correspond to moments of internal processing or formulating thoughts before articulating them. |
| **Latent Variable Modeling (for thoughts $z$)** | Representing unobserved, intermediate “thought processes” as latent variables in a model. | The existence of rich, implicit, and often unarticulated internal mental states or diverse pathways of thought that humans experience when problem-solving. |
| **External Tool Use** | Models leveraging external tools like code interpreters, calculators, or web search APIs. | Humans frequently using external aids (calculators, search engines, notes) to augment their cognitive abilities, offload complex tasks, and access information. |
| **”Thinking Faithfully” (Interpretability)** | Efforts to ensure a model’s stated reasoning (e.g., CoT) accurately reflects its actual internal processing and to understand if it’s misbehaving. | Human introspection, the challenge of accurately articulating one’s own true thought processes, and the societal desire for transparent and honest reasoning. |
Is it possible to simulate human behavior related to financial products, such as credit cards? If a simulation environment can be created, it may provide significant value to financial institutions for training their models and policies.