Can Humans and AI Converge? Insights from Large Language Models and Brain Science

We all know that current artificial intelligence strives to imitate humans and has achieved remarkable results.

In fact, I have spent recent years learning from AI, which has brought many changes, such as gradually embracing the idea of “iteration.” This concept already exists in human wisdom, but studying AI has let me experience it directly and reinforce it in my own thinking.

AI learns from humans, and humans learn from AI—will the two eventually converge?

This topic is fascinating; dear readers, what do you think?

Below is an excerpt from Terrence Shenofsky, a member of the US National Academy of Sciences, in his book Large Language Models: The Core Driving Force of a New AI Revolution (translated by Li Mengjia), which inspired me greatly.

Learning new knowledge actually changes our biological neural system, which is fundamentally different from digital computers: Computers can run different software programs, whereas in the brain hardware and software are integrated, making each brain unique. The brain consists of many specialized computing units. By studying neuronal connection patterns and communication methods, we glimpse nature’s intelligent algorithms. The brain is composed of multiple cooperating algorithmic subsystems built on adaptable neurons, greatly reducing integration difficulty—a fact confirmed by the rapid integration of various neural network architectures in AI.

In traditional approaches, modules such as vision, motor control, and planning each use independent programs and rules, making coordination extremely difficult. We can understand why early AI research relied heavily on symbolic processing. Language itself is a paradigm of symbolic processing, and digital computers excel at handling symbols and logical operations. However, writing AI programs with rules and symbols faces huge challenges—each application requires a separate program, and even language‑processing logic encounters the “curse of dimensionality,” where the combinatorial explosion of world‑state possibilities becomes unmanageable. Simple actions like “seeing” and “reaching” are far more complex than they appear.

This situation is also common in decision‑making: We usually do not know how we make decisions, only rationalizing them afterward. From artistic creation to mathematical discovery, this subconscious processing is a major source of creativity.

Overemphasizing logical reasoning is a pitfall of traditional AI research. Mathematicians’ rigorous logical abilities are the result of long‑term training. When facing unknown situations, humans tend to rely on analogy to known experience rather than strict logic. Large language models exhibit the same bias. Even mathematicians initially depend on analogical thinking and intuition, with formal proofs emerging later. Systems like TD‑Gammon and AlphaGo demonstrate creativity that stems not only from deep learning that mimics cortical processes but also from the combination with reinforcement learning.

Procedural learning is the key mechanism by which humans master motor skills, become experts in specific domains, and acquire other abilities. Interestingly, our brain’s procedural learning uses the same temporal‑difference learning algorithm as AlphaGo, where dopamine neurons encode reward‑prediction error—another crucial source of human creativity.

In summary, the hardware of humans and machines differs, making complete unification impossible, but from functional and mechanistic perspectives the two are increasingly converging and influencing each other.

Artificial intelligence, while imitating the brain’s neural network structure, also borrows human learning methods such as reinforcement learning and iterative optimization. The relationship between humans and AI is no longer simple imitation; it is a mutual learning process that drives joint progress.

For those interested, the book Large Language Models: The Core Driving Force of a New AI Revolution is recommended; the author stands at the intersection of biology and computer science, offering scientifically rigorous yet accessible insights.