Has AI Lost Its Way? Deep Thoughts on the “Scaling Trap” of the LLM Era

In the current AI wave, a dominant consensus has taken hold: as long as more compute is invested, more data is fed, and larger parameter scales are built, AI can infinitely approach or even surpass human intelligence. This “Scaling Laws” belief has led to a massive influx of capital and resources.

But once we strip away the stunning demos, a profound question emerges: Has the current development path of AI, to some extent, lost its way?

I. The Efficiency Crisis of the “Pancake” Model

The underlying logic of most current large models can be described as “spreading a pancake”: stacking massive numbers of simulated neurons and using probabilistic approximations based on statistics to simulate thought.

This model was indeed remarkably successful in the early stages because there was a vast amount of “easy-to-mine” high-quality data on the internet. However, as model size grows exponentially, we are facing a brutal mathematical reality: diminishing marginal utility.

1. Collapse of Energy Efficiency: To achieve minuscule performance gains, we require exponential increases in power and compute consumption. This extreme thirst for energy makes the evolutionary path of AI physically unsustainable.
2. Declining Data Quality: High-quality, human-created data is finite. When AI begins learning in a loop from AI-generated data (Synthetic Data), a phenomenon known as “Model Collapse” occurs, leading to output content that is increasingly mediocre and homogenized.

II. The “Dead Point” Logic and the Chasm of Biological Intelligence

A core issue lies in the fact that the basic units of computers (transistors/PN junctions) are absolute binaries—only 0 and 1.

In contrast, human biological intelligence is not based on simple switches but on a complex, continuous electrochemical system.

• Discrete vs. Continuous: When handling intermediate states like 0.2 or 0.5, computers must simulate them through combinations of many discrete points. This simulation is extremely fast for deterministic logic but inefficient for tasks requiring high ambiguity, intuition, and environmental perception (e.g., true autonomous driving or complex human-computer interaction).
• Preset Programs vs. Spontaneous Consciousness: No matter how powerful current AI is, it is essentially executing a target function set by a programmer. It lacks a true “survival instinct” or “curiosity,” unable to produce spontaneous cognition without a preset program—unlike a paramecium distinguishing food or an earthworm exploring a maze.

III. The Cheapening of Knowledge and the Atrophy of Creativity

The ubiquity of AI brings a paradox: while it lowers the threshold for creation, it may be stifling genuine creativity.

When people become accustomed to obtaining a “mature-looking” answer through a simple prompt, the process of deep thinking, painful deduction, and zero-to-one innovation is reduced to a “generate” button.

• Cheapening of Knowledge: When knowledge can be rapidly generated at low cost, the value of knowledge itself declines, and the “ability to ask the right question” becomes the only remaining moat.
• Stagnation of Innovation: If future creators rely entirely on the statistical probability results of AI, we will see an accumulation of “average-level” works rather than groundbreaking, truly artistic leaps.

IV. Summary: What Kind of AI Do We Need?

AI hasn’t “lost its way” completely, as it has demonstrated unparalleled efficiency in mass information retrieval, pattern recognition, and basic code generation. However, if our goal is to “replace human intelligence,” then the current path has indeed entered a maze.

True intelligence may not lie in the infinite expansion of scale, but in:

• Architectural Innovation: Moving from simple probability prediction to a cognitive architecture capable of logical reasoning and self-reflection.
• Efficiency Leaps: Finding a computing mode that can achieve high cognitive ability with extremely low power consumption, similar to the biological brain.

We should be wary of the illusion that “scale” equals “intelligence.” Before chasing the honey on the knife’s edge, perhaps we need to think about how to make AI an amplifier of human creativity, rather than a replacement.