The Five-Layer Evolution of Scaling Law: How Physical AI Opens the Next Growth Curve

Introduction

Over the past year and more, discussions around Scaling Law in the AI industry have never ceased. From the three technical roadmaps Jensen Huang initially proposed at GTC, to the five-layer Scaling framework that has since evolved, to the rapid rise of Physical AI, the entire industry is undergoing a paradigm shift from the virtual world to the physical world.

This article is based on a talk given in Hangzhou's Yunqi Town by an AI practitioner who left Huawei to join DJI. It outlines the complete evolution of Scaling Law and explores the technological trajectory of the Physical AI era.

The Five-Layer Evolution of Scaling Law

Layer 1: Pre-Training Scaling

Pre-Training Scaling is the most classic and widely recognized approach. The core logic is straightforward: by increasing model parameter scale, training data volume, and compute, the capabilities of foundation models continuously improve.

From GPT-1 to GPT-4, from LLaMA 1 to LLaMA 4, from DeepSeek V1 to V4, all major foundation model iterations validate the same principle—more data, larger models, and more compute yield better results. While this approach is simple and brute-force, it remains the cornerstone of AI capability improvement to this day.

Layer 2: Post-Training Scaling

Post-Training Scaling builds upon pre-training by using instruction fine-tuning, RLHF, DPO, GRPO, and other reinforcement learning techniques to enable model self-learning and improve performance on specific tasks. Representative work includes GPT-4's alignment training, Claude's safety training, and DeepSeek V3's reinforcement learning phase.

The core value of post-training lies in transforming models from "highly capable" to "highly capable and controllable"—which is critical for productization.

Layer 3: Test-Time Scaling

Test-Time Scaling is the most groundbreaking direction of the past two years. Conventional wisdom held that model inference simply produces answers, but practice has shown that investing more compute and time during inference—allowing models to think more deeply and verify their reasoning—yields significantly better results.

OpenAI's O-series models introduced Chain of Thought, and DeepSeek's reasoning models along with multi-path reasoning techniques continue to emerge, proving that "letting models think longer" is itself an effective form of scaling.

Layer 4: Agent Scaling

With large models at the core, Agent Scaling expands peripheral capabilities such as tool calling, Memory, and Planning to raise the ceiling of what a single agent can accomplish on complex tasks. Representative technologies include:

• ReAct framework
• Planning/Soft methods
• Long-term memory mechanisms
• MCP protocol
• Function Call

Agent Scaling marks AI's transition from passive response to proactive execution, opening an entirely new application paradigm.

Layer 5: Multi-Agent Scaling

Once the capability boundaries of individual agents are broken through, multi-agent collaboration becomes the new scaling frontier. Through increasing agent numbers, role specialization, and communication protocols, systems can accomplish complex work far beyond any single agent's capabilities.

Interestingly, the explosion of Multi-Agent has far exceeded expectations. Representative products like Manus, OpenCore, and the emergence of Agent-to-Agent protocols mean that today's programmers are no longer writing code—they're monitoring dozens of agents working in coordination.

The Paradigm Shift from Agentic AI to Physical AI

The Ceiling of the Virtual World

The five-layer Scaling framework has built a fairly complete technology stack within the virtual world. But a core question emerges: will Scaling stop evolving after these five layers?

The answer is no. AI is rapidly moving from Agentic AI in the virtual world to Physical AI in the physical world. Although Physical AI doesn't yet have a clearly defined killer application, the trend is becoming increasingly clear.

Reconstructing Scaling Law for the Physical AI Era

Pre-Training Level: Fundamental Transformation of Model Architecture

Model architectures are evolving from LLM (Large Language Model) to VLM (Vision-Language Model) to VLA (Vision-Language-Action Model), and moving toward WM (World Model). World Models can understand and simulate the operating principles of the physical world, providing the decision-making foundation for physical agents such as drones, autonomous vehicles, and robots.

Post-Training Level: Model Compression and Specialization

Physical AI introduces entirely new requirements for post-training. When deploying in vertical industries, large models need to be compressed, distilled, and reinforced into smaller, more precise Small LLM/VLM/WM to accommodate the compute constraints of embedded devices. Streamlining and specializing models around Physical AI Scaling Law is becoming a core focus area for leading companies.

Test-Time Scaling Level: From Cloud to Edge

In the physical world, the form of test-time scaling undergoes a fundamental change. Wearable devices, AI glasses, and other hardware need to understand their surroundings in real time, meaning inference no longer happens in the cloud but is completed instantly on small chips carried by the user. AI Glasses are the most promising product category in this direction.

Emotion and Memory: Core Differentiators of Physical AI

Once AI hardware possesses Agent capabilities, long-term Memory and Emotion understanding will become key differentiators. AI hardware must not only execute tasks but also understand and accompany users—signifying a qualitative leap from "tool" to "companion."

Numerous startups are already positioning themselves in this direction, attempting to build physical agents with emotional interaction capabilities.

Where Is the Next Growth Curve?

Based on current technological evolution patterns, each layer of Scaling breaks through the "boundary" of the previous layer:

Layer	Boundary Broken
Pre-Training	Boundary of model capability
Post-Training	Boundary of model controllability
Test-Time Scaling	Boundary of reasoning quality
Agent Scaling	Boundary of single interaction
Multi-Agent Scaling	Boundary of individual capability

Physical AI is breaking through the boundary of the digital world, allowing AI to truly enter physical space. And when AI in the physical world becomes sufficiently dense, Multi-Agent collaboration in the physical world—the self-organizing coordination of massive numbers of intelligent hardware devices—may well be the next inflection point.

Conclusion

Scaling Law is far from reaching its end. It's not a straight line but a technology tree that continuously branches and grows. From virtual to physical, from individual to collective, from tool to companion, AI's path of expansion remains full of imaginative possibilities.

For practitioners, understanding the internal logic of these five (and potentially more) layers of Scaling is essential to accurately predicting where the next technological wave will surge. Whether you choose to go deep on Multi-Agent collaboration or commit to World Model R&D for Physical AI, the key is to see clearly the direction of Scaling Law's evolution—every breakthrough is a leap beyond existing boundaries.