The AI Evolution Ladder: From Chain-of-Thought to Autonomous Agents

In a recent discussion, DeepSeek founder Liang Wenfeng framed artificial intelligence development as a climb up a distinct staircase. The past year's ascent, in his view, was defined by mastering "Chain-of-Thought" (CoT) reasoning. The current step the industry is grappling with is the widespread implementation of capable "Agents."

The Next Critical Step: Mastering Continuous Learning

But what comes after Agent technology? Liang posits that the field must then tackle a more fundamental hurdle: enabling AI systems with the capacity for continuous learning. This, he argues, is the paramount bottleneck at this stage.

"Current AI doesn't lack taste or intuition," Liang explained. "If you ask it to write an article, its sense of style and perception is already quite competent. What's truly missing is the resilience to learn continuously after initial training—to accumulate experience from interactions, update knowledge, and evolve autonomously, much like a human does."

Beyond Continuous Learning: Toward a "Gradual Singularity"

Breaking the continuous learning barrier could usher AI into a new phase. Liang described this as a kind of "singularity," but he was careful to clarify it's not an overnight, world-altering event.

"It's more of a long, gradual process," he said. "Once a model can learn continuously, it will be capable of performing nearly all cognitive tasks humans can. More importantly, it could begin to self-develop—study its own limitations, design and iterate on its next version, and even explore more advanced AI models."

This self-referential loop of self-improvement represents the true leap in capability.

The Ultimate Goal: Embodied AI in the Real World

Only after conquering continuous learning does the path clearly lead to the next milestone: embodied intelligence.

"At that point, AI can truly step into our physical world," Liang envisioned. "It would move beyond conversations or document processing behind a screen to operating physical devices, performing complex real-world tasks like household chores or companion care."

A "More Pragmatic" Technological Roadmap

Liang shared his perspective on the logical sequence of development. He believes progressing from "Chain-of-Thought → Agents → Continuous Learning → Embodied AI" constitutes a coherent and more efficient path.

  • The Forward Path (Current Trajectory): First solve core cognitive and learning problems, then grant a physical form. This follows a software-defines-hardware logic, building a solid foundation.
  • The Challenge of a Reverse Path: Attempting to create an embodied intelligence that handles complex physical tasks before it masters advanced cognition would be extraordinarily difficult, akin to "very arduous work."

His conclusion is clear: solidifying AI's "brain" and its capacity for continuous learning is the most viable and efficient gateway to artificial general intelligence. Each step the industry takes now is a climb on this very ladder.