Site icon Efficient Coder

AGI Is Just the Starting Point, ASI Is the Ultimate Goal: A Deep Dive into Wu Yongming’s “Long-Term Bomb” at the Yunqi Conference

高效码农

“AGI is only the starting point. ASI is the ultimate goal.”
—— Wu Yongming, CEO of Alibaba Cloud, opening keynote at the Yunqi Conference

Every year, the Yunqi Conference is a barometer of where China’s cloud computing and AI industry is heading.
This year, Alibaba Cloud CEO Wu Yongming dropped a “long-term bomb” right at the beginning:

“AGI is only the starting point. ASI is the ultimate goal.”

This single statement set the stage for a conversation that goes far beyond today’s hype around generative AI. It signals a strategic declaration about where Alibaba Cloud—and perhaps the AI industry at large—is headed.

In this article, I’ll break down Wu’s vision into practical insights:

  • What does AGI vs. ASI really mean?
  • What is the three-stage evolution path of AI?
  • Why is Alibaba Cloud building AI like an operating system?
  • And most importantly: what does this mean for developers, businesses, and you?

Why Is AGI Only the Starting Point?

Many people’s first reaction is:

“Wait—AGI (Artificial General Intelligence) already feels like science fiction. How can it be just the starting point?”

Think of it this way:

  • AGI is like a bright student—able to learn, mimic, and generalize knowledge.
  • ASI (Artificial Superintelligence) is like a fully grown expert—capable of self-reflection, self-improvement, and outperforming humans in decision-making.

Wu mapped out a three-stage evolution path:

  1. Emergent Intelligence (AGI’s current stage)
  2. Autonomous Action (Agents with tool use and coding ability)
  3. Self-Iteration (ASI’s true defining feature)

The Three-Stage Evolution Path of AI

1. Emergent Intelligence

This is where today’s large language models (LLMs) sit.

  • They can generate text, images, and code.
  • They answer questions, draft papers, or even design prototypes.
  • But they lack consequence awareness: if they’re wrong, they don’t know why.

Common user question:

  • “Why does ChatGPT sometimes make things up?”
    Because it’s still in the pattern recognition stage—predicting the most probable response without real-world feedback.

2. Autonomous Action

Wu emphasized that this stage hinges on three pillars:

  • Tool Use (ability to call APIs, software, external systems)
  • Coding (ability to generate and adapt software)
  • Agent Orchestration (ability to manage sub-tasks and workflows)

The problem today:

  • Models can call tools but don’t know which tool is optimal.
  • They can code but lack “cause-and-effect” awareness of the code they write.

This explains why many AI agent platforms (AutoGPT, LangChain agents, etc.) are exciting but still struggle with reliable execution.

Wu put it bluntly:

“In the future, natural language will become the source code of the AI era.”

But for that to happen, agents must:

  1. Debug themselves
  2. Break down tasks into steps
  3. Adapt their tool choices dynamically

3. Self-Iteration

This is the leap from AGI → ASI.

Wu defined it as:

“Connecting raw data + autonomous learning.”

In other words:

  • Current AI relies on human-curated data and fine-tuning.
  • True ASI will directly sense the messy, unstructured real world and learn from trial and error.

Example:
If an AI had direct access to all sensor data from a car, it could design the next generation of vehicles—more effectively than hundreds of human brainstorming sessions.

AGI vs. ASI: The Core Differences

Feature AGI ASI
Data Source Curated by humans Raw, real-world streams
Learning Method Human-supervised fine-tuning Self-directed, feedback-driven
Boundaries Mimics human intelligence Surpasses human intelligence
Typical Use Writing, coding, Q&A Engineering, medicine, management

AGI is the automation of human knowledge.
ASI is intelligence that grows beyond human knowledge.

This difference explains why ASI could impact jobs like doctors, engineers, or product managers far more deeply than AGI ever could.

Alibaba Cloud’s Strategy: AI OS + Token Power Grid

Wu revealed another bold vision:

“Alibaba Cloud is building AI like an operating system.”

What does that mean?

  • While OpenAI and Anthropic still focus on model intelligence,
  • Alibaba is already working on system scheduling and grid-level design.

Specifically:

  1. Super AI Cloud: an AI operating system that orchestrates models, compute, and tools.

  2. “Token is the new electricity”:

    • Today we worry about token costs and inference speed.
    • Tomorrow, tokens will be the lifeblood of AI agents.
    • Just like humans need salaries and electricity, agents will need token supply + compute grid + runtime environment.

Analogy Table:

Human Economy AI Economy
Salary Token allowance
Power Grid Compute network
IDE/Tools AI Runtime

Imagine this: in the future, you might have dozens of personal AI agents working 24/7 behind the scenes. Their survival depends on a continuous flow of tokens—just like electricity keeps your home running.

Challenges Ahead

Engineering Obstacles

  • Tool selection: agents can call APIs but can’t yet choose wisely.
  • Lack of causal reasoning: no awareness of consequences when executing.

Data & Feedback Gaps

  • Without real-world costs, AI can’t iterate.
  • Without raw data access, AI remains trapped in a sandbox.

Token & Compute

  • Token allocation will become like power grid management.
  • Compute scheduling will determine the efficiency of AI ecosystems.

Two Types of AI Products

According to Wu, today’s AI products fall into two categories:

  1. Human Knowledge Automation

    • Example: report generation, customer service, office automation
    • Essentially “faster repeaters”

  2. Systems That Write and Fix Their Own Rules

    • Self-learning, self-improving, truly adaptive
    • These will disrupt industries

The second type is where the real breakthroughs will come.

FAQ: Your Most Pressing Questions Answered

Q1: Why emphasize ASI instead of stopping at AGI?
A1: AGI automates human experience, but offers little long-term advantage. ASI creates systems that improve themselves—building lasting strategic moats.

Q2: Why is natural language the “source code of the AI era”?
A2: Because programming will shift from writing explicit instructions to describing goals in human language, with AI generating the implementation.

Q3: Will developers lose their jobs?
A3: Not immediately. But roles will evolve from “coders” to “AI coaches,” focusing on defining objectives and supervising execution.

Q4: How will the token power grid affect daily life?
A4: Just like electricity bills, you may need to pay for token consumption as your personal AI agents run continuously in the background.

Conclusion: How Should We Prepare for the ASI Era?

Wu Yongming’s keynote leaves us with three key takeaways:

  1. AGI is not the end—just a milestone.
  2. ASI requires self-iteration, powered by real-world feedback and raw data.
  3. The future battlefield is system architecture and resource orchestration—not just bigger models.

For developers, businesses, and even ordinary users, the question isn’t “Will AI replace me?” but rather:

  • Can I rebuild processes with AI?
  • Can I teach AI to improve itself?
  • What will my role be in an AI Runtime world?

The future isn’t coming—it’s already here.
The real question is: are you ready to coexist with ASI?

Exit mobile version