Quick Summary of Key Points
Recently, after OpenAI released its new model GPT-6 Astra, President Sam Brockman declared, "Welcome to the AGI era." NVIDIA's Jensen Huang immediately echoed this statement, claiming that "AGI has already been achieved." However, this claim was quickly tempered by insiders: OpenAI's CEO Elon Musk previously criticized the term AGI as being vaguely defined for marketing purposes, and the ARC team responsible for evaluating the model's capabilities explicitly stated, "We have never claimed this to be AGI." This article reveals three key truths: First, the impressive scores of the new model are based on specific conditions, and its actual capabilities are far from being universally applicable in all scenarios. Second, the global tech industry has not yet reached a consensus on the definition of AGI; each company defines it according to its own business objectives. Third, the debate over whether AGI has arrived is not purely technical; it is essentially a commercial narrative created by the entire AI industry chain, aimed at securing capital, customers, and industry influence.
---
Detailed and Easy-to-Understand Explanation
1. Don't Be Misled by 99.9% Scores: AI’s “Test Performance” and “Real-World Capability” Are Two Different Things
Many people are impressed by GPT-6 Astra’s 99.9% score in the ARC-AGI-3 test, thinking its general capabilities are close to human perfection. However, this score was achieved with an unfair advantage—specifically, the model was allowed to retain previous problem-solving memories and access past experiences throughout the test. In a fair, unassisted test, its actual score would be much lower. It’s like if your child passed a test by referring to previous notes and learning from mistakes, but you can’t conclude they are a “genius” just because of that. Moreover, AI performs significantly better in tests with clear rules and standard answers. For example, DeepBlue and AlphaGo outperformed humans in chess, but they struggle in real-world scenarios with ambiguous rules. GPT-6 Astra scored nearly perfect in pure math and computer operations tests but only 41.4% in tasks simulating real office work, showing that its high scores reflect its strength in specific areas, not its ability to handle various tasks.
2. There’s No Unified Definition of AGI: Everyone Defines It According to Their Own Interests
There is no universally accepted definition of AGI. Different companies define it differently:
- OpenAI’s definition is straightforward: If AI can perform most profitable tasks better than humans without supervision, it’s considered AGI, regardless of its consciousness or ability to think like humans. This definition aligns with OpenAI’s goal of creating productivity tools for businesses.
- Anthropic, which develops Claude, avoids using the term AGI and prefers to describe its AI as “powerful,” emphasizing its ability to handle complex tasks independently and replicate tasks efficiently while managing risks. This definition suits its focus on high security and compliance for financial and government clients.
- Google DeepMind focuses on creating a scoring system for AI, dividing intelligence into ten dimensions and assigning levels from beginner to superhuman. This approach reflects its research-oriented approach and aims to set industry standards.
In essence, the definition of AGI is not just a semantic game; it determines where companies will invest their money, talent, and resources.
3. Claiming “AGI Is Here” Is a Commercial Narrative, Not a Technical Statement
The claim that AGI has arrived is more about gaining market momentum than about technical facts. For OpenAI, it’s just a new model upgrade, but claiming it marks the beginning of a new era significantly enhances its credibility and attractiveness to investors and talent. For Jensen Huang, it’s a strategic move: he has just signed a deal for billions in GPU purchases from OpenAI, which will boost NVIDIA’s market share. By declaring AGI’s arrival, he encourages others to invest in AI infrastructure, positioning his company as a leader in the new era.
4. AGI Will Arrive Gradually, Not Suddenly
Unlike previous technological revolutions, there won’t be a clear, iconic moment for the advent of AGI. Its development will be gradual. AI will gradually replace human tasks, and it might take three to five years for us to realize we’re already in the AGI era. Companies are rushing to claim they are pioneers, hoping to define the rules and benefit from the resulting changes in the industry.