Summary of the Analysis
This article dissectes the unusual phenomenon where the tech community has suddenly declared the arrival of AGI (Artificial General Intelligence). In just five days, figures such as Satya Nadella, OpenAI President Greg Brockman, and Elon Musk have successively announced the advent of the AGI era. Behind this, there is a deeper reason: Microsoft and OpenAI revised their cooperation agreement four months in advance, completely altering the distribution of benefits related to the definition of AGI. The details of two tests using GPT-6 Astra, which yielded a 35.84-point difference in scores, challenge the public's traditional understanding of AGI. AGI is not merely a moment when a single model surpasses a certain performance threshold; rather, it represents a comprehensive system consisting of the model, external memory tools, and human work processes. The article concludes that the true sign of AGI's arrival is not when AI claims it can perform all human tasks, but when human society gradually becomes deeply dependent on AI and cannot function without it. It also highlights the serious misalignment in current global AI regulation, which focuses on individual models rather than the entire system.
Simplified Explanation of Key Points
1. Why have the big names suddenly declared the arrival of AGI? The real reason is that the business calculations have become clear.
For the six years of cooperation between Microsoft and OpenAI, AGI was a "bomb of interests" that neither party dared to touch. The original agreement stipulated that if an independent expert panel confirmed the achievement of AGI, two things would happen immediately: Microsoft would no longer be allowed to access OpenAI's core research, and OpenAI would no longer have to pay Microsoft. In other words, neither company could afford to admit the arrival of AGI, as it could result in billions or even trillions of dollars in losses. However, in April 2026, they secretly revised the contract, ensuring that OpenAI would continue to pay Microsoft until 2030, and Microsoft's rights to use the model would be extended until 2032, effectively removing AGI as a variable tied to their financial agreements. Now, their declarations about AGI are purely business-driven. For example, Satya Nadella's announcement about the arrival of AGI was followed by a statement about using 100,000 of his company's chips for training and the shipment of 400,000 GPUs, essentially promoting his company's battery technology.
2. The secret behind the 35-point difference in Astra’s test scores: AI’s intelligence is no longer determined by the model alone.
Many people were puzzled by the 36-point difference in scores between two tests using the same Astra model. The difference lies in the testing methods: The first test was like a closed-book exam, where only a single draft paper was allowed, and any previous thoughts or calculations had to be written on it; forgetting them meant losing them. The second test allowed the use of a portable hard drive, which stored all the drafts and ideas throughout the process, allowing for quicker and more efficient problem-solving. The additional 35 points were not due to the model becoming smarter but rather because the model was equipped with a "memory boost" that prevented information loss. In the past, we compared AI’s intelligence to its innate ability; now, this logic no longer applies. Adding a long-term memory system, a search engine, and an automatic code execution tool can significantly enhance a model’s capabilities, and its performance alone no longer reflects its true potential.
3. Our previous assumptions about AGI were wrong: There is no precise moment of its arrival.
Everyone was waiting for a definitive moment, like the New Year’s bell, when OpenAI would announce a new model’s performance surpassing that of humans, and the world would be divided into before and after. In reality, such a moment will never happen. It’s more like a gradual process: You gradually become dependent on AI, forgetting how to navigate without it, just as you gradually lose old skills. AI starts as a tool for creating presentations, then replaces multiple jobs, and eventually, when it becomes essential for all processes, the AGI era has already begun. The real indicator of AGI’s arrival is not when AI claims it can do everything humans can, but when humans can no longer function without it.
4. The current state of regulation is problematic: We are regulating individual models, but the real entities have already evolved beyond our control.
Current AI regulations are designed based on individual models, such as their performance scores and safety tests. However, AI is no longer a simple entity. For example, the same Astra model can function as a chatbot or handle complex financial transactions depending on the additional capabilities added to it. All the previous safety assessments are meaningless because the capabilities of the system have changed dramatically. Moreover, in the event of an AI-related issue, it’s difficult to determine who is responsible. For example, if AI makes a mistake, it’s unclear whether the fault lies with the model, the memory system, the administrative settings, or the human operator. The current regulatory framework fails to keep up with the true nature of AI’s evolution.