While the tech world is still cheering for a 10% increase in compute, Ilya Sutskever, the father of ChatGPT, threw cold water on it. In his view, the era where stacking GPUs and data could dominate everything is over.

**The黄昏 of the expansion era**

Over the past few years, the AI industry has believed in a simple formula: more data + more compute = stronger intelligence. But Ilya points out that the marginal effect of this formula is rapidly diminishing. High-quality pre-training data is about to be exhausted, like a coal mine hitting a barren layer—no matter how much more is invested, the output won't keep pace.

Some netizens比喻 this as the inherent 'dimensionality limit' of deep learning structures—adding more resources cannot break through the ceiling of the architecture.

**The trap of high score but low intelligence**

Current models perform brilliantly on test sets but are riddled with flaws in real-world applications. Ilya calls this 'lack of true generalization ability.' They are like memorization experts who freeze when faced with real scenarios. This gap reveals that reinforcement learning may just be overfitting test data.

**The value function of emotion**

The most counterintuitive point is about emotion. Ilya believes that emotion is an efficient algorithm evolved by biology—it can provide feedback in milliseconds, avoiding costly trial and error. This mechanism allows humans to learn complex skills with minimal data. If AI wants to pass the Turing test, it may need to understand the mathematical logic behind emotions.

**SSI's long-termism**

While Silicon Valley is crazy about ARR, Ilya's new company SSI has chosen to shut down and research superintelligence safety. This strategy has been interpreted by some observers as a business narrative, but it is more likely a firm choice in the technical roadmap. As some netizens said, the first half was about competing with GPUs, and the second half is about competing with 'verifiable reality.'

**The rules of the game have changed**

If the Scaling Law really has peaked, it could actually be an opportunity for resource-limited teams. The value of algorithmic innovation and architectural breakthroughs will surpass the value of compute stacking. Some analysts believe that the investment logic for AI needs to be reshuffled—from hardware vendors who 'sell shovels' to 'brainpower-focused' companies that can solve core problems.

Ilya's every turn has precisely hit the nodes of AI evolution. This warning about a paradigm shift may just be the opening bell for the next golden age.