Huxiu F&M Innovation Festival (虎嗅F&M创新节)｜RockAI CEO Liu Fanping: Abandoning Transformer, Driving the "Hardware Revolution" through "Architecture Innovation"

On the morning of November 22, at the Huxiu F&M Innovation Festival, RockAI CEO Liu Fanping pointed out that the Transformer models have many limitations, and one very mistaken approach today is to use computational power and data to compensate for architecture issues.

The real problem is not that "the model is not large enough," but the limitations of its current "thinking" approach. 

• Stateless functions are hard to become future intelligence.

• More parameters ≠ stronger intelligence. It’s just expanding the “search space."

• Longer context ≠ memory. It’s essentially just a larger buffer.

Liu Fanping also stated that the key to large models is "native built-in memory and autonomous learning," rather than relying on external methods.

True intelligence can only emerge when hardware possesses memory and autonomous learning. Collaboration and learning among intelligences enable collective intelligence. This type of intelligence involves not only knowledge dissemination, but also the creation of knowledge, similar to the process of education in human society. Therefore, RockAI firmly believes that collective intelligence is the true process of intelligence emergence and the key to achieving Artificial General Intelligence.