XNews

Scaling, Generalization, and Market Risks

Scaling, Generalization, and Market Risks

Several experts and reports indicate that continuing to scale computational power will enhance models, but there are risks related to generalization and training infrastructure; additionally, the adoption rate of AI is stabilizing in some enterprises.

Key insights

Continued Scaling Will Bring Improvements, but Key Capabilities Are Still Lacking

Ilya clarified that continuing to scale computational power and data will improve model and benchmark performance, but some important aspects are still missing, specifically referring to the model's generalization ability and efficient few-shot learning capability. This implies that scaling alone cannot solve all problems.

Risks Concentrated in Training Infrastructure and Underestimated Application Layer

Andrew Ng divides AI into three components: application layer, inference, and training. He believes the application layer is underestimated and presents significant opportunities; the supply on the inference side is limited; and the training side is capital-intensive and carries the highest risk. The rise of open-source models raises the training threshold, potentially weakening the training side's competitive edge and leading to investment return issues.

Signs of Stabilization or Stagnation in AI Adoption Rates Among Enterprises

A report from Apollo Academy cites census and Ramp data indicating that the speed of AI adoption among enterprises is not experiencing continuous exponential growth. There are signs of stabilization or even decline in adoption rates, especially among large enterprises, suggesting friction and bottlenecks in transitioning from experimentation to large-scale implementation.

Labor Replacement Due to Automation Is Occurring, but Unevenly Distributed

MIT research shows that AI has replaced approximately 11.7% of the U.S. workforce, primarily concentrated in clerical and information transmission roles. The study indicates that the labor structure is affected and needs to adapt to the widespread use of AI tools.

Discrepancy Between Benchmark Scores and Real-World Performance

It has been observed that models score high on benchmarks but perform inadequately in real-world scenarios. This may be due to overly targeted training or reinforcement learning optimization aimed at improving evaluation scores, leading to insufficient generalization ability and the phenomenon of 'high scores but low capability.'

Expert Users as Key Nodes for Dissemination and Best Use Case Discovery

The text points out that expert users provide the best use cases and play a leading role in community dissemination. An early development path centered around expert users can more robustly drive industry implementation, while directly targeting the general public may quickly lead to growth bottlenecks.

Source posts