Article contents:
- Introduction: From massive breakthroughs to growing concern
- Early indicators: GPT-5 and Gemini fall short of expectations
- Were “scaling laws” just a theory?
- The data wall: The main problem
- Synthetic data: A temporary fix or a new problem?
- Industry leaders’ response: “There is no wall”
- The next phase: The search for practical use cases
- Conclusion: A new race toward application, not invention
Introduction: From massive breakthroughs to growing concern
Over the past two years, generative artificial intelligence evolved so rapidly that major breakthroughs seemed inevitable. But in recent weeks, a growing anxiety has begun to creep into Silicon Valley: is progress starting to slow? Following a period of unbridled enthusiasm, the industry now faces difficult questions about whether the tremendous growth we have witnessed is sustainable.
Early indicators: GPT-5 and Gemini fall short of expectations
One early indicator is the lack of qualitative leaps between models released by major companies. Reports indicate that OpenAI is facing significantly less improvement in the quality of its upcoming GPT-5 model. Meanwhile, Anthropic has delayed the release of its powerful Opus model. Even at tech giant Google, an upcoming version of Gemini reportedly falls short of internal expectations. Dan Niles, founder of Niles Investment Management, says: “There was an enormous initial surge in what these new models could do, and what is happening now is that performance gains are beginning to plateau.”
Were “scaling laws” just a theory?
If progress has plateaued, it calls into question a core assumption that Silicon Valley treated as dogma: “scaling laws.” This idea posits that adding more computing power and more data infinitely guarantees better models. But recent developments suggest it may be more of a theory than a law. Simply increasing model size no longer appears sufficient to achieve massive leaps in performance.
The data wall: The main problem
The core problem may be that AI companies are running out of high-quality data to train their models, running straight into what experts call the “data wall.” These models have already consumed a massive portion of publicly available internet content. Now, there is no longer enough new text and imagery to feed their insatiable appetite.
Synthetic data: A temporary fix or a new problem?
To combat data scarcity, companies are turning to “synthetic data”—data generated by AI models themselves. But this is a temporary and hazardous solution. Alexander Wang, founder of Scale AI, says: “AI is a garbage-in, garbage-out industry. So if you feed these models a lot of AI-generated nonsense, the models will produce more AI-generated nonsense.” In other words, training AI on its own output could lead to degraded quality over time.
Industry leaders’ response: “There is no wall”
Of course, some industry leaders reject the notion that the rate of improvement has hit a dead end. Nvidia CEO Jensen Huang stated on the company’s latest earnings call, “Foundation model training scaling continues and is ongoing.” OpenAI CEO Sam Altman also posted simply on X: “There is no wall.” Nonetheless, their companies’ actions—such as delaying releases or encountering marginal improvements—may tell a different story.
The next phase: The search for practical use cases
If the AI revolution has exhausted its momentum in model development, the next phase of the race is the search for use cases—consumer applications that can be built on top of existing technology without requiring further radical improvements. For instance, the development and deployment of “AI agents”—autonomous software programs that can perform tasks on your behalf—is expected to be the next game-changer. Meta CEO Mark Zuckerberg says: “I think we are going to live in a world where there are hundreds of millions, if not billions, of AI agents.”
Conclusion: A new race toward application, not invention
The AI race is entering a new phase. The focus is no longer just on building larger and more powerful models, but has now shifted to how to use existing technology in useful and profitable ways. We may not see astonishing leaps like those witnessed with the emergence of ChatGPT, but innovation will continue in the form of practical applications that change our daily lives. Progress may not have slowed down; it is simply changing course.
Frequently asked questions
Q: What is meant by the “data wall”?
A: It is a term describing the point where AI models have consumed most of the high-quality data available on the internet, making it difficult to find new data to train larger and better models.
Q: What is “synthetic data” and what are its risks?
A: It is data generated by artificial intelligence itself to be used in training other models. The risk is that this data may contain errors or biases from the original model, leading to a degradation in the quality of new models over time.
Q: Does this mean AI development has stopped?
A: No, it does not mean a stop; rather, it may mean a slowdown in the pace of radical improvements in foundation models. Focus is likely to shift now from building larger models to finding practical and innovative applications for current technology.
