Former OpenAI Scientist: Computing Power has reached its limits, the AI industry must return to the core of research.

Former OpenAI Chief Scientist and current SSI co-founder Ilya Sutskever spoke in a recent interview about how current AI models are almost unbeatable in various tests, evaluations, and competitions, yet most people's daily lives have not been fundamentally changed. He believes that this gap of “superb performance in evaluation tests but ordinary experience” stems from the industry's over-reliance in recent years on the successful formula of “stacking computing power, stacking data, and stacking model parameters.” However, the natural language corpus available for pre-training is inherently limited, and this approach will inevitably encounter bottlenecks. The next phase of AI development will enter a new stage, where it will not be about who has more GPUs, but about who can find new learning principles, understand generalization, and make AI's learning methods closer to those of humans.

The sci-fi plot is unfolding, but life hasn't changed much.

At the beginning of the interview, the host described the current development of AI and the atmosphere in the San Francisco Bay Area as completely reminiscent of a science fiction novel. However, paradoxically, even though the global investment in AI amounts to tens of billions or even hundreds of billions of dollars, nearing 1% of the GDPs of various countries, the everyday experiences of the general public have not drastically changed.

Most people only see technology giants announcing further expansions of data centers in the news, or spending more budget to buy GPUs, but from the streets to the alleys, the noticeable changes in AI are still quite limited.

Sutskever accepted this statement and pointed out that the model's capabilities are indeed powerful, but in familiar living environments, the sensory experience is not as shocking as in laboratories or research reports; this gap is, in fact, a normal phenomenon.

The evaluation test ability is extremely strong, but the practical performance often has issues.

Sutskever then talked about the model's “dual nature.” He stated that AI often demonstrates a level of performance that exceeds human capabilities in program evaluations, standardized tests, or various benchmarks, but once applied to real-world demands, it encounters completely different situations. He gave an example where many developers ask the model to help fix bugs. The model usually confidently identifies the problem and offers a modification, but the next step often introduces new errors.

When you ask it to fix the second bug again, it may reintroduce the first bug, creating a loop between the two errors. Sutskever believes that this contradiction of being “superb in evaluations but unstable in the real world” is one of the most worthwhile AI phenomena to understand deeply at present.

To test training, causing the model to deviate from the real world.

When analyzing the reasons for the gap, Sutskever pointed out that major companies often adjust their model behavior based on public evaluation projects when doing reinforcement learning. Because as long as they perform well in evaluations, they can gain an advantage in launch events, investment presentations, and technical comparisons. This also makes the models incredibly powerful in those testing scenarios, but when faced with the ever-changing realities, they are unable to demonstrate the same capabilities.

He uses programming competitions as an analogy. If a contestant practices for 10,000 hours to win the championship, they can indeed achieve remarkable results in the competition. Another contestant who has only practiced for 100 hours but possesses a natural understanding of problems may actually be more adaptable in their long-term career. The model right now is like the former:

“Highly trained, extremely strong in fixed tasks, but lacking human-like deep generalization ability.”

The trend of model expansion is prevalent, and the research space has been significantly compressed.

Sutskever recalled that the period from 2012 to 2020 can be described as the “golden age of AI research,” with various sectors actively exploring different architectures and approaches, and many breakthroughs stemming from a variety of novel ideas. However, the success of GPT-3 and the law of model scaling has completely changed the direction. Since then, the entire industry has gradually formed a consensus:

“By enlarging the model, increasing the data, and adding computing power, the capabilities will naturally rise.”

This formulaic approach has become mainstream in the investment market and among major companies due to its low risk and predictable outcomes. However, because everyone adopts the same strategy, the actual space for research has been compressed.

Natural data is limited, and pre-training will eventually hit a wall.

Sutskever emphasized that the natural language data available online is limited, and training large language models relies on these texts. After years of expansion, various companies have nearly exhausted all available data “to the fullest.” When the amount of data can no longer double in growth, relying solely on the performance improvements brought by expanding the model and Computing Power will noticeably slow down.

He believes this represents that AI is about to enter the next stage, no longer just pursuing larger models, but rather re-understanding how humans learn, how generalization forms, whether models can self-correct with fewer samples, and whether they can have mid-course evaluation abilities like humans during the learning process, rather than solely relying on final feedback to determine the direction of behavior.

These problems cannot essentially be solved by the number of GPUs, but require real scientific research.

( Note: Pre-Training refers to allowing the model to read a large amount of online text to learn language structure and basic knowledge, forming the foundation for subsequent capabilities. Mainstream large models use pre-training as the first stage. )

Reinforcement learning has led to a computing power explosion, and efficiency is actually not as good as imagined.

Sutskever also mentioned that in recent years, many companies' reinforcement learning (RL) training scales have even surpassed pre-training. These long sequence inferences consume a large amount of Computing Power, but the effective learning brought by each inference is actually limited, resulting in a decrease in overall efficiency. If we continue to rely on the same training methods, we will only invest more resources without being able to break through the limits of the model's essence.

Therefore, he believes that the AI industry is gradually returning to the research stage of “exploring new methods,” with the focus no longer on who has the largest data center, but on who can find new learning principles.

( Note: Reinforcement Learning refers to a training method that allows the model to grow through trial and error. After completing a task, the model receives feedback or rewards and adjusts its behavior accordingly. )

Focus on understanding and learning, the business model can be discussed later.

Sutskever stated that SSI's strategy is to focus on research, particularly on understanding generalization, human-like learning methods, and how models can self-improve with a small number of demonstrations. He believes that rather than rushing to answer business models, it is better to concentrate on finding more fundamental learning architectures than pre-training. Once breakthroughs are made, various business applications will emerge one after another.

He estimates that in the next 5 to 20 years, there is a chance to create AI with “learning efficiency comparable to that of humans.” Once machines can quickly master new skills like humans, coupled with large-scale deployment, their overall capabilities will show exponential growth, approaching superintelligence.

The gap between evaluation and practicality will disappear with the emergence of new learning methods.

Sutskever finally stated that the current AI revolution seems less dramatic because there is still a significant gap between model capabilities and usage. When models evolve from being able to take exams to being able to learn, self-correct, continuously generalize, and grow steadily, AI will reshape global life at an extraordinarily fast pace.

By then, people will not only see massive investments in the news, but will also be able to truly feel the changes in their daily lives.

This article by former OpenAI scientist: Computing Power has reached its limit, and the AI industry must return to its research core. First appeared in Chain News ABMedia.