AI needs to be able to self-improve! More and more people in the AI community believe that "current AI training methods cannot break through."

A small but growing group of AI developers from companies like OpenAI and Google believe that the current technological path cannot achieve significant breakthroughs in fields such as biology and medicine, nor can it avoid simple errors. This viewpoint is raising questions in the industry about the direction of billions of dollars in investments.

According to The Information, at the Neural Information Processing Systems (NeurIPS) conference held in San Diego last week, many researchers discussed this topic. They believe that developers must create AI that can continuously acquire new capabilities after deployment, a "continuous learning" ability similar to human learning, which has not yet been realized in the AI field.

These doubts contrast with the optimistic predictions of some AI leaders. Anthropic CEO Dario Amodei stated last week that general artificial intelligence (AGI) could be achieved by scaling existing training techniques, while OpenAI CEO Sam Altman believes that AI will be able to self-improve in more than two years. However, if the skeptics are correct, this could put the billions of dollars that OpenAI and Anthropic are investing in technologies like reinforcement learning at risk next year.

Despite technological limitations, current AI performance in tasks such as writing, design, shopping, and data analysis continues to drive revenue growth. OpenAI expects its revenue to more than double this year to about $13 billion, while Anthropic anticipates its revenue will grow more than tenfold to about $4 billion.

Core Controversy: Can AI Learn Like Humans?

David Luan, head of Amazon's AI research department, clearly stated, "I can guarantee that the way we train models today will not last." Several researchers attending NeurIPS expressed similar views, believing that achieving human-like AI may require entirely new development techniques.

OpenAI co-founder and former chief scientist Ilya Sutskever stated last month that some of the most advanced AI training methods today cannot help models generalize, that is, handle a variety of tasks including those involving previously unseen topics. In the medical field, continuous learning could mean that ChatGPT can identify new types of tumors that do not exist in medical literature, rather than needing to be trained on a large number of precedents. This would enable it to perform like a human radiologist who can discover patterns based on a single case.

In a keynote speech at NeurIPS, Richard Sutton, a professor at the University of Alberta and known as the father of reinforcement learning, also stated that models should be able to learn from experience, and researchers should not try to enhance model knowledge through large amounts of specialized data created by human experts. He believes that when human experts reach the limits of their knowledge, AI's progress will "ultimately be hindered." Instead, researchers should focus on inventing AI that can learn from new information after handling real tasks.

Attempts at Technological Breakthroughs and Real-World Barriers

Several important research papers presented at NeurIPS explored this theme. **Researchers from the Massachusetts Institute of Technology and OpenAI proposed a new technology called "adaptive language models," which enables large models to acquire new knowledge or improve performance on new tasks by utilizing information encountered in the real world **

For example, when ChatGPT users request an analysis of previously unseen medical journal articles, the model may rewrite the article into a series of Q&A for self-training. The next time someone inquires about the topic, it can respond by integrating new information. Some researchers believe that this continuous self-updating is crucial for AI to achieve scientific breakthroughs, as it makes AI more like human scientists who can apply new information to old theories.

However, technological limitations have slowed down enterprise customers' procurement of new products like AI agents. The model continues to make mistakes on simple questions, and AI agents often perform poorly without significant efforts from AI providers to ensure their correct operation.

Business Impact: Revenue Growth and Investment Risks Coexist

If the views of skeptics like Luan and Sutskever are correct, this could cast doubt on the billions of dollars developers are expected to invest next year in popular technologies like reinforcement learning, including payments to companies like Scale AI, Surge AI, and Turing that assist with such work. Scale spokesperson Tom Channick disagrees, stating that using continuously learning AI still requires learning from human-generated data and the reinforcement learning products provided by Scale.

Nevertheless, even without new breakthroughs, AI developers seem capable of generating substantial revenue. OpenAI and Anthropic had almost no revenue three years ago but now generate considerable income from chatbot and AI model sales. Other startups developing AI applications, such as coding assistant Cursor, are expected to collectively generate over $3 billion in sales in the coming year.

Industry Competition: Google's Lead Sparks Turmoil

Researchers also discussed the AI race among major developers. Google's technology has surpassed competitors in certain metrics, and Altman has informed OpenAI to prepare for a "tough atmosphere" and "temporary economic headwinds."

During a Q&A session with the Google AI team, several attendees asked how the team improved the pre-training process—an issue OpenAI has been working to resolve for much of this year. Google Research Vice President Vahab Mirrokni stated that the company has improved the data combinations used for pre-training and found better ways to manage thousands of Google-designed tensor processing units, thereby reducing hardware failures' interference with the model development process.

OpenAI leadership recently indicated that they have similarly improved the pre-training process, developing a new model codenamed Garlic, and believe they can compete with Google in the coming months