Summary of Key Points
This article discusses the current state, challenges, future directions of AI, and the evolution of industry narratives: AI is essentially an advanced statistical model that has limited effectiveness in replacing tasks such as customer service, but it shows significant improvements in areas like drug design. The main bottleneck is the depletion of available data, which needs to be addressed through high-quality data filtering and the creation of synthetic data (a process akin to "AI distillation"). AI makes rapid progress in structured fields like mathematics and programming, yet it struggles in softer disciplines such as social sciences and the physical world. The industry narrative has shifted from emphasizing "superintelligence" to focusing on "AI agents" that can perform specific tasks, replacing white-collar jobs. However, for AI to become a transformative technology, it must break beyond mere improvements and offer something truly novel; otherwise, its added value will be limited.
1. AI is Not “True Intelligence”; It’s More Like an “Advanced Repetitor”
Conservative views suggest that current AI is merely an advanced statistical model—a machine that memorizes large amounts of text and images to predict answers based on statistical patterns, leading to errors (such as providing irrelevant information or fabricating facts), thus requiring human review. Its strengths lie in language, image processing, and code analysis, but it still falls short of human-like "thinking processes." For example, large language models (LLMs) are more like textsemblers that require lengthy prompts to complete complex tasks and have difficulty correcting errors.
A prime example is customer service, which was once thought to be easily replaceable by AI. However, users quickly realize the robots lack empathy and prefer human interaction. HR professionals also point out that AI cannot truly understand human emotions and complex needs.
2. The “Data Crisis” Facing AI
AI training requires a vast amount of data, similar to how humans need food to learn. The natural data available on the internet is gradually being depleted, prompting companies to use synthetic data generated by AI for retraining—although this can lead to repetitive errors and logical inconsistencies if the initial data is flawed.
To overcome this issue, two approaches are needed: (1) selecting high-quality data (e.g., from published books, academic papers, and reputable forums) and filtering out junk information; (2) using AI to create new data through techniques like "distillation," where a sophisticated "teacher AI" generates data and reasoning results for a smaller, more efficient "student AI" (such as ChatGPT’s smaller models). The challenge is that the quality of the "teacher AI" must first be improved for the student AI to learn something new.
3. AI’s Imbalance: Strong in Structured Fields, Weak in Soft Areas
AI’s progress is uneven, with clear strengths and weaknesses:
- Strengths: Structured fields (mathematics, programming, logic) have well-defined rules and standard answers, allowing AI to self-verify (e.g., using compilers to check code). AI can improve significantly in these areas without extensive human guidance, as seen with AlphaGo.
- Weaknesses: Soft fields (social sciences, public policy, business strategy) lack standard answers and require understanding of human behavior and complex contexts. The physical world also presents challenges, as collecting and verifying data for tasks like robot movement or household chores is costly and time-consuming.
4. A Sharp Change in Industry Narratives: From “Super AI” to “AI Agents”
In the past two years, companies like OpenAI touted concepts like superintelligence and general-purpose AI that could perform everything. However, these claims have faded, indicating a lack of confidence in such technologies. The current focus is on "AI agents" that can perform specific tasks (e.g., writing reports, organizing data, scheduling appointments), targeting the massive white-collar workforce worth trillions of dollars annually.
This shift is akin to replacing low-level jobs with automated machinery (e.g., robotic arms in factories). Whether this will succeed is crucial for AI’s practical application; only by replacing a large number of white-collar jobs can its value be fully realized.
5. For AI to Be a Transformative Technology, It Must Offer Something New
The U.S. government and tech companies describe AI as a transformative force, similar to electricity or the internet. However, historical transformative technologies have not been mere improvements but have introduced entirely new solutions (e.g., electricity led to the invention of light bulbs and telephones). Current AI applications (chatbots, image generators) are not enough to create significant value. If AI cannot break through these limitations, the investments made by governments and companies will face risks when the initial excitement fades. What is expected is a technology that truly changes the world, not just an improved version of existing tools.
In summary, while AI has great potential, it is not yet at the stage of revolutionizing everything. The challenges in data availability and understanding the real world, along with the shift in industry narratives, indicate that there is still a long way to go before AI becomes a truly transformative technology.