Summary of Key Points
The departure of Jeff Dean, the soul of Google's technology, to found the company AI for Science with Discovery Loop marks AI for Science as the next focal point of competition among global AI giants. In the past, AI was primarily a “question-answering machine,” but it is now evolving into a “research assistant” that can propose scientific hypotheses and design experiments, essentially becoming a “quasi-scientist.” This development not only reflects Google's choice of technological direction but also triggers profound discussions about the future of AI technology (the combination of symbolicism and connectionism), whether AI can generate new knowledge, and whether scientific discovery represents the final milestone towards Artificial General Intelligence (AGI).
1. Jeff Dean’s Departure: Google’s “Short-Term Interests” vs. AI for Science’s “Long-Term Appeal”
Who is Jeff Dean? Simply put, he was the key figure behind the development of Google Search’s core algorithms and later created TensorFlow, a tool used by researchers worldwide, akin to the fundamental infrastructure of the AI field. His departure stems from internal conflicts within Google regarding resource allocation. Google currently prioritizes profitable businesses such as advertising and cloud services, while AI for Science represents a “slow-burning” investment with long return periods (for example, drug development can take up to 10 years). Since Jeff Dean wants to focus on this area, he had no choice but to start his own company. This indicates that AI for Science has moved from being a laboratory concept to a field worthy of significant investment by industry leaders.
2. AI for Science: A New Arena for Giants – Who Can Make AI a “Scientist”?
Global AI giants are all vying for this space:
- Google DeepMind: Has already used AlphaFold to predict protein structures (what scientists used to take years to accomplish) and is now working on using AI to design new drugs.
- OpenAI: Collaborating with Microsoft, they utilize GPT-4 along with other tools to assist scientists in literature research and experiment design.
- Chinese Players: Baidu’s Wenxin Yiyuan is experimenting with large models to support scientific research, while Alibaba’s DAMO Academy is developing AI for weather forecasting.
Why are they all so interested? Because AI for Science can address critical challenges in scientific research, such as high costs of experimental trials and the overwhelming amount of data that needs analysis. AI has the potential to increase efficiency by ten to hundred times and could lead to the creation of new drugs, materials, and energy sources, representing markets worth hundreds of billions.
3. The Evolution of AI: From “Question Answering” to “Research Conducting”
Earlier versions of AI, like ChatGPT, could only answer questions or solve problems. However, modern AI systems like AI for Science can do two more advanced things:
1. Propose Hypotheses: For example, by analyzing large amounts of cancer data, AI might suggest that a certain genetic mutation could be linked to lung cancer.
2. Design Experiments: They can advise scientists on the steps and required instruments needed to test these hypotheses.
This represents a significant leap, as AI moves from being a student to an assistant, and potentially even a researcher in the future. For instance, DeepMind’s AlphaFold2 can not only predict protein structures but also design new proteins (such as enzymes that could degrade plastics), marking the beginning of AI’s ability to create novel solutions.
4. The Combination of Technologies: The Key to AI in Research
Let’s explain two technical terms in simple language:
- Connectionism: Refers to modern large models like GPT, which learn patterns by being fed data, similar to how humans draw conclusions from experience.
- Symbolism: Represents the earlier approach of writing rules (e.g., “If A, then B”) for logical reasoning.
Previously, these approaches were separate, but AI for Science requires a combination of both. For example, large models process vast amounts of research data (such as papers and experimental results), and then symbolic logic is used to explain why certain outcomes occur and what experiments should be conducted next. This approach ensures that AI’s conclusions are both data-driven and logically sound.
5. The Ultimate Question: Can AI Become a “Scientist”? Is Scientific Discovery the Final Step towards AGI?
The most contentious question is whether AI can generate entirely new knowledge. Scientists like Newton and Einstein made groundbreaking discoveries from scratch. Currently, AI can only combine existing knowledge, but it lacks the ability to develop intuition and creativity. However, scientific discovery could be the final step towards AGI. If AI can propose novel theories and concepts on its own, it would be a significant milestone toward achieving true artificial general intelligence. Jeff Dean and other industry leaders are working towards this goal, and major breakthroughs could occur in the next 5 to 10 years.
Conclusion
Jeff Dean’s departure is not an isolated event; it signals a shift in the AI industry from consumer-oriented applications (such as chatbots and video recommendations) to research-driven applications. AI for Science has the potential to transform how scientific research is conducted and may redefine innovation. In the future, collaboration between scientists and AI will become the norm, with AI potentially taking on a central role in scientific discoveries. The changes we can expect include faster drug development, breakthroughs in new energy technologies, and more precise disease treatments—all of which are closely related to the progress of AI for Science.