At the GenAI Summit 2024, held on August 18th in Ho Chi Minh City, Dr. Le Viet Quoc, dubbed the "AI prodigy" at Google, offered his insights on the evolution of AI. International Mathematical Olympiad This shows that machines have begun to learn to reason.
“"Don't underestimate the reasoning abilities of AI in the next few years. If you look deeper into mathematical datasets, there are many questions we can't solve, but AI can," he said.
The Google Brain co-founder said that when he received the math solution results... AI, He and his colleagues reviewed the data to see if there were any instructions or documents that the AI could learn from and use to develop solutions, but they found none. This shows that the AI has begun to reason independently like a human.
“"At some point, we really won't be able to distinguish between machine reasoning and human reasoning," he said, assessing AI's current capabilities at 3 out of 10—not yet outstanding, but promising "unimaginable" advancements in 10 years.
Based on practical experience, Dr. Thang Luong, a senior researcher at Google DeepMind, points out that there isn't always enough data to verify or analyze AI, forcing them to rely on their own reasoning. He argues that for complex problems like solving unsolved puzzles, GenAI It will use a logical reasoning method. Each step will be analyzed by the AI as either correct or incorrect. If correct, it will continue; if incorrect, it will go back and try another approach, thereby finding the formula to arrive at the answer.
“"If superintelligence involves too much reasoning, it will create too many steps, which can easily lead to..." illusion. "When data is unavailable, AI will take a roundabout approach. In AI problem-solving, it's very difficult to determine what is an illusion and what is AI reasoning, especially with classic problems that have yet to be solved," Mr. Thang said.
When asked how to reduce the effects of AI hallucinations, Jeff Dean, Gemini, Google's Chief Scientific Officer and co-founder of Google Brain and Google Translate, argues that current GenAI is largely trained using a "guess and match" model. This word-guessing "game" might produce results that fit one context but not another. When developers mix all the data together for training, the information becomes fragmented, leading to the AI creating an illusion.
“"If the model understands the context of words, it can predict the next word more accurately. In other words, the information being input and decoded must be placed within a specific context instead of being mixed up," Dean said, adding that while this won't eliminate AI hallucinations, it will be significantly effective.
Meanwhile, Dr. Le Viet Quoc proposed another approach: placing AI responses into search engines. To counter AI illusions, the model must first be able to detect what constitutes an "illusion." For example, with a GenAI response, developers can "break it down" into multiple statements, perform reverse queries to check, and thus identify which data in the response is incorrect. Through this, the models can provide users with additional information about "which data sections are illusory." This method has been adopted by several developers and has proven effective.
At the conference, Deputy Minister of Planning and Investment Nguyen Thi Bich Ngoc cited Google's forecast that by 2030, the annual economic impact of digital technology in Vietnam could reach VND 1,733 trillion (USD 74 billion), with a significant contribution from AI. She also cited research by Thundermark Capital showing that Vietnam and Singapore are two Southeast Asian representatives among the world's top 30 in AI research.
According to VNExpress
Contact EZ Fast Tech today for a free consultation and to get an effective, AI-optimized website!
- Website: https://ezfasttech.com
- Hotline: 0772869820
- E-mail: [email protected]
EZ Fast Tech – Take your website to the next level with the power of AI!
