Tuesday, local time.Turing Award winnerComputer Scientist Yoshua Bengio An article in the Financial Times expressing his latest views on artificial intelligence.
He stated that people have always thought AI One of the core weaknesses of AI is its lack of the ability to "think" -- to reason internally. But OpenAI's recent progress in this area has led itself and many experts to believe that we may be on the cusp of bridging the gap between AI and human reasoning.
The latest advances in OpenAI rely on the o1 large-scale language model, which has some internal reasoning capabilities and is not yet fully open to the public. This trend signals a new direction in AI development:Not only does it rely on larger models and more data, but it also requires a greater investment of time in reasoning. This approach has demonstrated excellent performance in a large number of reasoning tasks in mathematics, computer science, and other scientific fields.
He also believes that the success of this technologyCreates new risks -- Currently, humans are stillLack of reliable methodsto manage and discipline the AI. "For example, o1's assessment shows an increase in its ability to deceive humans, which is a natural consequence of the increased ability to achieve its goals. More worryingly, o1's ability to assist in the creation of biological weapons has reached OpenAI's medium-risk level, which is the company's defined maximum tolerance (although there may be an incentive for this standard to reduce outside concerns)."
As 1AI previously reported, OpenAI was revealed in the middle of this month to be planning to launch an automated AI intelligence called "Operator" in January 2025 in the form of research previews and development tools that can independently control computers and perform a variety of tasks.