Openai co -founder Ilya Sutskever, who created an artificial intelligence laboratory safe superinTelling inc., for the first time in a long time publicly at the conference of Neurips in Vancouver. During his speech, Sutkever made a number of resonant statements about the future development of artificial intelligence technologies.
Sutskeever said that the traditional methodology for preliminary training in AI models, when they study at the huge volumes of nemarkized data, is coming to an end. According to him, the volume of human content on the Internet, which can be used for training, is almost exhausted. This comparison was reinforced by an analogy with the exhaustion of fossil resources: “As oil is a final resource and the volume of content on the Internet is limited. We have only one Internet.” Sutkever emphasized that the industry will have to look for new approaches to training models.
Speaking of the prospects, Sutskever noted that the next generation of models will be “agent” and will be able to make decisions, perform tasks and interact with the software autonomously. In addition, the Sutzkever predicts that future AI systems will gain the ability to reason, which will make their work more similar to the process of human thinking. This will be a step forward compared to current models that are mainly based on a template comparison of information.
The Sutskever compared the unpredictability of such “reasoning” systems with the behavior of modern chess programs that are at a dead end of even the best human players. He also emphasized that such systems will be able to effectively work with limited data and avoid confusion in difficult tasks.
One of the key ideas of the performance was the parallel between the scaling of AI and evolutionary biology. Sutkever cited research data showing the dependence between the mass of the brain and body in various species of animals, and noted that hominids have a unique development trajectory compared to other mammals. Sutkever suggested that AI can find similar new ways of scaling, going beyond the traditional approach to learning.
After the performance, Sutkever answered questions from the audience. One of the participants asked how scientists can create incentives for the development of AI that can share people’s freedom. Sutkever recognized the complexity of the issue, noting that this may require a structural approach to the government level.
The idea of using cryptocurrency as a tool for such mechanisms also sounded. Sutkever did not exclude such an opportunity, adding that the creation of AI, who wants to coexist with people and have its own rights, could be the ultimate goal. Concluding, Ilya noted that the future is extremely unpredictable, but such reasoning and speculation are important for further research.