Former chief scientist and co-founder of OpenAI Ilya Sutzkever is confident that in the near future, developers of large language models will reconsider their approach to creating them. The main challenge, he believes, is the lack of new data, writes The Verge.
Currently, AI models are trained on information collected from the Internet and provided by partners. According to Sutzkever, the Internet, in this context, resembles a fossil fuel, such as oil, the resources of which are practically exhausted.
There will be no more data – we have reached the peak. We will have to deal with the data we have. There is only one Internet.
Most likely, Sutzkever estimates, the next generation models “will be agents in the real sense.” Although the researcher did not define them, agents are usually understood as autonomous AI systems that run on user systems and perform various tasks independently.
An example of such developments is the project of the company Anthropic, which has already presented the first version of its “agent”. OpenAI is expected to unveil its own similar system in early 2025.
Sutzkever himself, who left OpenAI in 2024, now heads his own startup, Safe Superintelligence (SSI). He will develop “safe superintelligence,” calling it “the most important technical problem of our time.”