Artificial intelligence companies have extensively collected internet data to train their models, including copyrighted music, as confirmed by Elon Musk at the DealBook Summit.

While Anthropic argues that training AI on copyrighted content qualifies as 'fair use,' major music industry players strongly oppose this position. The debate continues in courts, but a more pressing question has emerged: Is there any music left that hasn't been used for AI training?

According to former OpenAI Co-Founder and Chief Scientist Ilya Sutskever, speaking at NeurIPS in Vancouver, the answer is no. He states that pre-training AI on internet data will soon end because "We've achieved peak data and there'll be no more. We have to deal with the data that we have. There's only one internet."

The future of AI is shifting away from large language models (LLMs) toward what Sutskever calls an 'agentic' future - autonomous AI systems capable of performing tasks, making decisions, and interacting with software independently. Unlike current AI models that rely on pattern matching, future AI aims to develop reasoning capabilities similar to human thought processes.

Stanford's Fei-Fei Li adds that the next frontier is "the 3D ladder" or "spatial intelligence." She suggests that relying solely on 2D internet data is like building AI for a 'flat earth.' This indicates a fundamental shift in AI development: from models that simply regurgitate information to systems that can reason and understand in ways that more closely mirror human cognition.

Related Articles

Previous Articles