Elon Musk agrees that we’ve exhausted AI training data

Elon Musk concurs with other AI experts that there’s little real-world data left to train AI models on.

“We’ve now exhausted basically the cumulative sum of human knowledge … in AI training,” Musk said during a livestreamed conversation with Stagwell chairman Mark Penn on X late Wednesday. “That happened basically last year.”

Musk, who owns AI company xAI, echoed themes former OpenAI chief scientist Ilya Sutskever touched on at NeurIPS, the machine learning conference, during an address in December. Sutskever, who said the AI industry had reached what he called “peak data,” predicted a lack of training data will force a shift away from the way models are developed today.

Indeed, Musk suggested that synthetic data — data generated by AI models themselves — is the path forward. “The only way to supplement [real-world data] is with synthetic data, where the AI creates [training data],” he said. “With synthetic data … [AI] will sort of grade itself and go through this process of self-learning.”

Other companies, including tech giants like Microsoft, Meta, OpenAI, and Anthropic, are already using synthetic data to train flagship AI models. Gartner estimates 60% of the data used for AI and an­a­lyt­ics projects in 2024 were syn­thet­i­cally gen­er­ated.

Microsoft’s Phi-4, which was open sourced early Wednesday, was trained on synthetic data alongside real-world data. So were Google’s Gemma models. Anthropic used some synthetic data to develop one of its most performant systems, Claude 3.5 Sonnet. And Meta fine-tuned its most recent Llama series of models using AI-generated data.

Training on synthetic data has other advantages, like cost savings. AI startup Writer claims its Palmyra X 004 model, which was developed using almost entirely synthetic sources, cost just $700,000 to develop — compared to estimates of $4.6 million for a comparably sized OpenAI model.

But there as disadvantages as well. Some research suggests that synthetic data can lead to model collapse, where a model becomes less “creative” — and more biased — in its outputs, eventually seriously compromising its functionality. Because models create synthetic data, if the data used to train these models has biases and limitations, their outputs will be similarly tainted.