Today's large language models are deeply captive to human data, and the real breakthrough may lie in letting AI collect data through its own actions, the way babies do. This AI on Air clip is excerpted from a recent episode on the @Redpoint show. The speaker is Jürgen Schmidhuber @SchmidhuberAI, AI pioneer and co-inventor of LSTM. In the clip, he explains why training on human-generated web data is fundamentally limited. ▷ Every piece of data on the worldwide web exists only because at least one person, at some point, found it interesting. Models trained on it are therefore heavily biased toward human language, human-approved videos, and human behavior. ▷ His alternative is an "artificial scientist": an agent in an unknown environment that builds a world model by predicting the consequences of its own actions, then uses that model for planning, generating its own training data in the process. ▷ This mirrors how babies learn, not by downloading the web but by moving their fingers and watching what changes. He argues that artificial curiosity, an idea he proposed in 1990, points to the future: web data is a tiny fraction of what autonomous experiments could yield, and the resulting world models would be far less dependent on human language and far less human-biased.