Trend Tuesday: Turns Out, Reality Has Better Training Data

Much has been said about AI running out of text - “the internet has been read.” Maybe true, and irrelevant. Because the AI actually learns from patterns.

When AlphaZero learned chess, it didn’t study grandmaster books. It played millions of games against itself, and found moves which humans did not know. Self-driving systems improve through millions of kilometers of road experience. Robots learn from the physical constraints that the real world imposes.

An especially interesting case is Delphi-2M, a GPT-style model trained on the lifetime medical trajectories of over 400 000 people. Each record (age, diagnosis, lifestyle) becomes a token in a sequence. From this, Delphi-2M predicts what disease might occur next, and when.

The broader trend is clear: AI is shifting from learning about the world through language. Instead, it can learn directly from patterns. Text-based models mirror what humans already know. Pattern-based models learn from sequences of actions, measurements, and events - discovering structure in disease progression, traffic flow, market dynamics, or the laws of physics.

Even if every sentence has been read, learning continues. Because the world itself remains an open book.

What other domains do you think AI should “read” next?

• Shmatko, A., Jung, A. W., Gaurav, K. et al., Nature (2025), Learning the natural history of human disease with generative transformers, https://lnkd.in/exmkaFza

• Silver, D., Schrittwieser, J., Simonyan, K. et al., Science (2018), Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (AlphaZero), https://lnkd.in/ePiMP2Ry

• Lisa Schut, Nenad Tomasev et. al., Bridging the Human-AI Knowledge Gap: Concept Discovery and Transfer in AlphaZero, https://lnkd.in/eqest5DV