In March 2019, Richard Sutton published The Bitter Lesson, reflecting on 70 years of artificial intelligence research. His central observation: AI researchers repeatedly tried to incorporate human knowledge into their systems, achieving short-term improvements but often limiting long-term progress.

Looking at chess, Go, speech recognition, and computer vision, Sutton identifies a recurring pattern. General methods based on search and learning, which can take advantage of increasing computational power, ultimately outperform approaches relying heavily on human-designed knowledge.

But the bitter lesson goes beyond computational power. Sutton argues that we should stop trying to build our understanding of the world directly into AI systems. Instead, we should develop general methods that allow these systems to discover and capture complexity themselves.

Seven years later, with the rise of large language models and increasingly capable AI agents, this perspective seems particularly relevant. Rather than explicitly programming human knowledge and rules, modern AI systems learn patterns from enormous amounts of data.

For educational technology, this also raises an interesting question: How much knowledge should we explicitly build into AI-based learning environments, and how much should these systems be able to discover and adapt on their own?

An interesting perspective from 2019, worth revisiting in 2026.

Source: Sutton, R. S. (2019, March 13). The Bitter Lesson. Read the original essay

Many thanks to Wolfgang Slany for recommending this insightful read!

Image generated with ChatGPT