A team from MIT has published research on weight-space programming, a method that lets AI systems work directly on the weights of neural networks rather than relying on human methods like writing code and tuning hyperparameters. The article frames this development through the Bitter Lesson, Richard Sutton’s 2019 empirical observation that “handmade” methods always, eventually lose against scalable methods in AI. It notes that every neural network, with large language models particularly big and particularly verbose, exemplifies the lesson: use more data, more compute, make models big and scale up algorithms, and success is mostly guaranteed to follow. The author argues that AI research should not be granted an exemption from this principle, predicting that frontier AI models will eventually reach into the machinery of intelligence itself and weave its numerical DNA to their will, starting an unprecedented era of recursive self-improvement without so much as a human inside the loop. The article states this is no longer fiction, citing the MIT team’s publication as the basis for that claim.