KI und die schwierigen Probleme der Welt · Theoretische Informatik · Anthropic · 2026-09-28
Auf dem Weg zur Monosemantizität
Originaltitel: Towards monosemanticity
Individual neurons in neural networks often carry several meanings at once, which makes them hard to interpret. Using dictionary learning, the researchers decomposed a transformer layer of 512 neurons into more than 4,000 features, each corresponding to a pattern of neuron activations. The features separately represent things such as DNA sequences, legal language, HTTP requests, Hebrew text, and nutrition statements, properties that are not visible when looking at neurons one at a time.