Encodes branded items, manufactured goods, and products.
What It Does
Identity.Product neurons activate on named products, brands, and manufactured goods: technology products (iPhone, Windows, ChatGPT), vehicles (Tesla, Boeing 747), consumer goods (Coca-Cola, LEGO), and other branded items that have distinct identities as products. They encode the product as a specific named thing with a version, manufacturer, and use-context.
How It Behaves
Product neurons concentrate in the middle-to-late layers, where brand identity is fully resolved. They interact with Organization neurons (linking products to their manufacturers) and Time neurons (product release dates, version histories). An interesting property: Product neurons show higher activation on newer products in models trained on more recent data, reflecting the recency bias in training corpora — recently released products are overrepresented in internet text.
Research Example
In Llama 3.1 8B, Identity.Product neurons distinguish 'iPhone 15' (specific versioned product) from 'a smartphone' (generic category) — the branded proper noun triggers Product neuron firing while the generic category does not. This distinction is important for tasks involving product recommendations or technical specifications, where confusion between branded and generic references causes output errors.