Recently resurfaced on Hacker News, this 2021 paper from Anthropic's interpretability team remains a key reference in mechanistic interpretability research. The authors lay out a mathematical framework describing transformers as sequences of linear operations and attention mechanisms that can be decomposed and studied individually, rather than treated as opaque black boxes.
The central contribution centers on the idea of the residual stream as a shared communication channel across layers, where each attention block and each layer reads from and writes to this stream. This perspective helped identify structures such as QK (query-key) and OV (output-value) circuits, along with the phenomenon of induction heads, which reproduce patterns already seen earlier in context. These findings have since underpinned much of the subsequent work on interpreting large language models.
The paper has had an outsized influence relative to its technical density, offering a vocabulary and set of conceptual tools that researchers outside Anthropic have since adopted, notably to study phenomena like in-context learning and emergent model capabilities.
Its reappearance on forums such as Hacker News, years after original publication, reflects the staying power of certain foundational works in a field that otherwise moves quickly. It also signals renewed interest in mechanistic approaches as language models grow larger and less transparent, making the need to understand their internal workings, rather than simply observing outputs, increasingly urgent.