The BABEL codec: the first complete, certified decode of everything happening inside a production language model (GPT-2 small). It reads the model's internal state into English AND writes English back into the model. 94.7% of behavior reconstructed — and that holds at every layer depth and text regime tested, not just one spot. Everything is open: paper, the full lexicon, the grammar tables, the decoder/encoder weights, reproduction scripts, and a demo that shows you the model's thoughts on any sentence you type.
[link] [comments]