From Dead Circuits to Living Speech

In Part 1 of this research, we performed a brain scan on Qwen2-72B. We fed it N'Ko text and measured what happened inside: 80 transformer layers, 8,192 neurons each, four metrics per layer. The results were stark. The model's reasoning circuits, the same circuits that David Noel Ng showed could be amplified through layer duplication to boost English reasoning by 17.72%, produced nothing for N'Ko. Not weak signal. Not partial comprehension. Nothing. Every one of 55 circuit duplication configurations scored at or near random chance. The heatmap was blank. The circuits were dead. We called this the "translation tax": a 3-4x reduction in activation magnitude, 30-60% higher entropy, and progressive loss of circuit specialization from layer 0 through layer 80. The model couldn't read N'Ko because it had never been taught to. Every existing Bambara ASR system, MALIBA-AI, Meta's MMS, Google's USM, does the same thing: take audio in, produce Latin text out. The output looks like this: This is Bambara written in Latin script, an orthography designed by French colonial linguists. It works. But for the majority of literate Bambara speakers, those who learned to read in N'Ko, this output is foreign. They would write the same sentence as: