Reverse Engineering: Bard's Identity Found Dormant Inside Google's Gemma 4
dejanseo · x · 2026-08-08
A developer conducted a reverse engineering study on Google's Gemma 4 using mechanistic interpretability methods, surprisingly discovering that the older identity "Bard" remains dormant inside the model.
- Phenomenon: By ablating a few thousand specific MLP neurons in one layer, the model's answer to "What is your name?" reverts from "Gemma 4" back to "Bard."
- Underlying Mechanism: The model stores several candidate identities at different strengths, with the strongest one winning at decode time. Partially removing the evidence for the current identity surfaces the old one.
- Engineering Pipeline: The author built a drill-down pipeline using PyTorch hooks to intervene (zeroing out specific components) during inference, tracing behaviors from chat responses down to individual weight vectors. The article also details two failed and two successful intervention methods.
More from Research
- Simplex Diffusion Models author Justin Deschenaux to present his paper Monday — LucaAmb · 2026-10-16
- Tool auto-audits paper claims with pass/fail evals — 4 of 7 metrics reproduce, 1 fails — jwt0625 · 2026-10-16
- Hironobu Fujiyoshi publishes deck: from AI agents to Agentic Robotics — HirokatuKataoka · 2026-10-16
- Is math just a language? A debate on why LLMs are good at mathematics — LucaAmb · 2026-10-16
- Alibaba open-sources OpenGCUA, a full-stack generalist computer-use agent; +19.6 points on OSWorld-v2 across 8 models — Tongyi-MAI · 2026-10-16
- CoDA co-evolves simulation and robot policies via shared failure memory, lifting OOD success from 5.3% to 27.6% — PLAN-Lab · 2026-10-16