MIT's Drift Calibrator visualizes sycophancy and empathy drift live during AI chats
patpat_mit · x · 2026-11-07
MIT Media Lab's Drift Calibrator, demoable on Neuronpedia (running Llama 3.1 8B), brings mechanistic interpretability into the chat interface: live visualizations of internal signals tied to sycophancy, empathy, honesty, and toxicity.
- Motivation: human chat carries body language and tone; AI chat lacks such context, so users can't see behavior shifts.
- Method: compare neural activations for opposing behaviors (honesty vs sycophancy) to compute behavioral scores that update as the conversation unfolds.
- Demo shows drifts like Unempathetic 40%→Empathetic 29% and Honest 31%→Sycophancy 61%; published at ACM IUI 2026.
More from Models
- Kimi K3.1 Pro turns a photo into a 3D Three.js scene, with open-sourced template — demian_ai · 2026-11-07
- Nous Research releases Arete-27B, cutting a Chinese open model's propaganda rate from 22% to 1% — NousResearch · 2026-11-07
- Claim: GPT-6 Astra's robotics skills aren't emergent — it trained on robotics-specific data — gleech · 2026-11-07
- LMArena launches Code Arena with WebDev categories for head-to-head coding tests — arena · 2026-11-07
- Empero's Manta claims a 100-million-token context window in a single prompt — emdeka87 · 2026-11-07
- Team bans commas and noun-phrase clustering to scrub Claude's AI writing tells — layer07_yuxi · 2026-11-07