Morning Edition · Friday, June 26, 2026
New Interpretability Work Locates Sycophancy and Refusal in the Same Activation Geometry
Two papers argue that misbehaviors engineers try to suppress are governed by linear directions in the residual stream, and that persona and refusal are not independent.
Two new papers extend a line of research that treats model behaviors as adjustable directions in activation space. The first, Detecting and Controlling Sycophancy with Cascading Linear Features, addresses a practical limitation of activatio…
Continue the AI Intelligence Brief
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
- 5 AI intelligence signals a day
- Frontier labs, compute, and chips
- Model releases and AI infrastructure
- Source-grounded analysis with confidence labels
The Global Intelligence Brief stays free.
Part of a tracked trend
Mechanistic Interpretability of Transformer Internals
Research increasingly characterizes the internal structure of trained transformers — outlier 'massive activations,' residual-stream features — as load-bearing architectural phenomena rather than artifacts, with recurring downstream consequences for quantization, compression, and interpretability tooling.
More from this edition
- US-Led Pax Silica Bloc Widens to 24 States as AI Supply Chains Split Along Geopolitical Lines
- Agents Become the First Real Economic Test of the AI Buildout
- IBM Discloses 0.7-Nanometer Transistor Process Using Vertically Stacked Devices
- MindOn Shows a Single Model Controlling Robots of Different Body Types
- Unitree Prices a Humanoid Robot From 4,900 Dollars
- Papers Ask What to Do After a Benchmark Saturates and How to Tell Knowing From Guessing
- A New Benchmark Tests Whether AI Agents Can Actually Secure Smart Contracts
- Study Finds Standard Post-Training Can Erode Values Instilled During Pretraining
- Anthropic Pushes Policy Proposals Built Around an AI Capability Curve It Calls Exponential
- Survey Data Show Gen Z Both the Heaviest AI Users and the Most Skeptical
- Researchers Propose Turning Editing Models' Own Moderation Filters Into a Defense