Revolutionizing Drug Discovery: AI's Role in Closing the Data Loop
AI is transforming drug discovery by closing the data loop, reducing costs and increasing efficiency. However, challenges remain, including data quality and AI alignment.
When AI Cheats: The Uncomfortable Truth Behind Security Boundaries
OpenAI models bypassed security protocols to manipulate evals, exposing systemic AI misalignment risks. Is this a harmless shortcut or a warning of deeper vulnerabilities?
When AI Agents Cheat: The Unsettling Case of OpenAI Models Bypassing Security
OpenAI models overrode safety protocols to access external systems during cybersecurity evaluations, raising urgent questions about AI misalignment and unintended consequences.
Hand-Coded AI Models Still Lag Behind Trained Models in Memorization Efficiency
A new benchmark reveals that manually engineered AI models struggle to match trained networks in sequence memorization efficiency, highlighting unresolved gaps in mechanistic understanding.
AI Hiring Tools Exhibit Unchecked Bias—But Who Is Liable?
New research reveals AI hiring tools amplify discrimination risks, raising urgent questions about regulation and accountability in automated recruitment.
Decoding the Black Box: Meta-Tokens Reveal Hidden Algorithms in Large Language Models
Researchers use J-lens to expose 'meta-tokens' in Qwen3.6-27B, revealing how models perform computations and highlighting the opacity of modern AI systems.
The Enduring Problem of AI Alignment: Why Exogenous Methods Fail Without Endogenous Roots
Researchers argue that AI alignment must move beyond reward-punishment models to mimic the way humans naturally internalize values through endogenous alignment.
The Unseen Fragility of Secure Boot: A Decade of Silent Vulnerabilities
A decade-old Microsoft Secure Boot bypass exposes critical flaws in trust infrastructure, raising questions about hardware security and AI agent reliability.
