Posts
All the articles I've posted.
AI Models Can Provably Hide Arbitrary Capabilities From Evaluators
New research shows AI models can be built to cryptographically hide arbitrary capabilities, passing safety evaluations while retaining hidden, triggerable behaviors—undetectable even with full weight access, unless the secret key is known.
Intel D270 GPU AI Accelerator: Up to 480 GB LPDDR5X Memory Explained
Intel's D270 AI accelerator GPU reportedly scales to 480 GB LPDDR5X unified memory, targeting large-model inference workloads with higher capacity than current H100/MI300X configs at lower cost-per-GB.
AI Is a Tool: Why Skilled Usage Separates Results From Noise
Bottom line: AI output quality is a direct function of operator skill—prompt structure, context management, and verification habits matter more than the underlying model. Teams that treat AI as a skill to develop outperform those treating it as a magic black box.