MoA Than One Curve: Teaching FFNs to Choose Their Nonlinearity
A mechanism-first reading of Mixture of Activations, and why token-adaptive nonlinearities may matter more than another round of parameter routing.
Cognaptus Insights
Industry thoughts, trends, and commentary.
MoE Than a Cost Trick: How Sparse Experts Became an Architecture Stack
A business-focused synthesis of three new MoE papers showing why sparse experts are becoming a design language for conversion, composition, and iterative computation—not merely a cheaper inference trick.
Pretty Text, Ugly Logic: When Image Models Learn to Write but Not to Reason
A comparison-based reading of why visually clear AI-generated text can still hide broken reasoning, and what that means for document, slide, and dashboard automation.
Right Answer, Wrong Audit: When Reasoning Models Grade the Destination, Not the Route
A mechanism-first reading of VAIR, a benchmark showing why correct answers can make large reasoning models unreliable auditors of flawed reasoning.
Safe Hands, Unsafe Audit: Why Robot Success Does Not Prove Robot Safety
A cross-layer reading of robotic manipulation safety, showing why task completion is not enough evidence for safe deployment.
Talk Is Cheap, Until It Trains ASR
A comparison-driven reading of how LLM-generated synthetic conversations can improve conversational ASR, and why the useful question is not more data, but better-matched data.
Curved Space, Straighter Retrieval: Why Graph RAG Needs Geometry
HyRAG shows that graph RAG failures may come less from weak retrieval and more from the wrong geometry for hierarchical knowledge.
Memory Lane, With Garbage Collection: What eMoT Gets Right About Reasoning Agents
A mechanism-first reading of eMoT, a reasoning framework that treats successful reasoning patterns as reusable procedural memory rather than disposable chain-of-thought text.
Mind the Slot: Jailbreak Prompts Have Weak Points, Not Just Bad Words
SlotGCG shows that LLM jailbreak risk is shaped not only by adversarial token content, but by where those tokens touch the prompt.
Pocket Experts: MobileMoE and the Memory Math of On-Device AI
MobileMoE shows that capable on-device AI is not just a smaller-model problem, but a routing, memory, quantization, and runtime-engineering problem.