Not Every Spike Is Positive: The MTJ Neuron Built for Signed Signals
A mechanism-first reading of an orthogonal-easy-axis MTJ neuron that implements signed LIF dynamics, with a useful device concept and very real commercialization boundaries.
Cognaptus Insights
Industry thoughts, trends, and commentary.
The Mole Is Not the Model: Dermoscopy AI Needs a Chain of Custody
A mechanism-first reading of a dermoscopic dataset methodology that shows why medical AI reliability starts before model training.
The Receipt Is in the Pixels: Model Attribution After the Watermark Fantasy
A paper on unified image generators shows that model identity can often be recovered from generated images, but not because the images carry an obvious watermark or reveal the prompt.
Don’t Miss the Bus: AlphaTransit and the Value of Learned Lookahead
A mechanism-first reading of AlphaTransit, showing why learned search matters more than simply drawing wider bus routes.
Local Fluency Is Not Local Fairness: IndoBias and the Indonesian Bias Problem
IndoBias shows why Indonesian-language AI fairness cannot be inferred from general benchmarks, model fluency, or local adaptation alone.
One Step, Not One Trick: SOM and the Q-Guided MeanFlow Policy
A mechanism-first analysis of how SOM turns critic gradients into one-step generative policies for online reinforcement learning.
Stop Signs Are Not Steering Wheels: TRIAD and the Case for Repairable Agent Guardrails
A business-focused analysis of why agent guardrails need to revise unsafe plans, not merely block them.
The Missing Ingredient Wasn’t Vision: NutriMLLM and the Data Recipe for Micronutrient AI
NutriMLLM shows that comprehensive nutrition AI needs specialised supervision, not just bigger vision-language models.
Think Before You Click: Test-Time AI Is the New Control Surface
Why inference-time adaptation is becoming both a performance lever and a governance risk.
You Can’t Reweight a Dead End: TRD and the Prefix Failure Problem
Trajectory-Refined Distillation shows why repairing failed reasoning paths may matter more than tuning token-level distillation losses.