Gradient Customs: AlphaToken Checks Which Tokens Are Allowed to Train
AlphaToken reframes post-training as selective gradient routing, showing how token-level valuation can improve adaptation while reducing retention loss.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Share the Trunk, Spare the Averaging: Federated Actor-Critic Gets Personal
A mechanism-first reading of pFedAC: why federated reinforcement learning needs shared representations, local policy heads, and fewer fantasies about one global policy.
Small Moves, Big Models: The Quiet Discipline of Bounded AI
Two papers show why AI creates value in structured systems when it is scoped as a precise intervention, not promoted into an all-purpose replacement.
The Rule Is the Model: DEM’s Case for Bedside Anomaly Detection Without Explainer Theatre
A mechanism-first reading of DEM, a glass-box anomaly detector that turns residual distillation into an operational governance dial for physiological monitoring.
The Solver Isn’t the Strategy: FrontierOR’s Reality Check for AI Optimisation Agents
FrontierOR shows why runnable optimisation code is not the same as scalable algorithm design, and why enterprise AI agents need harder tests than solver demos.
Kitchen Confidential: FoodMonitor and the Compliance AI Reality Check
FoodMonitor shows why real compliance AI needs auditable evidence, not just video understanding with a rulebook attached.
Memory Foam: When AI Stops Storing Everything and Starts Learning From It
A business-focused reading of two new arXiv papers showing why long-horizon AI needs grounded abstraction, validated experience, and selective internalisation rather than ever-larger memory stores.
Mind the Readout: Why AI Gets Smarter When We Stop Worshipping the Output
A business-focused synthesis of three arXiv papers showing why AI reliability depends on representation, readout, and compute discipline—not just bigger outputs or heavier architectures.
Mixed Feelings: When LLM Batching Stops Being Obviously Better
A systems paper shows why mixed batching is not a universal default for LLM inference, and why bandwidth-aware scheduling may matter more than scheduler fashion.
Relight at Your Own Risk: WildRelight and the Synthetic Vision Reality Check
WildRelight shows why real-world relighting needs measurement infrastructure, not just prettier synthetic demos.