Applications, News & Interviews

Create a clean editorial infographic about fair division of indivisible goods among groups. Center panel: a horizontal bar labeled "Total pool of goods" being divided into k uneven sections, each section representing a group (draw as simple icons like a chair, a truck, a laptop, a box distributed unevenly across labeled zones "Group 1", "Group 2", "Group 3", "Group k").

How Fair Can a Group Split Ever Really Be? A Tight Answer, at Last

A new result pins down, almost exactly, how badly an individual can be shortchanged when indivisible goods are divided fairly among groups rather than individuals — and gets there with a new trick for one-sided discrepancy.

A clean editorial hero infographic comparing two recommendation-system architectures, split into a left panel and right panel with a clear vertical divider.

Netflix Swaps Thousands of Features for a Language Model — and Its Ranker Gets Better

GenRec, Netflix's new LLM-backed recommendation ranker, beat a mature production system while training on a fraction of the labeled data — by trading feature engineering for context engineering.

A wide editorial hero illustration in a clean minimalist digital art style, depicting a growing cluster of glowing spherical cells rendered as translucent 3D particle clusters, expanding across a dark grid-like surface divided into several distinct colored zones representing GPU partitions.

Applications, News & Interviews

Teaching a Neural Network to Referee GPU Workloads in Growing Cell Simulations

A recurrent neural network that learns from simulated history, not real traces, keeps multi-GPU tissue-growth simulations balanced without the constant repartitioning that slows them down.

Clean editorial hero infographic in three horizontal panels on a light background with navy, slate gray, and one gold accent color.

Applications, News & Interviews

Fixing a Blind Spot in the Analytic Hierarchy Process: How Regularization Tames Unstable Priority Rankings

A new optimization model, ARDLS, patches a long-standing flaw in one of the most widely used decision-making frameworks — where the "best" priority ranking could depend on nothing more than a solver's starting guess.

Editorial hero illustration in a minimalist financial-editorial style: a lone investor figure standing at a fork where a smooth upward-sloping path splits into two divergent trails — one trail narrows sharply and drops into a shadowed steep ravine (representing a bad outcome), the other trail widens and rises gently into open light (representing a good outcome).

Applications, News & Interviews

When the Downside Is Capped: Rethinking Portfolio Rules Under a CVaR Limit

A new continuous-time analysis shows that capping expected losses in the tail doesn't just make investors more cautious across the board — it makes them cautious in an asymmetric, state-dependent way, and it comes with a provably convergent algorithm to compute the optimal policy.

Editorial hero illustration of a hospital/university laboratory scheduling scene: a wide corridor with three glass-walled lab rooms visible side by side, each labeled by subtle department signage (no text), technicians in white coats moving between rooms, one lab room glowing softly to show it is in active use while a matching lab type across the hall sits empty and dim, suggesting careful allocation of scarce staff.

Applications, News & Interviews

When Two Goals Can't Both Come First: Lexicographic Scheduling for Shared Lab Space

A worked MILP example shows how to book scarce laboratory rooms without splitting scarce staff across departments — by solving one objective completely before letting a second one even matter.

Applications, News & Interviews

Three Ways to Define "Risk" When Building a Portfolio — and How Optimization Picks the Best Mix

A hands-on AMPL notebook shows how the same optimization engine can build very different "safe" portfolios depending on how you define risk.