Fred Wilson

True Human-AI Synergy Is Rare: A Six-Dimensional Design Space to Close the Gap

Despite widespread AI deployment in high-stakes domains, genuine human-AI synergy — where combined performance exceeds either party alone — remains uncommon. The authors term this persistent shortfall the "synergy gap" and argue it persists because current work narrowly focuses on interpretability, trust calibration, and interface design. They propose a structured six-element design space (sociotechnical context, decision-making frameworks, human decision participants, AI capabilities, interaction, and holistic evaluation) as a shared framework for practitioners and researchers to systematically address the conditions under which human-AI combination actually works.

Calcium-Omics Features Outperform Agatston Score Alone in ML-Based Myocardial Ischemia Prediction from Non-Contrast CT

A machine learning framework trained on 987 patients demonstrates that augmenting standard Agatston scoring with "calcium-omics" features — quantitative descriptors extracted from non-contrast CT calcium scoring (CTCS) scans — significantly improves myocardial ischemia prediction over clinical variables or Agatston score alone. The final XGBoost model, with SHAP-guided feature selection, achieved 98.9% precision and 79.2% sensitivity using only 10 features (Agatston score, 8 calcium-omics features, and age). Notably, a structural tension emerged between feature importance methods: the number of calcified arteries ranked lowest by SHAP yet showed the strongest univariate association with ischemia via logistic regression (OR: 3.63), highlighting the interpretive complexity of ensemble vs. marginal feature attribution.

CatBoost + SHAP on CT Calcium Scans Predicts Obstructive CAD Without Contrast or Invasion

A CatBoost gradient boosting model trained on calcium-omics and epicardial fat-omics features extracted from non-contrast CT calcium scoring (CTCS) scans achieves strong predictive performance for obstructive coronary artery disease (CAD), with 83.1% sensitivity, 93.8% specificity, and 85.3% accuracy on 1,324 SCOT-HEART trial patients. SHAP-driven feature selection reduced 424 candidate features to 14, with the top two being fat-omics derived — highlighting the underutilized diagnostic signal in epicardial adipose tissue. Critically, the model retains predictive accuracy even in zero-calcium-score patients, a known blind spot of conventional CTCS risk stratification. This approach could reduce reliance on contrast-enhanced CTA or invasive angiography in low-to-intermediate risk populations.

First Explicit Multi-Monopole Constructions on Non-Product 3-Manifolds via Adiabatic Limit Theorem

The multi-spinor generalization of the Seiberg-Witten equations on 3-manifolds — known as multi-monopoles — fails to yield a topological invariant because solution counts jump discontinuously across parameter-space boundaries called chambers (wall-crossing). Existing constructions were too limited to exhibit this phenomenon concretely. Wilson resolves this by proving an adiabatic limit theorem on mapping tori, constructing multi-monopoles by perturbing fixed points of the monodromy map on multi-vortex moduli spaces. This yields the first explicit multi-monopole solutions on non-product 3-manifolds across multiple chambers, providing a concrete testing ground for wall-crossing phenomena tied to multi-valued harmonic spinors and higher-dimensional gauge theory.

A study on Type-2 isomorphic circulant graphs. PART 9: Computer programs to show Type-1 $\&$ -2 isomorphic circulant graphs

TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation

An Ultra-Short Period Super-Earth and a Sub-Neptune Orbiting the K dwarf TOI-4311

Fred Wilson and Marco Mignano Discuss the Arc of NYC Startups and VC in Recorded Walking Conversation

Fred Wilson, veteran VC and co-founder of Union Square Ventures, recorded a walking conversation through the Union Square neighborhood with new partner Marco Mignano. The discussion spans the historical evolution, current state, and future trajectory of startups, venture capital, and the New York City tech ecosystem. The format — an ambient, location-grounded dialogue — suggests an intent to make the conversation accessible and public-facing rather than institutional.

Fred Wilson on 40 Years in Venture — and Why USV Is Automating Itself

Sentinel-Strategist Architecture Minimizes RAG Security-Utility Trade-off via Adaptive Defenses

The Sentinel-Strategist architecture enables context-aware defense orchestration in RAG systems by detecting anomalous retrieval via a Sentinel and selectively activating defenses through a Strategist. This approach counters multi-vector attacks like membership inference and data poisoning while preserving retrieval utility. Experiments across benchmarks show it eliminates membership inference leakage, reduces poisoning success near zero, and recovers contextual recall to 75-100% of undefended baselines versus 60% for static stacks.

ADMM Solves Bilinear Minimax Problems via Exact Generalized Projections

For bilinear minimax problems max_{c ∈ C} min_{β ∈ S} c^T A β over compact convex sets, standard ADMM's proximal operator is intractable. The paper proves this proximal reduces exactly to a generalized projection onto S, without approximations. The resulting algorithm alternates Euclidean projection onto C and generalized projection onto S, with established convergence.

Class-Sized Logic Fragments Characterize Large Cardinals via Model-Theoretic Properties

Restricted fragments of class-sized L_{∞∞}, second-order, and sort logics regain compactness and Löwenheim-Skolem properties by banning certain infinitary quantifier-boolean combinations. These properties characterize Weak Vopěnka's Principle and Ord is Woodin via downwards Löwenheim-Skolem, Shelah cardinals via compactness, with some provable in ZFC. Results strengthen set-sized logic properties by extending to class-sized versions.