Three big things we still don't know about AI's energy burden AI companies are revealing the one number that researchers have long sought. Earlier this year, when my colleague Casey Crownhart and I spent six months researching the climate and energy burden of AI, we came to see one number in particular as our white whale: how much energy the leading AI models, like ChatGPT or Gemini, use up when generating a single response. This fundamental number remained elusive even as the scramble to power AI escalated to the White House and the Pentagon, and as projections showed that in three years AI could use as much electricity as 22% of all US households. The problem with finding that number, as we explain in our piece published in May, was that AI companies are the only ones who have it. We pestered Google, OpenAI, and Microsoft, but each company refused to provide its figure. Researchers we spoke to who study AI's impact on energy grids compared it to trying to measure the fuel efficiency of a car without ever being able to drive it, making guesses based on rumors of its engine size and what it sounds like going down the highway.

Help! My therapist is secretly using ChatGPT Some patients have discovered their private confessions are being quietly fed into AI. In Silicon Valley's imagined future, AI models are so empathetic that we'll use them as therapists. They'll provide mental-health care for millions, unimpeded by the pesky requirements for human counselors, like the need for graduate degrees, malpractice insurance, and sleep. Down here on Earth, something very different has been happening. Last week, we published a story about people finding out that their therapists were secretly using ChatGPT during sessions. In some cases it wasn't subtle; one therapist accidentally shared his screen during a virtual appointment, allowing the patient to see his own private thoughts being typed into ChatGPT in real time.

Eye-tracking technology can aid in understanding neurodevelopmental disorders and tracing a person's identity. However, this technology poses a significant risk to privacy, as it captures sensitive information about individuals and increases the likelihood that data can be traced back to them. This paper proposes a human-centered framework designed to prevent identity backtracking while preserving the pedagogical benefits of AI-powered eye tracking in interactive learning environments. We explore how real-time data anonymization, ethical design principles, and regulatory compliance (such as GDPR) can be integrated to build trust and transparency. We first demonstrate the potential for backtracking student IDs and diagnoses in various scenarios using serious game-based eye-tracking data. We then provide a two-stage privacy-preserving framework that prevents participants from being tracked while still enabling diagnostic classification. The first phase covers four scenarios: I) Predicting disorder diagnoses based on different game levels. II) Predicting student IDs based on different game levels. III) Predicting student IDs based on randomized data. IV) Utilizing K-Means for out-of-sample data. In the second phase, we present a two-stage framework that preserves privacy. We also employ Federated Learning (FL) across multiple clients, incorporating a secure identity management system with dummy IDs and administrator-only access controls. In the first phase, the proposed framework achieved 99.3% accuracy for scenario 1, 63% accuracy for scenario 2, and 99.7% accuracy for scenario 3, successfully identifying and assigning a new student ID in scenario 4. In phase 2, we effectively prevented backtracking and established a secure identity management system with dummy IDs and administrator-only access controls, achieving an overall accuracy of 99.40%.

In this position paper, we address the persistent gap between rapidly growing AI capabilities and lagging safety progress. Existing paradigms divide into ``Make AI Safe'', which applies post-hoc alignment and guardrails but remains brittle and reactive, and ``Make Safe AI'', which emphasizes intrinsic safety but struggles to address unforeseen risks in open-ended environments. We therefore propose \textit{safe-by-coevolution} as a new formulation of the ``Make Safe AI'' paradigm, inspired by biological immunity, in which safety becomes a dynamic, adversarial, and ongoing learning process. To operationalize this vision, we introduce \texttt{R$^2$AI} -- \textit{Resistant and Resilient AI} -- as a practical framework that unites resistance against known threats with resilience to unforeseen risks. \texttt{R$^2$AI} integrates \textit{fast and slow safe models}, adversarial simulation and verification through a \textit{safety wind tunnel}, and continual feedback loops that guide safety and capability to coevolve. We argue that this framework offers a scalable and proactive path to maintain continual safety in dynamic environments, addressing both near-term vulnerabilities and long-term existential risks as AI advances toward AGI and ASI.

Targeted data poisoning attacks pose an increasingly serious threat due to their ease of deployment and high success rates. These attacks aim to manipulate the prediction for a single test sample in classification models. Unlike indiscriminate attacks that aim to decrease overall test performance, targeted attacks present a unique threat to individual test instances. This threat model raises a fundamental question: what factors make certain test samples more susceptible to successful poisoning than others? We investigate how attack difficulty varies across different test instances and identify key characteristics that influence vulnerability. This paper introduces three predictive criteria for targeted data poisoning difficulty: ergodic prediction accuracy (analyzed through clean training dynamics), poison distance, and poison budget. Our experimental results demonstrate that these metrics effectively predict the varying difficulty of real-world targeted poisoning attacks across diverse scenarios, offering practitioners valuable insights for vulnerability assessment and understanding data poisoning attacks.

Electronic Health Records (EHRs), comprising diverse clinical data such as diagnoses, medications, and laboratory results, hold great promise for translational research. EHR-derived data have advanced disease prevention, improved clinical trial recruitment, and generated real-world evidence. Synthesizing EHRs across institutions enables large-scale, generalizable studies that capture rare diseases and population diversity, but remains hindered by the heterogeneity of medical codes, institution-specific terminologies, and the absence of standardized data structures. These barriers limit the interpretability, comparability, and scalability of EHR-based analyses, underscoring the need for robust methods to harmonize and extract meaningful insights from distributed, heterogeneous data. To address this, we propose MASH (Multi-source Automated Structured Hierarchy), a fully automated framework that aligns medical codes across institutions using neural optimal transport and constructs hierarchical graphs with learned hyperbolic embeddings. During training, MASH integrates information from pre-trained language models, co-occurrence patterns, textual descriptions, and supervised labels to capture semantic and hierarchical relationships among medical concepts more effectively. Applied to real-world EHR data, including diagnosis, medication, and laboratory codes, MASH produces interpretable hierarchical graphs that facilitate the navigation and understanding of heterogeneous clinical data. Notably, it generates the first automated hierarchies for unstructured local laboratory codes, establishing foundational references for downstream applications.

Representation-based multi-task learning (MTL) improves efficiency by learning a shared structure across tasks, but its practical application is often hindered by contamination, outliers, or adversarial tasks. Most existing methods and theories assume a clean or near-clean setting, failing when contamination is significant. This paper tackles representation MTL with an unknown and potentially large contamination proportion, while also allowing for heterogeneity among inlier tasks. We introduce a Robust and Adaptive Spectral method (RAS) that can distill the shared inlier representation effectively and efficiently, while requiring no prior knowledge of the contamination level or the true representation dimension. Theoretically, we provide non-asymptotic error bounds for both the learned representation and the per-task parameters. These bounds adapt to inlier task similarity and outlier structure, and guarantee that RAS performs at least as well as single-task learning, thus preventing negative transfer. We also extend our framework to transfer learning with corresponding theoretical guarantees for the target task. Extensive experiments confirm our theory, showcasing the robustness and adaptivity of RAS, and its superior performance in regimes with up to 80\% task contamination.

Beginning with text and images, generative AI has expanded to audio, video, computer code, and molecules. Yet, if generative AI is the answer, what is the question? We explore the foundations of generation as a distinct machine learning task with connections to prediction, compression, and decision-making. We survey five major generative model families: autoregressive models, variational autoencoders, normalizing flows, generative adversarial networks, and diffusion models. We then introduce a probabilistic framework that emphasizes the distinction between density estimation and generation. We review a game-theoretic framework with a two-player adversary-learner setup to study generation. We discuss post-training modifications that prepare generative models for deployment. We end by highlighting some important topics in socially responsible generation such as privacy, detection of AI-generated content, and copyright and IP. We adopt a task-first framing of generation, focusing on what generation is as a machine learning problem, rather than only on how models implement it.

We introduce the Smoothed Online Optimization for Target Tracking (SOOTT) problem, a new framework that integrates three key objectives in online decision-making under uncertainty: (1) tracking cost for following a dynamically moving target, (2) adversarial perturbation cost for withstanding unpredictable disturbances, and (3) switching cost for penalizing abrupt changes in decisions. This formulation captures real-world scenarios such as elastic and inelastic workload scheduling in AI clusters, where operators must balance long-term service-level agreements (e.g., LLM training) against sudden demand spikes (e.g., real-time inference). We first present BEST, a robust algorithm with provable competitive guarantees for SOOTT. To enhance practical performance, we introduce CoRT, a learning-augmented variant that incorporates untrusted black-box predictions (e.g., from ML models) into its decision process. Our theoretical analysis shows that CoRT strictly improves over BEST when predictions are accurate, while maintaining robustness under arbitrary prediction errors. We validate our approach through a case study on workload scheduling, demonstrating that both algorithms effectively balance trajectory tracking, decision smoothness, and resilience to external disturbances.

Motivated by the need for rigorous and scalable evaluation of large language models, we study contextual preference inference for pairwise comparison functionals of context-dependent preference score functions across domains. Focusing on the contextual Bradley-Terry-Luce model, we develop a semiparametric efficient estimator that automates the debiased estimation through aggregating weighted residual balancing terms across the comparison graph. We show that the efficiency is achieved when the weights are derived from a novel strategy called Fisher random walk. We also propose a computationally feasible method to compute the weights by a potential representation of nuisance weight functions. We show our inference procedure is valid for general score function estimators accommodating the practitioners' need to implement flexible deep learning methods. We extend the procedure to multiple hypothesis testing using a Gaussian multiplier bootstrap that controls familywise error and to distributional shift via a cross-fitted importance-sampling adjustment for target-domain inference. Numerical studies, including language model evaluations under diverse contexts, corroborate the accuracy, efficiency, and practical utility of our method.