Rising seniors are the last class of students who remember high school before ChatGPT. But only just barely: OpenAI's chatbot was released months into their freshman year. Ever since then, writing essays hasn't required, well, writing. By the time these students graduate next spring, they will have completed almost four full years of AI high school. Gone already are the days when using AI to write an essay meant copying and pasting its response verbatim.

A US medical journal has warned against using ChatGPT for health information after a man developed a rare condition following an interaction with the chatbot about removing table salt from his diet. An article in the Annals of Internal Medicine reported a case in which a 60-year-old man developed bromism, also known as bromide toxicity, after consulting ChatGPT. The article described bromism as a "well-recognised" syndrome in the early 20th century that was thought to have contributed to almost one in 10 psychiatric admissions at the time. The patient told doctors that after reading about the negative effects of sodium chloride, or table salt, he consulted ChatGPT about eliminating chloride from his diet and started taking sodium bromide over a three-month period. This was despite reading that "chloride can be swapped with bromide, though likely for other purposes, such as cleaning".

While everyone else has been talking about Nvidia's GPUs, Lisa Su has discreetly turned AMD into a chipmaking phenom. Su, the leader of AMD, moves fast these days, though I suspect that's always been the case. Her company's chips underpin the artificial intelligence that's changing the world at breakneck speeds. To hear Su and literally everyone else in semiconductors talk about it, the US is in an AI with China--and the rules keep changing . The Trump administration has once again shifted its stance on what kind of chips can and can't be shipped to China, with the latest decree being that the US will take a 15 percent cut of AMD and Nvidia chip sales to China. Meanwhile, on the home front, Su has claimed that AMD's newest AI chips can outperform Nvidia's--part of her strategy to keep eroding Nvidia's dominance in the market. So, yeah: Be ready to keep up. Under Lisa Su, the stalwart American semiconductor company has reasserted itself as a force in the age of AI. "Reasserted" doesn't do it justice: Su took a struggling AMD and executed a 10-year turnaround that has been, as one economist put it, nothing short of remarkable. Since 2014, when Su took over as CEO, AMD's market cap has risen from around $2 billion to nearly $300 billion. Aside from her well-known bona fides, Su herself--what drives her, what inspires her, what irritates her, where her politics lie--is less known. This is what I was hoping to learn when I visited AMD's offices and labs in the hills of Austin, Texas, on a day in late June when the wind seemed to do little more than push heat around. Our conversation kicked off with China, which accounts for nearly a quarter of AMD's business. Su now travels frequently to Washington, DC, to grease the wheels. "We've come to realize that export controls are a bit of a fact of life," she told me, "just given how critical the chips that we make are." In other words, it's precisely because AMD's chips are so darn important--to national security, to national economies--that they're now at the heart of modern statecraft.

Against those expectations, the model has mostly underwhelmed. People have highlighted glaring mistakes in GPT-5's responses, countering Altman's claim made at the launch that it works like "a legitimate PhD-level expert in anything any area you need on demand." Early testers have also found issues with OpenAI's promise that GPT-5 automatically works out what type of AI model is best suited for your question--a reasoning model for more complicated queries, or a faster model for simpler ones. Altman seems to have conceded that this feature is flawed and takes away user control. However there is good news too: the model seems to have eased the problem of ChatGPT sucking up to users, with GPT-5 less likely to shower them with over the top compliments.

In this paper, we introduce a score-based diffusion model appr oach for adaptively learning the time-evolving solutions of stochastic partial differential equat ions (SPDEs) through recursive Bayesian inference. Partial differential equations (PDEs) are fundamental tools for modeling the dynamic behavior of complex physical systems. While they have been widely suc cessful in scientific and engineering applications, many practical scenarios involve inherent unc ertainties due to limited physical knowledge and environmental variability. For example, in climate and meteorological modeling, uncertainties in initial conditions, boundary data, and subgrid-scale ph ysical processes can significantly affect the accuracy of predictions governed by PDEs such as the Navier-Stokes or advection-diffusion equations. Similarly, in porous media flow problems, spatial het erogeneity and limited characterization of subsurface properties -- such as permeability or porosity -- i ntroduce substantial uncertainty into models governed by Darcy's law and related PDEs, making accu rate prediction particularly challenging. To capture these uncertainty effects and support rel iable predictive analysis, it is essential to incorporate SPDEs into mathematical modeling framework . The numerical solution of SPDEs has thus become a central focus of the uncertainty quantifica tion (UQ) community, where significant efforts have been dedicated to developing efficient solvers tha t can accurately characterize and propagate uncertainty in high-dimensional, nonlinear dynamica l systems (see, e.g., [1, 2, 13, 21, 36, 42, 53] and the reference therein). Despite advances in SPDE solvers capable of quantifying unc ertainty, significant challenges remain.

Recent advances in learning rate (LR) scheduling have demonstrated the effectiveness of decay-free approaches that eliminate the traditional decay phase while maintaining competitive performance. Model merging techniques have emerged as particularly promising solutions in this domain. We present Warmup-Stable and Merge (WSM), a general framework that establishes a formal connection between learning rate decay and model merging. WSM provides a unified theoretical foundation for emulating various decay strategies-including cosine decay, linear decay and inverse square root decay-as principled model averaging schemes, while remaining fully compatible with diverse optimization methods. Through extensive experiments, we identify merge duration-the training window for checkpoint aggregation-as the most critical factor influencing model performance, surpassing the importance of both checkpoint interval and merge quantity. Our framework consistently outperforms the widely-adopted Warmup-Stable-Decay (WSD) approach across multiple benchmarks, achieving significant improvements of +3.5% on MATH, +2.9% on HumanEval, and +5.5% on MMLU-Pro. The performance advantages extend to supervised fine-tuning scenarios, highlighting WSM's potential for long-term model refinement.

Machine learning (ML) systems are increasingly deployed in high-stakes domains where reliability is paramount. This thesis investigates how uncertainty estimation can enhance the safety and trustworthiness of ML, focusing on selective prediction -- where models abstain when confidence is low. We first show that a model's training trajectory contains rich uncertainty signals that can be exploited without altering its architecture or loss. By ensembling predictions from intermediate checkpoints, we propose a lightweight, post-hoc abstention method that works across tasks, avoids the cost of deep ensembles, and achieves state-of-the-art selective prediction performance. Crucially, this approach is fully compatible with differential privacy (DP), allowing us to study how privacy noise affects uncertainty quality. We find that while many methods degrade under DP, our trajectory-based approach remains robust, and we introduce a framework for isolating the privacy-uncertainty trade-off. Next, we then develop a finite-sample decomposition of the selective classification gap -- the deviation from the oracle accuracy-coverage curve -- identifying five interpretable error sources and clarifying which interventions can close the gap. This explains why calibration alone cannot fix ranking errors, motivating methods that improve uncertainty ordering. Finally, we show that uncertainty signals can be adversarially manipulated to hide errors or deny service while maintaining high accuracy, and we design defenses combining calibration audits with verifiable inference. Together, these contributions advance reliable ML by improving, evaluating, and safeguarding uncertainty estimation, enabling models that not only make accurate predictions -- but also know when to say "I do not know".

Integrating multi-omics data, such as DNA methylation, mRNA expression, and microRNA (miRNA) expression, offers a comprehensive view of the biological mechanisms underlying disease. However, the high dimensionality and complex interactions among omics layers present major challenges for predictive modeling. We propose Multi-Omics integration with Tree-generated Graph Neural Network (MOTGNN), a novel and interpretable framework for binary disease classification. MOTGNN employs eXtreme Gradient Boosting (XGBoost) to perform omics-specific supervised graph construction, followed by modality-specific Graph Neural Networks (GNNs) for hierarchical representation learning, and a deep feedforward network for cross-omics integration. On three real-world disease datasets, MOTGNN outperforms state-of-the-art baselines by 5-10% in accuracy, ROC-AUC, and F1-score, and remains robust to severe class imbalance (e.g., 87.2% vs. 33.4% F1 on imbalanced data). The model maintains computational efficiency through sparse graphs (2.1-2.8 edges per node) and provides built-in interpretability, revealing both top-ranked biomarkers and the relative contributions of each omics modality. These results highlight MOTGNN's potential to improve both predictive accuracy and interpretability in multi-omics disease modeling.

Deep learning approaches are increasingly relevant for time series forecasting tasks. Methods such as N-BEATS, which is built on stacks of multilayer perceptrons (MLPs) blocks, have achieved state-of-the-art results on benchmark datasets and competitions. N-BEATS is also more interpretable relative to other deep learning approaches, as it decomposes forecasts into different time series components, such as trend and sea-sonality. In this work, we present N-BEATS-MOE, an extension of N-BEATS based on a Mixture-of-Experts (MoE) layer. N-BEATS-MOE employs a dynamic block weighting strategy based on a gating network which allows the model to better adapt to the characteristics of each time series. We also hypothesize that the gating mechanism provides additional inter-pretability by identifying which expert is most relevant for each series. We evaluate our method across 12 benchmark datasets against several approaches, achieving consistent improvements on several datasets, especially those composed of heterogeneous time series.

Continual learning aims to enable models to retain existing knowledge while continuously acquiring new knowledge, typically derived from new data, under the constraint of limited or restricted access to data related to previously learned knowledge. Depending on the defined scenarios, continual learning is typically categorized into task-incremental learning, class-incremental learning [4], and domain-incremental learning. Among these three settings, class-incremental learning has been the most extensively studied. Taking classification tasks as an example, class-incremental learning divides a dataset into multiple sessions, where the classes in different sessions do not overlap. The model is required to learn the classes of each session over time and is evaluated on all classes after each update. In contrast, task-incremental learning assumes knowledge of which session the data belongs to during evaluation, allowing the model to classify only within the corresponding task. On the other hand, in domain-incremental learning, different sessions involve data from the same set of classes but with different distributions. The model is required to correctly identify the domain of the test data while performing classification. Research targeting these settings has led to significant improvements in model performance in scenarios requiring multi-stage fine-tuning.