We introduce and study an online problem arising in question answering systems. In this problem, an agent must sequentially classify user-submitted queries represented by $d$-dimensional embeddings drawn i.i.d. from an unknown distribution. The agent may consult a costly human expert for the correct label, or guess on her own without receiving feedback. The goal is to minimize regret against an oracle with free expert access. When the time horizon $T$ is at least exponential in the embedding dimension $d$, one can learn the geometry of the class regions: in this regime, we propose the Conservative Hull-based Classifier (CHC), which maintains convex hulls of expert-labeled queries and calls the expert as soon as a query lands outside all known hulls. CHC attains $\mathcal{O}(\log^d T)$ regret in $T$ and is minimax optimal for $d=1$. Otherwise, the geometry cannot be reliably learned without additional distributional assumptions. We show that when the queries are drawn from a subgaussian mixture, for $T \le e^d$, a Center-based Classifier (CC) achieves regret proportional to $N\log{N}$ where $N$ is the number of labels. To bridge these regimes, we introduce the Generalized Hull-based Classifier (GHC), a practical extension of CHC that allows for more aggressive guessing via a tunable threshold parameter. Our approach is validated with experiments, notably on real-world question-answering datasets using embeddings derived from state-of-the-art large language models.

Recent interpretability work on large language models (LLMs) has been increasingly dominated by a feature-discovery approach with the help of proxy modules. Then, the quality of features learned by, e.g., sparse auto-encoders (SAEs), is evaluated. This paradigm naturally raises a critical question: do such learned features have better properties than those already represented within the original model parameters, and unfortunately, only a few studies have made such comparisons systematically so far. In this work, we revisit the interpretability of feature vectors stored in feed-forward (FF) layers, given the perspective of FF as key-value memories, with modern interpretability benchmarks. Our extensive evaluation revealed that SAE and FFs exhibits a similar range of interpretability, although SAEs displayed an observable but minimal improvement in some aspects. Furthermore, in certain aspects, surprisingly, even vanilla FFs yielded better interpretability than the SAEs, and features discovered in SAEs and FFs diverged. These bring questions about the advantage of SAEs from both perspectives of feature quality and faithfulness, compared to directly interpreting FF feature vectors, and FF key-value parameters serve as a strong baseline in modern interpretability research.

Text watermarking plays a crucial role in ensuring the traceability and accountability of large language model (LLM) outputs and mitigating misuse. While promising, most existing methods assume perfect pseudorandomness. In practice, repetition in generated text induces collisions that create structured dependence, compromising Type I error control and invalidating standard analyses. We introduce a statistical framework that captures this structure through a hierarchical two-layer partition. At its core is the concept of minimal units -- the smallest groups treatable as independent across units while permitting dependence within. Using minimal units, we define a non-asymptotic efficiency measure and cast watermark detection as a minimax hypothesis testing problem. Applied to Gumbel-max and inverse-transform watermarks, our framework produces closed-form optimal rules. It explains why discarding repeated statistics often improves performance and shows that within-unit dependence must be addressed unless degenerate. Both theory and experiments confirm improved detection power with rigorous Type I error control. These results provide the first principled foundation for watermark detection under imperfect pseudorandomness, offering both theoretical insight and practical guidance for reliable tracing of model outputs.

OpenAI has released new estimates of the number of ChatGPT users who exhibit possible signs of mental health emergencies, including mania, psychosis or suicidal thoughts. The company said that around 0.07% of ChatGPT users active in a given week exhibited such signs, adding that its artificial intelligence (AI) chatbot recognizes and responds to these sensitive conversations. While OpenAI maintains these cases are extremely rare, critics said even a small percentage may amount to hundreds of thousands of people, as ChatGPT recently reached 800 million weekly active users, per boss Sam Altman. As scrutiny mounts, the company said it built a network of experts around the world to advise it. Those experts include more than 170 psychiatrists, psychologists, and primary care physicians who have practiced in 60 countries, the company said. They have devised a series of responses in ChatGPT to encourage users to seek help in the real world, according to OpenAI.

OpenAI claimed that its recent GPT-5 update improved user safety in a model evaluation involving more than 1,000 self-harm and suicide conversations. OpenAI claimed that its recent GPT-5 update improved user safety in a model evaluation involving more than 1,000 self-harm and suicide conversations. More than a million ChatGPT users each week send messages that include "explicit indicators of potential suicidal planning or intent", according to a blogpost published by OpenAI on Monday. The finding, part of an update on how the chatbot handles sensitive conversations, is one of the most direct statements from the artificial intelligence giant on the scale of how AI can exacerbate mental health issues. In addition to its estimates on suicidal ideations and related interactions, OpenAI also said that about 0.07% of users active in a given week - about 560,000 of its touted 800m weekly users - show "possible signs of mental health emergencies related to psychosis or mania".

The All-Clad pizza oven was one of my biggest surprises of the summer. All products featured on WIRED are independently selected by our editors. However, we may receive compensation from retailers and/or from purchases of products through these links. Cookware brand All-Clad surprised me this year. This summer, it breezed into the backyard pizza world with a debut pizza oven that I like as well as any oven I've tested this year.

Open AI CEO Sam Altman speaks during a summit on June 2, 2025 in San Francisco, California. Open AI CEO Sam Altman speaks during a summit on June 2, 2025 in San Francisco, California. Founded in 2015 as a nonprofit, rather than a for-profit company, it promised to develop AI "in the way that is most likely to benefit humanity." With billions of dollars in investments from Microsoft, Japanese bank SoftBank, and chipmaker Nvidia, however, OpenAI has proposed changing its corporate structure to give investors more control over its technology. Critics of the change include cofounder-turned-competitor, Elon Musk, and nonprofits concerned about OpenAI's adherence to its mission.

OpenAI released initial estimates about the share of users who may be experiencing symptoms like delusional thinking, mania, or suicidal ideation, and says it has tweaked GPT-5 to respond more effectively. For the first time ever, OpenAI has released a rough estimate of how many ChatGPT users globally may show signs of having a severe mental health crisis in a typical week. The company said Monday that it worked with experts around the world to make updates to the chatbot so it can more reliably recognize indicators of mental distress and guide users toward real-world support. In recent months, a growing number of people have ended up hospitalized, divorced, or dead after having long, intense conversations with ChatGPT. Some of their loved ones allege the chatbot fueled their delusions and paranoia.

The Download: what to make of OpenAI's Atlas browser, and how to make climate progress I tried OpenAI's new Atlas browser but I still don't know what it's for OpenAI rolled out a new web browser last week called Atlas. It comes with ChatGPT built in, along with an agent, so that you can browse, get answers, and have automated tasks performed on your behalf all at the same time. I've spent the past several days tinkering with Atlas. I've used it to do all my normal web browsing, and also tried to take advantage of the ChatGPT functions--plus I threw some weird agentic tasks its way to see how it did with those. My impression is that Atlas is fine? But my big takeaway is that it's pretty pointless for anyone not employed by OpenAI.

Are Kids Still Looking for Careers in Tech? AI is changing what careers are possible for students interested in STEM subjects. WIRED spoke with five aspiring scientists to find out how they're preparing for the future. Today's high school students face an uncertain road ahead. AI is changing what skills are valued in the job market, and the Trump administration's funding cuts have stalled scientific research across disciplines.