Others say they don't need us. It's time for us to think of ourselves not as the high priests of health care but as what we have always been: healers. Not long ago, I cared for a middle-aged man I'll call Jim, who was generally healthy but had recently started to feel sluggish. One of his friends told him to try a hormone supplement. After Jim saw on social media that Robert F. Kennedy, Jr., the Trump Administration's Secretary of Health and Human Services, had endorsed supplements as a part of an "anti-aging" regimen, he ordered one from a telehealth company. A few months later, he noticed swelling and pain in his calf. ChatGPT warned him that he might have a blood clot.

We introduce Stackelberg Learning from Human Feedback (SLHF), a new framework for preference optimization. SLHF frames the alignment problem as a sequential-move game between two policies: a Leader, which commits to an action, and a Follower, which responds conditionally on the Leader's action. This approach decomposes preference optimization into a refinement problem for the Follower and an optimization problem against an adversary for the Leader. Unlike Reinforcement Learning from Human Feedback (RLHF), which assigns scalar rewards to actions, or Nash Learning from Human Feedback (NLHF), which seeks a simultaneous-move equilibrium, SLHF leverages the asymmetry of sequential play to capture richer preference structures. The sequential design of SLHF naturally enables inference-time refinement, as the Follower learns to improve the Leader's actions, and these refinements can be leveraged through iterative sampling. We compare the solution concepts of SLHF, RLHF, and NLHF, and lay out key advantages in consistency, data sensitivity, and robustness to intransitive preferences. Experiments on large language models demonstrate that SLHF achieves strong alignment across diverse preference datasets, scales from 0.5B to 8B parameters, and yields inference-time refinements that transfer across model families without further fine-tuning.

Data is a critical asset for training large language models (LLMs), alongside compute resources and skilled workers. While some training data is publicly available, substantial investment is required to generate proprietary datasets, such as human preference annotations or to curate new ones from existing sources. As larger datasets generally yield better model performance, two natural questions arise. First, how can data owners make informed decisions about curation strategies and data sources investment? Second, how can multiple data owners collaboratively pool their resources to train superior models while fairly distributing the benefits? This problem, data valuation, which is not specific to large language models, has been addressed by the machine learning community through the lens of cooperative game theory, with the Shapley value being the prevalent solution concept. However, computing Shapley values is notoriously expensive for data valuation, typically requiring numerous model retrainings, which can become prohibitive for large machine learning models. In this work, we demonstrate that this computational challenge is dramatically simplified for LLMs trained with Direct Preference Optimization (DPO). We show how the specific mathematical structure of DPO enables scalable Shapley value computation. We believe this observation unlocks many applications at the intersection of data valuation and large language models.

China figured out how to sell EVs. Now it has to bury their batteries. As early electric cars age out, hundreds of thousands of used batteries are flooding the market, fueling a gray recycling economy even as Beijing and big manufacturers scramble to build a more orderly system. In August 2025, Wang Lei decided it was finally time to say goodbye to his electric vehicle. Wang, who is 39, had bought the car in 2016, when EVs still felt experimental in Beijing. It was a compact Chinese brand.

Though the two countries are now in a race to develop atomic technology, China's most advanced reactor was the result of collaboration with American scientists. This April, in a speech given at the Shanghai branch of the Chinese Academy of Sciences, the physicist Xu Hongjie announced a breakthrough. For over a decade, his team had been working on an experimental nuclear reactor that runs on a lava-hot solution of fissile material and molten salt, rather than on solid fuel. The reactor, which went online two years ago, was a feat in itself. It is still the only one of its kind in operation in the world, and has the potential to be both safer and more efficient than the water-cooled nuclear plants that dominate the industry. Now, Xu explained, his team had been able to refuel the reactor without shutting it down, demonstrating a level of mastery over their new system. As dazzling as that was, the timing of Xu's speech also freighted the topic with geopolitical import. Only a few months earlier, DeepSeek, the Chinese artificial-intelligence company, had set alarms ringing through the U.S. tech world when it became clear that the relatively small Chinese startup, operating under U.S. export controls, had created a large language model that rivalled anything devised by the behemoths of Silicon Valley.

George Osborne has a new job in tech, and it doesn't bode well for Britain OpenAI is the latest to make a political hire as big tech spreads its tentacles around the world. Since leaving frontline politics, the former chancellor has served as the chair of the Northern Powerhouse Partnership, edited (not entirely successfully) the Evening Standard, advised asset manager BlackRock, joined boutique advisory firm Robey Warshaw, been appointed as the chair of the British Museum and taken on roles including advising crypto firm Coinbase . But Osborne's latest job is the most eye-opening - and is an alarming augur of what is to come. OpenAI, the maker of ChatGPT, has become the latest organisation to employ Osborne . He will run OpenAI for Countries, a unit tasked with working directly with governments while expanding the company's Stargate datacentre programme beyond the US.

Warner Bros. rejects Paramount's hostile bid Some help you pull-in locally-stored data and another lets you organize Apple Music songs. OpenAI has introduced an app directory that's now available right inside ChatGPT, the company announced. Apps extend ChatGPT conversations by bringing in new context and letting users take action like order groceries, turn an outline into a slide deck or search for an apartment, the company wrote in a blog post . OpenAI also noted in a help document that connector apps like Google Drive are now simply called apps. The new apps section (on iOS, Android and web) is divided into Feature, Lifestyle and Productivity categories, letting you connect to commonly used apps and sites like Booking.com,

Welcome to our annual list of the worst, least successful, and simply dumbest technologies of the year. We like to think there's a lesson in every technological misadventure. But when technology becomes dependent on power, sometimes the takeaway is simpler: it would have been better to stay away. Here are some of the more notable ones . Each time you've heard a borderline outlandish idea of what AI will be capable of, it often turns out that Sam Altman was, if not the first to articulate it, at least the most persuasive and influential voice behind it. For more than a decade he has been known in Silicon Valley as a world-class fundraiser and persuader.

The Cybertruck, sycophantic AI, and humanoid robots all made this year's list of the biggest technology failures. Welcome to our annual list of the worst, least successful, and simply dumbest technologies of the year. This year, politics was a recurring theme. Donald Trump swept back into office and used his executive pen to reshape the fortunes of entire sectors, from renewables to cryptocurrency. The wrecking-ball act began even before his inauguration, when the president-elect marketed his own memecoin, $TRUMP, in a shameless act of merchandising that, of course, we honor on this year's worst tech list. We like to think there's a lesson in every technological misadventure.

Autoregressive models (ARMs) currently constitute the dominant paradigm for large language models (LLMs). Energy-based models (EBMs) represent another class of models, which have historically been less prevalent in LLM development, yet naturally characterize the optimal policy in post-training alignment. In this paper, we provide a unified view of these two model classes. Taking the chain rule of probability as a starting point, we establish an explicit bijection between ARMs and EBMs in function space, which we show to correspond to a special case of the soft Bellman equation in maximum entropy reinforcement learning. Building upon this bijection, we derive the equivalence between supervised learning of ARMs and EBMs. Furthermore, we analyze the distillation of EBMs into ARMs by providing theoretical error bounds. Our results provide insights into the ability of ARMs to plan ahead, despite being based on the next-token prediction paradigm.