OpenAI has banned the accounts of a group of Chinese users who had attempted to use ChatGPT to debug and edit code for an AI social media surveillance tool, the company said Friday. The campaign, which OpenAI calls Peer Review, saw the group prompt ChatGPT to generate sales pitches for a program those documents suggest was designed to monitor anti-Chinese sentiment on X, Facebook, YouTube, Instagram and other platforms. The operation appears to have been particularly interested in spotting calls for protests against human rights violations in China, with the intent of sharing those insights with the country's authorities. "This network consisted of ChatGPT accounts that operated in a time pattern consistent with mainland Chinese business hours, prompted our models in Chinese, and used our tools with a volume and variety consistent with manual prompting, rather than automation," said OpenAI. "The operators used our models to proofread claims that their insights had been sent to Chinese embassies abroad, and to intelligence agents monitoring protests in countries including the United States, Germany and the United Kingdom."

Threat actors, some likely based in China and Iran, are formulating new ways to hijack and utilize American artificial intelligence (AI) models for malicious intent, including covert influence operations, according to a new report from OpenAI. The February report includes two disruptions involving threat actors that appear to have originated from China. According to the report, these actors have used, or at least attempted to use, models built by OpenAI and Meta. In one example, OpenAI banned a ChatGPT account that generated comments critical of Chinese dissident Cai Xia. The comments were posted on social media by accounts that claimed to be people based in India and the U.S.

Sitting in Lincoln Center awaiting the curtain for Ayad Akhtar's McNeal--a much anticipated theater production starring Robert Downey Jr., with ChatGPT in a supporting role--I mused how playwrights have been dealing with the implications of AI for over a century. In 1920--well before Alan Turing devised his famous test and decades before the 1956 summer Dartmouth conference that gave artificial intelligence its name--a Czech playwright named Karel Čapek wrote R.U.R.--Rossum's Universal Robots. Not only was this the first time the word "robot" was employed, but Čapek may qualify as the first AI doomer, since his play dramatized an android uprising that slaughtered all of humanity, save for a single soul. Also on the boards in New York City this winter was a small black-box production called Doomers, a thinly veiled dramatization of the weekend where OpenAI's nonprofit board gave Sam Altman the boot, only to see him return after an employee rebellion. Neither of these productions have the pizzazz of a splashy Broadway extravaganza--maybe later we'll buy tickets to a musical where Altman and Elon Musk have a dance-off--but both grapple with issues that reverberate in Silicon Valley conference rooms, Congressional hearings, and late-night drinking sessions at the annual NeurIPS conference.

Previously, we have outlined three guiding principles for developing embodied artificial intelligence (EAI) systems.1 EAI systems should not depend on predefined, complex logic to handle specific scenarios. Instead, they must incorporate evolutionary learning mechanisms, enabling continuous adaptation to their operational environments. Additionally, the environment significantly influences not only physical behaviors but also cognitive structures. While the third principle focuses on simulation, the first two principles emphasize building EAI foundation models capable of learning from the EAI systems' operating environments. A common approach for EAI foundation models is to directly utilize pretrained large models.

Operator is now out in Australia, Brazil, Canada, India, Japan, Singapore, South Korea, the UK and most places where ChatGPT is also available, OpenAI has announced. The company launched Operator in the US back in January, introducing it as an "agent that can go to the web to perform tasks" for the user. Operator can handle various browser-based tasks for users, such as filling out forms, making restaurant reservations and ordering groceries. At the moment, it's still a research preview in its early stages that comes with limitations, but the company said it hopes to roll out improvements based on user feedback. Operator is now rolling out to Pro users in Australia, Brazil, Canada, India, Japan, Singapore, South Korea, the UK, and most places ChatGPT is available.

Google's thoughtful AI research partner, Deep Research, is now available to Google Workspace users, Google said Thursday. And that's not all: Google Workspace users who sign up for Gemini Advanced can try other, experimental AI models, too. On Tuesday, Google announced that Gemini Advanced with Deep Research is now accessible by mobile users. On Thursday, Google migrated Gemini Advanced with Deep Research to Workspace users as well, provided that the Workspace subscription includes the extra Gemini Advanced subscription. Before this, Deep Research was only available in Gemini Advanced on the web, for 20 per month. Google's Workspace blog suggests that a salesperson could use Gemini Advanced with Deep Research to prepare a report on a prospective client, or a teacher could use it with lesson planning -- similar to what Google said Deep Research could do when it was originally announced.

We introduce the Contrastive Similarity Space Embedding Algorithm (ContraSim), a novel framework for uncovering the global semantic relationships between daily financial headlines and market movements. ContraSim operates in two key stages: (I) Weighted Headline Augmentation, which generates augmented financial headlines along with a semantic fine-grained similarity score, and (II) Weighted Self-Supervised Contrastive Learning (WSSCL), an extended version of classical self-supervised contrastive learning that uses the similarity metric to create a refined weighted embedding space. This embedding space clusters semantically similar headlines together, facilitating deeper market insights. Empirical results demonstrate that integrating ContraSim features into financial forecasting tasks improves classification accuracy from WSJ headlines by 7%. Moreover, leveraging an information density analysis, we find that the similarity spaces constructed by ContraSim intrinsically cluster days with homogeneous market movement directions, indicating that ContraSim captures market dynamics independent of ground truth labels. Additionally, ContraSim enables the identification of historical news days that closely resemble the headlines of the current day, providing analysts with actionable insights to predict market trends by referencing analogous past events.

Transformers and large language models (LLMs) have revolutionized machine learning, with attention mechanisms at the core of their success. As the landscape of attention variants expands, so too do the challenges of optimizing their performance, particularly across different hardware platforms. Current optimization strategies are often narrowly focused, requiring extensive manual intervention to accommodate changes in model configurations or hardware environments. In this paper, we introduce AttentionEngine, a comprehensive framework designed to streamline the optimization of attention mechanisms across heterogeneous hardware backends. By decomposing attention computation into modular operations with customizable components, AttentionEngine enables flexible adaptation to diverse algorithmic requirements. The framework further automates kernel optimization through a combination of programmable templates and a robust cross-platform scheduling strategy. Empirical results reveal performance gains of up to 10x on configurations beyond the reach of existing methods. AttentionEngine offers a scalable, efficient foundation for developing and deploying attention mechanisms with minimal manual tuning. Our code has been open-sourced and is available at https://github.com/microsoft/AttentionEngine.

Robotics has long sought to develop visual-servoing robots capable of completing previously unseen long-horizon tasks. Hierarchical approaches offer a pathway for achieving this goal by executing skill combinations arranged by a task planner, with each visuomotor skill pre-trained using a specific imitation learning (IL) algorithm. However, even in simple long-horizon tasks like skill chaining, hierarchical approaches often struggle due to a problem we identify as Observation Space Shift (OSS), where the sequential execution of preceding skills causes shifts in the observation space, disrupting the performance of subsequent individually trained skill policies. To validate OSS and evaluate its impact on long-horizon tasks, we introduce BOSS (a Benchmark for Observation Space Shift). BOSS comprises three distinct challenges: "Single Predicate Shift", "Accumulated Predicate Shift", and "Skill Chaining", each designed to assess a different aspect of OSS's negative effect. We evaluated several recent popular IL algorithms on BOSS, including three Behavioral Cloning methods and the Visual Language Action model OpenVLA. Even on the simplest challenge, we observed average performance drops of 67%, 35%, 34%, and 54%, respectively, when comparing skill performance with and without OSS. Additionally, we investigate a potential solution to OSS that scales up the training data for each skill with a larger and more visually diverse set of demonstrations, with our results showing it is not sufficient to resolve OSS. The project page is: https://boss-benchmark.github.io/

Large language models have demonstrated remarkable progress in mathematical reasoning, leveraging chain-of-thought and test-time compute scaling. However, many open questions remain regarding the interplay between reasoning token usage and accuracy gains. In particular, when comparing models across generations, it is unclear whether improved performance results from longer reasoning chains or more efficient reasoning. We systematically analyze chain-of-thought length across o1-mini and o3-mini variants on the Omni-MATH benchmark, finding that o3-mini (m) achieves superior accuracy without requiring longer reasoning chains than o1-mini. Moreover, we show that accuracy generally declines as reasoning chains grow across all models and compute settings, even when controlling for difficulty of the questions. This accuracy drop is significantly smaller in more proficient models, suggesting that new generations of reasoning models use test-time compute more effectively. Finally, we highlight that while o3-mini (h) achieves a marginal accuracy gain over o3-mini (m), it does so by allocating substantially more reasoning tokens across all problems, even the ones that o3-mini (m) can already solve. These findings provide new insights into the relationship between model capability and reasoning length, with implications for efficiency, scaling, and evaluation methodologies.