I am a forty-five-year-old journalist who, for many years, didn't read the news. In high school, I knew about events like the O. J. Simpson trial and the Oklahoma City bombing, but not much else. In college, I was friends with geeky economics majors who read The Economist, but I'm pretty sure I never actually turned on CNN or bought a paper at the newsstand. I read novels, and magazines like Wired and Spin. If I went online, it wasn't to check the front page of the Times but to browse record reviews from College Music Journal. Somehow, during this time, I thought of myself as well informed.

You might think that such AI companionship bots--AI models with distinct "personalities" that can learn about you and act as a friend, lover, cheerleader, or more--appeal only to a fringe few, but that couldn't be further from the truth. A new research paper aimed at making such companions safer, by authors from Google DeepMind, the Oxford Internet Institute, and others, lays this bare: Character.AI, the platform being sued by Garcia, says it receives 20,000 queries per second, which is about a fifth of the estimated search volume served by Google. Interactions with these companions last four times longer than the average time spent interacting with ChatGPT. One companion site I wrote about, which was hosting sexually charged conversations with bots imitating underage celebrities, told me its active users averaged more than two hours per day conversing with bots, and that most of those users are members of Gen Z. The design of these AI characters makes lawmakers' concern well warranted.

The integration of large language models (LLMs) into information retrieval systems introduces new attack surfaces, particularly for adversarial ranking manipulations. We present StealthRank, a novel adversarial ranking attack that manipulates LLM-driven product recommendation systems while maintaining textual fluency and stealth. Unlike existing methods that often introduce detectable anomalies, StealthRank employs an energy-based optimization framework combined with Langevin dynamics to generate StealthRank Prompts (SRPs)-adversarial text sequences embedded within product descriptions that subtly yet effectively influence LLM ranking mechanisms. We evaluate StealthRank across multiple LLMs, demonstrating its ability to covertly boost the ranking of target products while avoiding explicit manipulation traces that can be easily detected. Our results show that StealthRank consistently outperforms state-of-the-art adversarial ranking baselines in both effectiveness and stealth, highlighting critical vulnerabilities in LLM-driven recommendation systems.

Google is bringing AI Mode to more people in the US. The company announced on Monday it would make the new search tool, first launched at the start of last month, to millions of more Labs users across the country. For uninitiated, AI Mode is a new dedicated tab within Search. It allows you to ask more complicated questions of Google, with a custom version of Gemini 2.0 doing the legwork to deliver a nuanced AI-generated response. Labs, meanwhile, is a beta program you can enroll your Google account in to gain access to new Search features before the company rolls them out to the public.

Microsoft unveiled a swath of new AI features during its 50th birthday bash in Redmond last week, and I got to try many of them. Copilot Vision holds the most potential for consumers, but Copilot's shopping features aren't far behind. If you're concerned about Copilot's new Memory and Personality traits, I wouldn't be -- it looks way too early to get a sense of what it will eventually be. Those were my impressions after sampling most of the new Copilot demonstrations on display at Microsoft's 50th anniversary celebration. Microsoft talked about its vision for a personable Copilot that could shop for you and walk you through tech issues step-by-step.

The 39th Annual AAAI Conference on Artificial Intelligence (AAAI 2025) took place in Philadelphia from Tuesday 25 February to Tuesday 4 March 2025. The programme featured eight invited talks. Susan works at the intersection of computer science and economics. In the past she has researched problems relating to mechanism design, auctions, pricing, and causal inference, but recently she has turned her attention to modelling worker career transitions using transformer models. In her talk, Susan described the research in a few of her recent papers covering topics such as the gender wage gap and economic prediction of labour sequence data.

No one's ever walked out of a meeting thinking, "Wow, that couldn't have been an email." If you're tired of hearing Steve and Chris ramble on for an hour, just stop going and send this AI voice recorder along instead. Even if you aren't in attendance, this gadget captures the whole conversation, uses ChatGPT-4 to provide an accurate transcript, and delivers smart summaries so you can get actual work done during that "important" meeting. Save 76 on the Focais meeting recorder and get free shipping while supplies last. You might not be allowed to skip the meeting outright, but this smart device at least allows you to secretly do work on your laptop--or just zone out into your happy place.

The year that ChatGPT went viral, only two US companies--OpenAI and Google--could boast truly cutting-edge artificial intelligence. Three years on, AI is no longer a two-horse race, nor is it purely an American one. A new report published today by Stanford University's Institute for Human-Centered AI (HAI) highlights just how crowded the field has become. OpenAI and Google are still neck and neck in the race to build bleeding-edge AI, the report shows. But several other companies are closing in.

Computational phenotyping is essential for biomedical research but often requires significant time and resources, especially since traditional methods typically involve extensive manual data review. While machine learning and natural language processing advancements have helped, further improvements are needed. Few studies have explored using Large Language Models (LLMs) for these tasks despite known advantages of LLMs for text-based tasks. T o facilitate further research in this area, we developed an evaluation framework, Evaluation of PHEnotyping for Observational Health Data (PHEONA), that outlines context-specific considerations. W e applied and demonstrated PHEONA on concept classification, a specific task within a broader phenotyping process for Acute Respiratory Failure (ARF) respiratory support therapies. From the sample concepts tested, we achieved high classification accuracy, suggesting the potential for LLM-based methods to improve computational phenotyping processes.

Inference-time computation offers a powerful axis for scaling the performance of language models. However, naively increasing computation in techniques like Best-of-N sampling can lead to performance degradation due to reward hacking. Toward a theoretical understanding of how to best leverage additional computation, we focus on inference-time alignment, which we formalize as the problem of improving the quality of responses drawn from a pre-trained policy, given a prompt of interest and access to an imperfect reward model. We analyze the performance of inference-time alignment algorithms in terms of (i) response quality, and (ii) compute, and provide new results that highlight the importance of the pre-trained policy's coverage over high-quality responses for performance and compute scaling: 1. We show that Best-of-$N$ alignment with an ideal choice for $N$ can achieve optimal performance under stringent notions of coverage, but provably suffers from reward hacking when $N$ is large, and fails to achieve tight guarantees under more realistic coverage conditions. 2. We introduce $\texttt{InferenceTimePessimism}$, a new algorithm which mitigates reward hacking through deliberate use of inference-time compute, implementing the principle of pessimism in the face of uncertainty via rejection sampling; we prove that its performance is optimal and does not degrade with $N$, meaning it is scaling-monotonic. We complement our theoretical results with an experimental evaluation that demonstrate the benefits of $\texttt{InferenceTimePessimism}$ across a variety of tasks and models.