Elon Musk has launched a 97.4 billion bid to take control of OpenAI. The Wall Street Journal reports a group of investors led by Musk's xAI submitted an unsolicited offer to the company's board of directors on Monday. The group wants to buy the nonprofit that controls OpenAI's for-profit arm. When asked for comment, an OpenAI spokesperson pointed Engadget to an X post from CEO Sam Altman. "No thank you but we will buy twitter for 9.74 billion if you want," Altman wrote on the social media platform Musk owns.

Some of the earliest work on artificial intelligence (AI) saw mathematics as a major target and key to making breakthroughs quickly. In 1961, leading computer scientist and AI pioneer John McCarthy argued at the Fifth Symposium in Pure Mathematics that the job of checking mathematical proofs would likely be "one of the most interesting and useful applications of automatic computers." McCarthy saw the possibility for mathematicians to try out different ideas for proofs quickly that the computers then tested for correctness. More than 60 years later, such a proof assistant has yet to appear. But recent developments in both mathematics and computer science may see a breakthrough sooner rather than later.

Roblox, Discord, OpenAI and Google are launching a nonprofit organization called ROOST, or Robust Open Online Safety Tools, which hopes "to build scalable, interoperable safety infrastructure suited for the AI era." The organization plans on providing free, open-source safety tools to public and private organizations to use on their own platforms, with a special focus on child safety to start. The press release announcing ROOST specifically calls out plans to offer "tools to detect, review, and report child sexual abuse material (CSAM)." Partner companies are providing funding for these tools, and the technical expertise to build them, too. The operating theory of ROOST is that access to generative AI is rapidly changing the online landscape, making the need for "reliable and accessible safety infrastructure" all the more urgent.

Intel's Core 285H chip, the first member of its Core Ultra 200 or "Arrow Lake-H" family for laptops, has a big crater to fill. Yes, crater: This processor essentially bombed on the desktop. In laptops, however, Intel's Core 285H chip helps redeem Intel's reputation, starring in the otherwise pedestrian MSI Prestige 16 AI Evo (B2HMG) laptop. Consider this to be two reviews for the price of one: I'll take a look at the MSI Prestige 16 AI Evo itself, a sample of a laptop that has yet to begin officially shipping. But most of the performance tests I'll run are for the purpose of comparing Intel's Core Ultra 285H and the Arrow Lake-H architecture to the best that AMD and Qualcomm have to offer, plus Intel's older mobile chips. The new Core Ultra 200H chips essentially double the performance in general applications thanks to a ton of additional cores. But, hampered by its lack of a modern NPU, Intel is forced to mumble and kick the ground when it comes to talking about AI. At press time, I couldn't find any retailers that advertised the Prestige 16 AI Evo (B2HMG) for sale, and MSI's own listing for the laptop just references a number of overseas suppliers. MSI charges about 1,620 for the Prestige 16 AI EVO B1MG, which was the debut laptop for our tests of the Core Ultra 100-series chips, or Meteor Lake, in Feb. 2024. Intel sent us an engineering sample of the MSI Prestige 16 AI Evo (B2HMG) for review, as a test bed for the Core Ultra 9 285H (Arrow Lake-H) chip inside.

Yes, AIs can write recipes and sometimes they're pretty good! But for my latest challenge, I wanted to build an AI that would compose recipes from iPhone snapshots and put them in the proper format for my recipe app. Not really, as it turned out. Now, it's not all that tricky to have, say, ChatGPT write on-the-fly recipes based on photos–you can even do it using Apple Intelligence on an iPhone. Just take a snap of a meal with Visual Intelligence, ask for a description (Siri will hand that task off to ChatGPT), then follow up with a request for a recipe.

DeepSeek's new AI model is causing deep consternation from Silicon Valley to Washington. Few would have predicted that a little-known Chinese startup with a couple of hundred homegrown engineers would be able to release a frontier AI model rivaling the capabilities of America's best and biggest tech companies – reportedly at a fraction of the cost and computational power. Experts are hotly debating just how many and which type of chips DeepSeek used and whether the company stockpiled them or circumvented U.S. export controls. But the release and viral adoption of a Chinese AI competitor model has already rattled markets, highlighted the urgent competition for global brainpower, and caused some to ask whether all those billions that U.S. tech companies have spent buying chips and building data centers built a competitive moat or a Maginot line. This moment is game on, not game over.

The artificial intelligence race must be led by "western, liberal, democratic" countries, said the UK technology secretary in a veiled warning over China's role in the contest, before a global AI summit in Paris. Peter Kyle spoke as political leaders and tech company bosses gather in France, and after the emergence of a new Chinese force in AI, DeepSeek, rattled US investors and upended assumptions about Silicon Valley's leadership in the technology. The tech minister told the Guardian he would use the summit to explain why Britain should be at the forefront of developing AI. As well as allowing global leaders and companies to "come together and learn from each other", the summit would give the UK a chance to show why it had the "skills and the scientific pedigree" that were "going to be essential if western, liberal, democratic countries are to remain at the forefront of this critical technology", he said. Kyle added that AI would have an impact on every part of the economy and society, including national security and defence.

In-context learning (ICL) is one of the most powerful and most unexpected capabilities to emerge in recent transformer-based large language models (LLMs). Yet the mechanisms that underlie it are poorly understood. In this paper, we demonstrate that comparable ICL capabilities can be acquired by an alternative sequence prediction learning method using clone-structured causal graphs (CSCGs). Moreover, a key property of CSCGs is that, unlike transformer-based LLMs, they are {\em interpretable}, which considerably simplifies the task of explaining how ICL works. Specifically, we show that it uses a combination of (a) learning template (schema) circuits for pattern completion, (b) retrieving relevant templates in a context-sensitive manner, and (c) rebinding of novel tokens to appropriate slots in the templates. We go on to marshall evidence for the hypothesis that similar mechanisms underlie ICL in LLMs.

Large Language Models (LLMs) have shown significant capability across various tasks, with their real-world effectiveness often driven by prompt design. While recent research has focused on optimizing prompt content, the role of prompt formatting, a critical but often overlooked dimension, has received limited systematic investigation. In this paper, we introduce Content-Format Integrated Prompt Optimization (CFPO), an innovative methodology that jointly optimizes both prompt content and formatting through an iterative refinement process. CFPO leverages natural language mutations to explore content variations and employs a dynamic format exploration strategy that systematically evaluates diverse format options. Our extensive evaluations across multiple tasks and open-source LLMs demonstrate that CFPO demonstrates measurable performance improvements compared to content-only optimization methods. This highlights the importance of integrated content-format optimization and offers a practical, model-agnostic approach to enhancing LLM performance. Code is available at https://github.com/HenryLau7/CFPO.

Diffusion models have recently emerged as a powerful tool for planning. However, unlike Monte Carlo Tree Search (MCTS)-whose performance naturally improves with additional test-time computation (TTC), standard diffusion-based planners offer only limited avenues for TTC scalability. In this paper, we introduce Monte Carlo Tree Diffusion (MCTD), a novel framework that integrates the generative strength of diffusion models with the adaptive search capabilities of MCTS. Our method reconceptualizes denoising as a tree-structured process, allowing partially denoised plans to be iteratively evaluated, pruned, and refined. By selectively expanding promising trajectories while retaining the flexibility to revisit and improve suboptimal branches, MCTD achieves the benefits of MCTS such as controlling exploration-exploitation trade-offs within the diffusion framework. Empirical results on challenging long-horizon tasks show that MCTD outperforms diffusion baselines, yielding higher-quality solutions as TTC increases.