Understanding the curvature evolution of the loss landscape is fundamental to analyzing the training dynamics of neural networks. The most commonly studied measure, Hessian sharpness ($λ_{\max}^H$) -- the largest eigenvalue of the loss Hessian -- determines local training stability and interacts with the learning rate throughout training. Despite its significance in analyzing training dynamics, direct measurement of Hessian sharpness remains prohibitive for Large Language Models (LLMs) due to high computational cost. We analyze $\textit{critical sharpness}$ ($λ_c$), a computationally efficient measure requiring fewer than $10$ forward passes given the update direction $Δ\mathbfθ$. Critically, this measure captures well-documented Hessian sharpness phenomena, including progressive sharpening and Edge of Stability. Using this measure, we provide the first demonstration of these sharpness phenomena at scale, up to $7$B parameters, spanning both pre-training and mid-training of OLMo-2 models. We further introduce $\textit{relative critical sharpness}$ ($λ_c^{1\to 2}$), which quantifies the curvature of one loss landscape while optimizing another, to analyze the transition from pre-training to fine-tuning and guide data mixing strategies. Critical sharpness provides practitioners with a practical tool for diagnosing curvature dynamics and informing data composition choices at scale. More broadly, our work shows that scalable curvature measures can provide actionable insights for large-scale training.

Language diffusion models aim to improve sampling speed and coherence over autoregressive LLMs. We introduce Neural Flow Diffusion Models for language generation, an extension of NFDM that enables the straightforward application of continuous diffusion models to discrete state spaces. NFDM learns a multivariate forward process from the data, ensuring that the forward process and generative trajectory are a good fit for language modeling. Our model substantially reduces the likelihood gap with autoregressive models of the same size, while achieving sample quality comparable to that of previous latent diffusion models.

Altman's campaigning for his company coincides with its use of enormous present resources to serve an imagined future OpenAI CEO Sam Altman poses during the Artificial Intelligence (AI) Action Summit, at the Grand Palais, in Paris, on February 11, 2025. Sam Altman has claimed over the years that the advancement of AI could solve climate change, cure cancer, create a benevolent superintelligence beyond human comprehension, provide a tutor for every student, take over nearly half of the tasks in the economy and create what he calls "universal extreme wealth". In order to bring about his utopian future, Altman is demanding enormous resources from the present. As CEO of OpenAI, the world's most valuable privately owned company, he has in recent months announced plans for $1tn of investment into datacenters and struck multibillion-dollar deals with several chipmakers. If completed, the datacenters are expected to use more power than entire European nations .

Tests conducted by the Guardian show that GPT-5.2 sourced some of its info from the AI-generated online encyclopedia from Elon Musk's xAI. OpenAI may have called GPT-5.2 its most advanced frontier model for professional work, but tests conducted by the cast doubt on its credibility. According to the report, OpenAI's GPT-5.2 model cited Grokipedia, the online encyclopedia powered by xAI, when it came to specific, but controversial topics related to Iran or the Holocaust. As seen in the's report, ChatGPT used Grokipedia as a source for claims about the Iranian government being tied to telecommunications company MTN-Irancell and questions related to Richard Evans, a British historian who served as an expert witness during a libel trial for Holocaust denier David Irving. However, the noted ChatGPT didn't use Grokipedia when it came to a prompt asking about media bias against Donald Trump and other controversial topics. A study done by US researchers also showed that the AI-generated encyclopedia cited questionable and problematic sources.

Microsoft Copilot security vulnerability called "Reprompt" discovered by Varonis researchers allowed attackers to hijack Copilot sessions through malicious links.

ChatGPT cited Grokipedia when repeating information that the Guardian has debunked. ChatGPT cited Grokipedia when repeating information that the Guardian has debunked. Guardian found OpenAI's platform cited Grokipedia on topics including Iran and Holocaust deniers The latest model of ChatGPT has begun to cite Elon Musk's Grokipedia as a source on a wide range of queries, including on Iranian conglomerates and Holocaust deniers, raising concerns about misinformation on the platform. In tests done by the Guardian, GPT-5.2 cited Grokipedia nine times in response to more than a dozen different questions. These included queries on political structures in Iran, such as salaries of the Basij paramilitary force and the ownership of the Mostazafan Foundation, and questions on the biography of Sir Richard Evans, a British historian and expert witness against Holocaust denier David Irving in his libel trial.

Technology AI This AI thinks it's the 1800s What happens when you train an LLM only on limited historical data? Breakthroughs, discoveries, and DIY tips sent six days a week. An interesting thing about contemporary artificial intelligence models, specifically large language models (LLMs): They can only output text based on what's in their training dataset. Models, including ChatGPT and Claude, are "trained" on large databases of text. The models, when asked a question, statistically create a response by calculating, one word at a time, what the most likely next word should be.

Researchers have succeeded in developing the smallest fully autonomous robot in history. It measures less than 1 millimeter and can swim underwater for months powered only by light. Miniaturization has long been a challenge in the history of robotics . While engineers have made great strides in the miniaturization of electronics in the past few decades, builders of miniature autonomous robots have not been able to meet the goal of getting them under 1 millimeter in size. This is because small arms and legs are fragile and difficult to manufacture.

On this episode of, our hosts unpack the news from Davos, where Trump and major AI companies shared the stage at the World Economic Forum. This week, WIRED's Brian Barrett and Leah Feiger are joining the show as the new cohosts, alongside Zoë Schiffer. And our attention has been drawn to the drama going down in the quaint little town of Davos. Zoë tells us how at the World Economic Forum's event, major AI players like Anthropic have been the protagonists--sharing the spotlight with President Donald Trump, who insists on invading Greenland. Brian has been looking at how ICE activity is developing, and Leah is forcing us to think about this year's midterms because tech giants are already pouring millions into it. Plus, we dive into why OpenAI's decision to roll out ads in ChatGPT was a long time coming. Ads Are Coming to ChatGPT. Write to us at uncannyvalley@wired.com . You can always listen to this week's podcast through the audio player on this page, but if you want to subscribe for free to get every episode, here's how: If you're on an iPhone or iPad, open the app called Podcasts, or just tap this link . Today, we're starting a bit of a new chapter here on the show, and I want to introduce you to my brand new cohost, Brian Barrett, our executive editor here at WIRED, and Leah Feiger, our senior politics editor. So thrilled to be here. So longtime listeners know the show has taken on a bunch of different formats since it launched. We had the Gadget Lab days, the roundtable, news episodes. We really created this podcast because we want to bring you the best stories and the best takes about what's happening in tech and politics. That's all going to stay the same, but this time we're going to go even deeper. What trends you should be watching for, the news that's already happened or about to break, and how we are thinking about all of it.

TikTok Is Now Collecting Even More Data About Its Users. According to its new privacy policy, TikTok now collects more data on its users, including their precise location, after majority ownership officially switched to a group based in the US. When TikTok users in the US opened the app today, they were greeted with a pop-up asking them to agree to the social media platform's new terms of service and privacy policy before they could resume scrolling. These changes are part of TikTok's transition to new ownership. In order to continue operating in the US, TikTok was compelled by the US government to transition from Chinese control to a new, American-majority corporate entity.