#large-language-models

#large-language-models

For the first time, speech has been decoupled from consequence. We now live alongside AI systems that converse knowledgeably and persuasively-deploying claims about the world, explanations, advice, encouragement, apologies, and promises-while bearing no vulnerability for what they say. Millions of people already rely on chatbots powered by large language models, and have integrated these synthetic interlocutors into their personal and professional lives. An LLM's words shape our beliefs, decisions, and actions, yet no speaker stands behind them.

Fifty-four seconds. That's how long it took Raphael Wimmer to write up an experiment that he did not actually perform, using a new artificial-intelligence tool called Prism, released by OpenAI last month. "Writing a paper has never been easier. Clogging the scientific publishing pipeline has never been easier," wrote Wimmer, a researcher in human-computer action at the University of Regensburg in Germany, on Bluesky. Large language models (LLMs) can suggest hypotheses, write code and draft papers, and AI agents are automating parts of the research process.

Scientists are increasingly turning to artificial-intelligence systems for help drafting the grant proposals that fund their careers, but preliminary data indicate that these tools might be pulling the focus of research towards safe, less-innovative ideas. These data provide evidence that AI-assisted proposals submitted to the US National Institutes of Health (NIH) are consistently less distinct from previous research than ones written without the use of AI - and are also slightly more likely to be funded.

Drawing on more than 22,000 LLM prompts designed to reflect the kind of questions people would ask artificial intelligence (AI)-powered chatbots, such as, "How do I apply for universal credit?", the data raises concerns about whether chatbots can be trusted to give accurate information about government services. The publication of the research follows the UK government's announcement of partnerships with Meta and Anthropic at the end of January 2026 to develop AI-powered assistants for navigating public services.

Vibe coding is a relatively new programming paradigm that emerged with the rise of AI-powered development tools. The term was coined by Andrej Karpathy, a prominent AI researcher and former Director of AI at Tesla, to describe an intuitive way of coding where developers interact with AI models using natural language commands rather than traditional coding syntax. Instead of meticulously writing every line of code, developers simply "vibe" with the AI, describing what they want, and letting the AI generate the necessary code.

ElevenLabs co-founder and CEO Mati Staniszewski says voice is becoming the next major interface for AI - the way people will increasingly interact with machines as models move beyond text and screens. Speaking at Web Summit in Doha, Staniszewski told TechCrunch voice models like those developed by ElevenLabs have recently moved beyond simply mimicking human speech - including emotion and intonation - to working in tandem with the reasoning capabilities of large language models.

In essence, Lotus is building an AI doctor that functions like a real medical practice, equipped with a license to operate in all 50 states, malpractice insurance, HIPAA-compliant systems, and full access to patient records. The key difference is that the majority of the work is done by AI, which is trained to ask the same questions a doctor would.

This process, becoming aware of something not working and then changing what you're doing, is the essence of metacognition, or thinking about thinking. It's your brain monitoring its own thinking, recognizing a problem, and controlling or adjusting your approach. In fact, metacognition is fundamental to human intelligence and, until recently, has been understudied in artificial intelligence systems. My colleagues Charles Courchaine, Hefei Qiu, Joshua Iacoboni, and I are working to change that.

United States Immigration and Customs Enforcement is leveraging Palantir's generative artificial intelligence tools to sort and summarize immigration enforcement tips from its public submission form, according to an inventory released Wednesday of all use cases the Department of Homeland Security had for AI in 2025. The "AI Enhanced ICE Tip Processing" service is intended to help ICE investigators "to more quickly identify and action tips" for urgent cases, as well as translate submissions not made in English, according to the inventory.

"The future will be, for sure, that you are not typing any data information into an SAP system. You can instead ask certain analytical questions with your voice. You can trigger operational task workflows. You can also make entries in the system with your voice-performance feedback, pipeline entries, etc. The technological capabilities are there, it really is now about the execution."

Anti-intelligence is not stupidity or some sort of cognitive failure. It's the performance of knowing without understanding. It's language severed from memory, context, and and even intention. It's what large language models (LLMs) do so well. They produce coherent outputs through pattern-matching rather than comprehension. Where human cognition builds meaning through the struggle of thought, anti- intelligence arrives fully formed.

Generalist models "fail miserably" at the benchmarks used to measure how AI performs scientific tasks, Alex Zhavoronkov, Insilico's founder and CEO, told Fortune. " You test it five times at the same task, and you can see that it's so far from state of the art...It's basically worse than random. It's complete garbage." Far better are specialist AI models that are trained directly on chemistry or biology data.

To work around those rules, the Humanizer skill tells Claude to replace inflated language with plain facts and offers this example transformation: Before: "The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain." After: "The Statistical Institute of Catalonia was established in 1989 to collect and publish regional statistics." Claude will read that and do its best as a pattern-matching machine to create an output that matches the context of the conversation or task at hand.

In this book, you will learn how to use artificial intelligence to create mini-games. You will attempt to recreate the look and feel of various classic video games. The intention is not to violate copyright or anything of the sort, but instead to learn the limitations and the power of AI. Instead, you will simply be learning about whether or not you can use AI to help you know how to create video games.

He's the perfect outsider figure: the eccentric loner who saw all this coming and screamed from the sidelines that the sky was falling, but nobody would listen. Just as Christian Bale portrayed Michael Burry, the investor who predicted the 2008 financial crash, in The Big Short, you can well imagine Robert Pattinson fighting Paul Mescal, say, to portray Zitron, the animated, colourfully obnoxious but doggedly detail-oriented Brit, who's become one of big tech's noisiest critics.

If you want to win in AI - and I mean win in the biggest, most lucrative, most shape-the-world-in-your-image kind of way - you have to do a bunch of hard things simultaneously. You need to have a model that is unquestionably one of the best on the market. You need the nearly infinite resources required to continue to improve that mode and deploy it at massive scale.

Tokamak fusion reactors rely on heated plasma that is extremely densely packed inside a doughnut-shaped chamber. But researchers thought that plasma could not exceed a certain density - a boundary called the Greenwald limit - without becoming unstable. In a new study, scientists pushed beyond this limit to achieve densities 30% to 65% higher than those normally reached by EAST while keeping the plasma stable.

In fact, when prompted strategically by researchers, Claude delivered the near-complete text of Harry Potter and the Sorcerer's Stone, The Great Gatsby, 1984, and Frankenstein, in addition to thousands of words from books including The Hunger Games and The Catcher in the Rye. Varying amounts of these books were also reproduced by the other three models. Thirteen books were tested.

Three major large language models (LLMs) generated responses that, in humans, would be seen as signs of anxiety, trauma, shame and post-traumatic stress disorder. Researchers behind the study, published as a preprint last month, argue that the chatbots hold some kind of "internalised narratives" about themselves. Although the LLMs that were tested did not literally experience trauma, they say, their responses to therapy questions were consistent over time and similar in different operatingmodes, suggesting that they are doing more than "role playing".

Researchers have developed a tool that they say can make stolen high-value proprietary data used in AI systems useless, a solution that CSOs may have to adopt to protect their sophisticated large language models (LLMs). The technique, created by researchers from universities in China and Singapore, is to inject plausible but false data into what's known as a knowledge graph (KG) created by an AI operator. A knowledge graph holds the proprietary data used by the LLM.

Meta has applied large language models to mutation testing to improve compliance coverage across its software systems. The approach integrates LLM-generated mutants and tests into Meta's Automated Compliance Hardening system (ACH), addressing scalability and accuracy limits of traditional mutation testing. The system is intended to keep products and services safe while meeting compliance obligations at scale, helping teams satisfy global regulatory requirements more efficiently.

Contextual integrity defines privacy as the appropriateness of information flows within specific social contexts, that is, disclosing only the information strictly necessary to carry through a given task, such as booking a medical appointment. According to Microsoft's researchers, today's LLMs lack this kind of contextual awareness and can potentially disclose sensitive information, thereby undermining user trust. The first approach focuses on inference-time checks, i.e., safeguards applied when a model generates its response.