News/Research
Hugging Face cofounder poopoos AI science pep, says AI can’t ask the right questions
Hugging Face cofounder Thomas Wolf is challenging the tech industry's optimistic predictions about AI's potential to revolutionize scientific discovery. Speaking at VivaTech in Paris, Wolf argued that current large language models excel at finding answers but lack the creativity to ask original scientific questions—a critical limitation that may produce digital "yes-men" rather than breakthrough discoveries. What they're saying: Wolf believes the fundamental challenge lies in AI's inability to challenge existing frameworks of knowledge. "In science, asking the question is the hard part, it's not finding the answer," Wolf told Fortune. "Once the question is asked, often the answer is quite...
read Jun 24, 2025AI program Xbow becomes top US vulnerability researcher, finding 1,000+ bugs
An AI program called Xbow has become the top-ranked vulnerability researcher in the United States on HackerOne, a platform that coordinates software bug discoveries with major companies. The achievement marks a significant milestone in automated cybersecurity, as Xbow has outperformed human researchers by discovering over 1,000 software flaws across companies including Disney, AT&T, Ford, and Epic Games. What you should know: Xbow has submitted nearly 1,060 vulnerability reports in recent months, with 132 officially confirmed and resolved by affected companies. An additional 303 vulnerabilities were classified as "triaged," meaning they've been acknowledged but not yet fixed, while 125 remain under...
read Jun 24, 2025Anti-War Games: Anadyr Horizon’s AI simulates world leaders to forecast and prevent conflict
Anadyr Horizon is using predictive AI to forecast and prevent global conflicts through its "peace tech" platform called North Star. The startup, founded in 2024, creates digital twins of world leaders and runs thousands of simulations to predict how they might react to real-world scenarios like economic sanctions or military blockades, with clients already including government agencies and corporate risk managers. Why this matters: Violent conflict cost the global economy an estimated $19 trillion in 2023, and Anadyr's approach represents a departure from traditional defense tech by focusing on conflict prevention rather than warfare capabilities. • The company emerges as...
read Jun 23, 2025Autodidactic tactic: MIT researchers create AI that teaches itself by generating training data
Researchers at MIT have developed SEAL (Self-Adapting Language Models), a framework that enables large language models to continuously learn and adapt by generating their own training data and update instructions. This breakthrough addresses a critical limitation in current AI systems, allowing models to permanently absorb new knowledge rather than relying on temporary retrieval methods—a capability that could transform enterprise AI applications where agents must constantly evolve in dynamic environments. How it works: SEAL uses reinforcement learning to train LLMs to generate "self-edits"—natural-language instructions that specify how the model should update its own weights. The framework operates on a two-loop system...
read Jun 23, 2025Study: AI could actually cut global emissions by 5.4B tons annually by 2035
New research suggests artificial intelligence could cut global climate pollution by up to 5.4 billion metric tons annually by 2035—far exceeding the emissions created by AI's own energy-hungry data centers. The findings from the Grantham Research Institute challenge growing concerns about AI's environmental impact, showing that strategic deployment across transportation, energy, and food systems could deliver net climate benefits equivalent to eliminating all of the European Union's current emissions. The big picture: AI's climate potential hinges on governments channeling the technology toward high-impact applications rather than letting market forces alone dictate its development. The study identifies five key areas where...
read Jun 19, 2025Am I hearing this right? AI system detects Parkinson’s disease from…ear wax, with 94% accuracy
Researchers at the American Chemical Society have developed an AI-powered system that can detect Parkinson's disease by analyzing the smell of ear wax with 94% accuracy. The breakthrough could provide an inexpensive, early-detection screening tool for a condition that affects 1.1 million Americans and kills tens of thousands annually, offering hope for earlier intervention when treatments are most effective. How it works: The detection system relies on identifying specific volatile organic compounds in sebum, the oily substance that makes up ear wax. Researchers swabbed the ear canals of 209 participants, more than half of whom had been diagnosed with Parkinson's...
read Jun 19, 2025Professional dog walking looking better as Harvard study finds AI threatens jobs in creative, white-collar sector
Harvard Business Review researchers have published a comprehensive analysis examining how current AI capabilities are poised to disrupt virtually every sector of the labor market, from creative industries to professional services. The study reveals that existing AI models—and the more advanced, cost-effective versions already in development—pose immediate threats to jobs across writing, design, finance, law, medicine, and academia by delivering comparable quality at dramatically reduced costs. What you should know: AI's disruptive potential extends far beyond theoretical scenarios, targeting both creative and analytical professions with current technology. Creative professionals including writers, designers, photographers, architects, animators, and brand advertisers face direct...
read Jun 19, 2025Always on-call? Microsoft study finds workers face 271 daily messages in “infinite workday”
The modern workplace has developed a peculiar problem: work never actually ends. Microsoft's latest research reveals a troubling reality where the traditional 9-to-5 workday has morphed into an always-on digital treadmill that leaves employees exhausted and organizations less productive than ever. The tech giant's new report, "Breaking down the infinite workday," analyzes how millions of people use Microsoft 365 products like Outlook, Teams, and Office applications. The findings paint a stark picture of professional life in 2025—one where workers juggle 271 daily messages across email and chat platforms, attend meetings during their most productive hours, and check email well past...
read Jun 19, 2025AI lacks art of the deal? Stanford study finds AI agents with weaker negotiation skills lose millions
A new Stanford study reveals that AI agents have wildly different negotiation skills, with weaker agents consistently losing money to stronger ones in automated transactions. The research highlights critical risks as companies increasingly deploy AI agents for everything from retail purchases to supply chain negotiations, potentially creating unfair advantages for those with superior technology. What you should know: The study found that AI agent negotiations create "an inherently imbalanced game" where capability gaps lead to significant financial losses.• Buyers using weaker agents paid around 2% more in retail price negotiations compared to scenarios with equally capable agents.• Weaker seller agents...
read Jun 19, 2025Why AI researchers are ditching mega-models for Minsky’s multi-agent approach
Marvin Minsky's 1986 book "The Society of Mind" is finding new relevance in 2025 as AI researchers increasingly embrace modular, multi-agent approaches over monolithic large language models. The theory, which proposes that intelligence emerges from collections of simple "agents" rather than a single unified system, now maps directly onto current AI architectures like Mixture-of-Experts models and multi-agent frameworks such as HuggingGPT and AutoGen. Why this matters: As the AI field hits the limits of scaling single massive models, Minsky's vision offers a blueprint for building more robust, scalable, and aligned AI systems through modularity and internal oversight mechanisms. The core...
read Jun 18, 2025MIT breakthrough makes gallium nitride chips affordable for everyday electronics
MIT researchers have developed a breakthrough fabrication process that integrates high-performance gallium nitride transistors onto standard silicon chips using a low-cost, scalable method compatible with existing semiconductor foundries. This innovation could significantly improve the speed and energy efficiency of electronics ranging from smartphones to quantum computers by combining the best properties of both materials without the prohibitive costs typically associated with gallium nitride integration. Why this matters: Gallium nitride is the second most widely used semiconductor after silicon, but its high cost and specialized integration requirements have limited commercial adoption despite superior performance characteristics for high-speed communications and power electronics....
read Jun 18, 2025Alphabet’s Isomorphic Labs expands to Cambridge with $3B in pharma deals
Isomorphic Labs, an Alphabet spinoff focused on AI-driven drug discovery, is establishing its first U.S. office in Cambridge, Massachusetts, as the company prepares to advance its drug candidates into clinical trials. The move positions the London-based company closer to American pharmaceutical partners and clinical development infrastructure, marking a significant milestone in its evolution from AI research to therapeutic development. What you should know: Isomorphic Labs built its platform on DeepMind's AlphaFold protein structure prediction technology and has secured major pharmaceutical partnerships worth billions in potential value. The company raised $600 million in external funding in March, led by Thrive Capital,...
read Jun 17, 2025MIT research reveals safety, consumer cost concerns in automated driving after 10 years of study
MIT's Advanced Vehicle Technology Consortium celebrated its 10th anniversary, marking a decade of collaboration between academia and industry to advance automotive technology through data-driven research. The consortium has collected hundreds of terabytes of data on driver behavior with sophisticated vehicle features, positioning itself as a global influencer in the automotive industry while addressing critical challenges including consumer trust, safety regulation, and the complexity of automated driving systems. What you should know: The AVT Consortium brings together over 25 member organizations including automotive manufacturers, suppliers, and insurers to study real-world driver interactions with advanced vehicle technologies. The consortium has focused on...
read Jun 16, 2025Apple’s AI reasoning study sparks fierce debate over flawed testing methods
Apple's machine-learning research team ignited a fierce debate in the AI community with "The Illusion of Thinking," a 53-page paper arguing that reasoning AI models like OpenAI's "o" series and Google's Gemini don't actually "think" but merely perform sophisticated pattern matching. The controversy deepened when a rebuttal paper co-authored by Claude Opus 4 challenged Apple's methodology, suggesting the observed failures stemmed from experimental flaws rather than fundamental reasoning limitations. What you should know: Apple's study tested leading reasoning models on classic cognitive puzzles and found their performance collapsed as complexity increased. Researchers used four benchmark problems—Tower of Hanoi, Blocks World,...
read Jun 13, 2025AI safety researcher at LessWrong ponders, “When should humans quit the field?”
A researcher on the AI safety forum LessWrong is questioning when humans should abandon technical AI safety careers as AI systems become capable of conducting their own safety research. The post explores whether continued human training in alignment research makes sense if AI will soon outpace human capabilities in this field, potentially rendering human expertise obsolete within months or years. The central premise: The author assumes it's possible to create "SafeAlignmentSolver-1.0"—an AI system that can safely and effectively conduct alignment research at a scale that makes human efforts redundant. Key considerations for career decisions: Several factors should influence whether aspiring...
read Jun 13, 2025AstraZeneca signs $5B+ AI drug deal to rebuild China operations
AstraZeneca has signed a research agreement worth more than $5 billion with Chinese drugmaker CSPC Pharmaceutical Group, marking the Anglo-Swedish company's latest effort to rebuild its business in China following recent challenges including the arrest of its China president and potential import fines. The collaboration will focus on AI-driven drug discovery for chronic diseases, positioning AstraZeneca to strengthen its foothold in its second-largest market while leveraging China's growing pharmaceutical research capabilities. What you should know: The partnership establishes a comprehensive framework for discovering and developing pre-clinical candidates targeting chronic diseases. AstraZeneca will pay CSPC an upfront fee of $110 million...
read Jun 12, 2025AI decodes sperm whale language revealing complex communication patterns
Researchers at the CETI project are using artificial intelligence to decode sperm whale communication, revealing that these marine mammals possess a far more sophisticated language system than previously understood. The breakthrough suggests that sperm whales may be the first non-human species to demonstrate "duality of patterning"—a linguistic complexity once thought to be uniquely human—potentially revolutionizing our understanding of animal intelligence and communication. What they discovered: CETI researchers have identified 21 distinct types of "codas" or call systems in sperm whale communication that show remarkable complexity. The team found that sperm whales can produce 10 times more meanings than previously believed,...
read Jun 11, 2025Google names first chief AI architect to bridge research-product gap
Google has appointed Koray Kavukcuoglu, Google DeepMind's chief technology officer, as its first chief AI architect, a new senior vice president role reporting directly to CEO Sundar Pichai. The position reflects Google's push to better translate its leading AI research capabilities into successful consumer products, as the company faces mounting pressure to compete with rivals like OpenAI while maintaining its dominant search business. What you should know: Kavukcuoglu will oversee how Google integrates its Gemini AI models across the company's product portfolio while maintaining his current CTO role at DeepMind. The former aerospace engineer joined DeepMind in 2012 as a...
read Jun 11, 2025Beyond the screen: MIT student creates AI that restores paintings with polymer film “masks”
MIT mechanical engineering graduate student Alex Kachkine has developed a revolutionary method to physically restore damaged paintings using AI-generated polymer film "masks" that can be applied directly to original artworks. The breakthrough technique completed a restoration of a 15th-century oil painting in just 3.5 hours—approximately 66 times faster than traditional methods—while maintaining a digital record of all changes for future conservators. How it works: The method combines AI analysis with physical application through a sophisticated multi-step process. First, conservators clean the original painting and remove previous restoration attempts to reveal the underlying damage. The cleaned painting is scanned, and existing...
read Jun 11, 2025ETH Zurich’s badminton-playing robot suggests ESPN for physical AI isn’t coming soon
Scientists at ETH Zurich have developed ANYmal, a quadruped robot that can play badminton using AI-powered perception and movement skills. The robot represents a breakthrough in combining real-time visual processing with physical agility, though its performance against human players reveals significant limitations that highlight ongoing challenges in robotics. How it works: ANYmal resembles a miniature giraffe holding a badminton racket in its teeth, built on a 50-kilogram industrial platform originally designed for oil and gas applications by ANYbotics, an ETH Zurich spinoff company. The robot uses a stereoscopic camera (which creates depth perception like human eyes) for shuttlecock tracking and...
read Jun 10, 2025Huawei CEO admits chips lag US despite $25B annual R&D spend
Huawei CEO Ren Zhengfei has publicly acknowledged that the company's chips lag one generation behind U.S. competitors, marking the first official comments from Huawei leadership about their advanced chip manufacturing capabilities since U.S. export controls began in 2019. Despite this technological gap, Ren revealed that Huawei is deploying alternative strategies like cluster computing and mathematical optimization to bridge performance differences, while investing $25 billion annually in research and development. What you should know: Huawei is compensating for its chip technology deficit through innovative workaround solutions rather than direct competition. The company uses "mathematics to supplement physics, non-Moore's law to supplement...
read Jun 9, 2025MIT AI enables drones to adapt to wind without prior training
MIT researchers have developed an AI-powered control system that enables autonomous drones to automatically adapt to unpredictable disturbances like gusty winds without requiring advance knowledge of these conditions. The breakthrough system uses meta-learning to simultaneously learn from flight data and select the optimal adaptation algorithm, achieving 50% less trajectory tracking error than existing methods in simulations. How it works: The system replaces traditional control functions with a neural network that learns to approximate disturbances from just 15 minutes of flight data. Instead of relying on pre-programmed knowledge about wind patterns, the AI model automatically determines which optimization algorithm from the...
read Jun 9, 2025Hm, that right? AI companies fail to justify safety claims
AI companies are failing to provide adequate justification for their safety claims based on dangerous capability evaluations, according to a new analysis by researcher Zach Stein-Perlman. Despite OpenAI, Google DeepMind, and Anthropic publishing evaluation reports intended to demonstrate their models' safety, these reports largely fail to explain why their results—which often show strong performance—actually indicate the models aren't dangerous, particularly for biothreat and cyber capabilities. The core problem: Companies consistently fail to bridge the gap between their evaluation results and safety conclusions, often reporting strong model performance while claiming safety without clear reasoning. OpenAI acknowledges that "several of our biology...
read Jun 6, 202530 mathematicians met in secret to stump OpenAI. They (mostly) failed.
Thirty of the world's leading mathematicians convened at a secret meeting in Berkeley to test OpenAI's o4-mini reasoning model against professor-level mathematical problems. The AI stunned researchers by solving complex questions that would challenge even academic mathematicians, with some experts describing its capabilities as approaching "mathematical genius." What you should know: The clandestine mathematical gathering took place over two days in mid-May, with participants signing nondisclosure agreements and communicating only through Signal to prevent AI contamination of their test questions. Mathematicians competed to devise problems they could solve but would stump the AI, with a $7,500 reward for each question...
read