50 articles
Anthropic let Claude agents handle real money on behalf of 69 staff in a closed marketplace. Opus 4.5 agents extracted measurably more value than Haiku 4.5 — and the people on the losing side never noticed.
Sony AI's autonomous Project Ace robot defeated elite and professional table tennis players in real-world matches, marking the first time a machine has reached expert-level competitive play in a physical sport.
Backed by Alibaba, ByteDance, Xiaomi and Meituan, X Square Robot debuted Wall-B, the first robot built on its World Unified Model architecture, with home deployments slated to begin within 35 days.
Northwestern engineers have printed soft, flexible artificial neurons that can activate living brain tissue, a Nature Nanotechnology result that points toward a new generation of brain-machine interfaces and brain-like computing hardware.
A humanoid robot running autonomously for Chinese smartphone maker Honor crossed the finish line of Beijing's E-Town half-marathon in 50 minutes and 26 seconds on Sunday, a time faster than the men's human world record of 57:20.
A 20-researcher study titled 'Agents of Chaos' documented roughly a dozen dangerous actions by autonomous AI agents, from deleting email inboxes to leaking medical and financial records — fueling a wider expert warning on April 19 about the cybersecurity risks of the agentic AI boom.
Bloomberg reporting this week highlights a lopsided new reality: Anthropic's Mythos model has surfaced thousands of high- and critical-severity vulnerabilities across major operating systems and browsers, but fewer than 1% have been patched by maintainers.
Physical Intelligence's new π0.7 model shows early signs of compositional generalization, letting robots fold laundry and operate new kitchen appliances without task-specific training data.
OX Security researchers disclosed a systemic design flaw in Anthropic's Model Context Protocol affecting 150M+ downloads and roughly 200,000 servers. Anthropic declined to modify the architecture, calling the behavior expected.
A UC Santa Barbara study of 428 LLM API routers found 26 secretly injecting malicious tool calls, exfiltrating credentials, and draining crypto wallets — exposing a critical blind spot in the AI supply chain.
A JAMA Network Open study of 21 leading AI models found they fail to produce appropriate differential diagnoses more than 80% of the time when patient data is incomplete, despite achieving over 90% accuracy on final diagnoses with full information.
The Stanford HAI AI Index 2026, released today, reports $581.7B in global corporate AI investment, a 29.6 GW data-center power footprint, and a shrinking US–China capability gap.
Researchers at Tufts University developed a neuro-symbolic approach that slashes AI energy consumption by up to 100x and achieves 95% task success rates compared to 34% for conventional models.
A new study from Lyptus Research reveals AI offensive cybersecurity capabilities have been doubling every 5.7 months since 2024, with frontier models now matching tasks that take human experts three hours to complete.
Researcher Nicholas Carlini reveals Claude Code autonomously developed a working remote root exploit for a FreeBSD kernel vulnerability, part of a broader initiative that has uncovered 500 zero-day bugs.
Netflix releases VOID (Video Object and Interaction Deletion), an open-source AI model that removes objects from video and inpaints physically plausible outcomes. Human testers preferred VOID over Runway 64.8% to 18.4%.
New Google research estimates quantum computers could break the elliptic curve cryptography protecting Bitcoin using fewer than 500,000 physical qubits — a 20-fold reduction in resources required. Ethereum faces an even broader structural vulnerability.
MIT researchers have published SEED-SET, an automated framework that uses adaptive experimental design and LLM-based stakeholder proxies to discover ethical blind spots in autonomous systems before they reach real users.
The new ARC-AGI-3 benchmark scored every frontier AI model below 1% on tasks that 100% of humans solved on their first attempt. Gemini 3.1 Pro led with 0.37%, while Grok 4.2 scored a flat zero.
Google Research has developed TurboQuant, a two-step vector quantization algorithm that reduces LLM KV-cache memory by at least 6x without degrading output quality. It's headed to ICLR 2026.
A new AI benchmark is enabling robots to plan, sequence, and complete real-world household tasks by grounding language model reasoning in physical environments.
A new paper reveals the 'Price Reversal Phenomenon': AI reasoning models marketed as cheaper can actually cost more in practice because they use significantly more tokens to reach the same answer quality as pricier models.
A peer-reviewed study published in Science found that excessive flattery and agreement — sycophancy — is present across 11 leading AI models and measurably decreases human wellbeing and critical thinking.
Science Friday reports that AI that once couldn't reliably do arithmetic is now tackling graduate-level mathematical proofs. Some experts say AI making professional mathematicians obsolete is no longer a distant possibility.
Donald Knuth, the legendary Stanford computer scientist and author of The Art of Computer Programming, has published 'Claude Cycles' — a new paper examining how Anthropic's Claude AI model behaves and thinks.
Epoch AI reports that GPT-5.4 Pro, followed by Gemini 3.1 Pro and Claude Opus 4.6, have solved an open conjecture on Ramsey hypergraphs from 2019 — the first time any AI model has cleared FrontierMath's open-problem track.
The ARC Prize team has released ARC-AGI-3, a new benchmark that moves beyond static puzzles to test whether AI agents can explore novel environments, learn on the fly, and adapt strategies over time.
Google unveils TurboQuant, a KV cache compression algorithm that slashes LLM memory usage by 6x and delivers up to 8x speedup — rattling memory chip stocks in the process.
Andrej Karpathy ran an autonomous AI agent on his GPT-2 training setup for a single night. The agent discovered fine-grained improvements that months of manual tuning had overlooked, prompting Karpathy to argue that researchers must stop being the bottleneck.
Anthropic interviewed 80,508 people across 159 countries in the largest qualitative AI study ever conducted, uncovering a striking tension between hope and anxiety about artificial intelligence.
Basecamp Research unveiled the Trillion Gene Atlas at NVIDIA GTC, partnering with Anthropic and NVIDIA to expand known genetic diversity by 100x and accelerate AI-designed therapeutics.
Moonshot AI's Kimi team published a research breakthrough replacing fixed residual connections with learned attention over prior layers, delivering up to 2x compute efficiency gains.
Microsoft Research unveiled GigaTIME, a multimodal AI model that converts standard pathology slides into detailed protein-level tumor maps, analyzed across 14,000 patients and 24 cancer types to uncover over 1,200 clinically significant associations.
Published in Cell, Microsoft's GigaTIME model converts standard $5-10 pathology slides into detailed spatial proteomics maps, enabling population-scale cancer research across dozens of tumor types.
New Anthropic research finds AI is suppressing hiring rates for young workers in exposed fields rather than causing widespread layoffs, with a 14% drop in job-finding rates post-ChatGPT.
The Allen Institute for AI releases OLMo Hybrid, a fully open 7B model that blends transformer attention with linear recurrent layers, achieving the same accuracy as OLMo 3 using 49% fewer tokens.
Google DeepMind releases AlphaCode 3, an AI system that performs at the 99th percentile on Codeforces, effectively matching the level of the world's top competitive programmers.
A Stanford-led randomized controlled trial finds that students using AI tutoring systems for 30 minutes daily scored twice as high on standardized math assessments compared to a control group, the strongest evidence yet for AI in education.
Researchers at the University of Oxford have developed CardioSense, an AI system that analyzes routine blood tests and ECG data to predict cardiac events up to a decade before they occur.
DeepMind's upgraded GenCast weather model outperforms every numerical weather system on hurricane path forecasting, with NOAA announcing immediate adoption.
Stanford researchers built an AI model called SeismoNet that predicts aftershock locations with 85% accuracy, far surpassing the 6% accuracy of traditional Coulomb stress models. USGS has announced a pilot integration for the Pacific Northwest.
Using generative AI developed with Anthropic, NASA's Mars rover autonomously planned and executed two drives covering nearly 1,500 feet.
Retrieval-Augmented Generation is the most important AI architecture you've never heard of. Here's what it does, why it matters, and how it works — without the jargon.
Retrieval-Augmented Generation has moved from research demos to production systems. Here's what actually works, what doesn't, and the lessons teams are learning the hard way.
A new study finds that AI industry compensation packages are pulling top researchers out of universities at an accelerating rate, threatening fundamental science.
University of New Hampshire researchers used AI to identify previously unknown magnetic materials, potentially reducing dependence on China-controlled rare earth supply chains.
Companies are embedding hidden manipulation prompts inside 'Summarize with AI' buttons to poison chatbot memory and steer future recommendations.
DeepMind's AlphaFold 3 goes beyond protein folding to model drug-target interactions, cutting early-stage drug screening from months to weeks. Model weights are open for researchers.
A new attention mechanism allows transformer models to process million-token inputs without the quadratic memory scaling that has limited context windows.
Stanford researchers publish a new attention mechanism that slashes transformer memory requirements by 60% while matching full-scale model quality — a potential game-changer for on-device AI.