All Updates
Page 1723 of 1962
March 7, 2026
AI Inner Speech Speeds Up Learning
OIST researchers demonstrate that AI models with 'self-mumbling' internal dialogue and working memory achieve superior learning efficiency, multitask generalization, and adaptation on sparse data. Mimicking human inner speech, the method excels in complex tasks like sequence reversal. Findings published in Neural Computation.
Anthropic Finds Scant AI Labor Impact
Anthropic economists Maxim Massenkoff and Peter McCrory have reworked an AI 'damage yardstick' metric. Their analysis reveals AI is not eliminating jobs as much as experts predicted. Current data shows minimal labor market disruption.
March 6, 2026
Jammer Blocks AI Wearables, Physics Foils It
Deveillance's Spectre I jammer, developed by a recent Harvard grad, aims to block always-listening AI wearables and restore user control over pervasive surveillance. The device targets constant audio monitoring in daily life. However, the article argues physics will likely prevent it from working effectively.
Analysts: Don't Blame AI for Poor Jobs Yet
US economy lost 92,000 jobs in February, missing expectations of 50,000 gains and pushing unemployment to 4.4%. This downturn fueled fears that AI is driving higher unemployment. Analysts, however, advise against blaming AI at this stage.
Oracle, OpenAI Scrap Texas AI Data Center Expansion
Oracle and OpenAI have ended plans to expand a flagship AI data center in Abilene, Texas, due to stalled financing talks and OpenAI's evolving needs. Meta Platforms is now eyeing the site for lease from developer Crusoe, with Nvidia facilitating discussions. The development highlights the intricate challenges in constructing multi-billion-dollar AI infrastructure.
NEC, UTokyo, NTT Integrate AI Tech on 6G/IOWN
NEC, University of Tokyo, and NTT announced integrated technologies for AI agents on the 6G/IOWN platform to handle traffic surges from always-on AI. They optimized large-scale data communication and computation. A real-time AR support demo confirmed low latency and high accuracy.
Bilibili Profits First Time via AI Ads Boom
Bilibili reported 2025 first full-year GAAP profit of 11.9B RMB, reversing prior losses, fueled by ads (up 27% Q4), value-added services, and games. AI ads surged 180% YoY via AIGC and automation. DAU hit 1.13B, up 10%.
Apple Skips AI Arms Race, Guards User Gate
Apple avoids massive AI infrastructure spending like rivals Microsoft, Google, Meta, and Amazon, focusing instead on devices, user experience, and privacy. It partners with Google on Gemini models for cloud capabilities while prioritizing on-device AI via M-series chips across 2.5B devices. This retains control over user interfaces and long-term relationships.
TerraPower Wins Approval for Wyoming Nuclear Build
US NRC approved TerraPower, Bill Gates' nuclear firm, to construct a novel reactor next to an aging Wyoming coal plant. This marks the first new reactor construction permit in nearly a decade.
Oracle & OpenAI Scrap Texas Data Center Expansion
Oracle and OpenAI have abandoned expansion of their flagship AI data center in Texas due to financing issues and OpenAI's shifting demands. The plan was part of the Stargate joint venture with SoftBank. Trump announced the project in January with up to $500B investment for AI infrastructure.
Claude Users Surge Post-Pentagon Clash
Claude's mobile daily active users and new installs are rising despite US DoD dispute. Anthropic CEO Dario Amodei rejected gov use for mass surveillance or autonomous weapons. This led to supply chain risk label but hasn't slowed growth.
Xbox Needs 3 Changes to Return to Glory
Xbox faces challenges including 32% hardware revenue drop, Phil Spencer's departure, and speculation of Microsoft pivoting gaming to AI, but it's not ending. The article outlines three key changes needed for Xbox to regain its former glory.
LangChain Launches Deep Agents for Prod AI
LangChain CEO Harrison Chase argues better models need evolved 'harnesses' for production AI agents. Deep Agents, built on LangChain/LangGraph, enables planning, subagents, progress tracking for long-running tasks. It allows LLMs greater autonomy via loops, tools, and coherent execution.
45% Security Leads Work 6-Day Weeks in AI Era
45% of security leaders report workloads equivalent to six-day weeks. AI advancements are creating a turning point for their roles amid high mental pressure. The article explores organizational strategies to sustain motivation without added burdens.
Khan Invites Anthropic to London Amid US Risks
London Mayor Sadiq Khan has sent a letter inviting AI firm Anthropic to expand operations in the city. The invitation comes as the US moves to designate Anthropic a supply chain risk. This reflects London's ambition to attract AI companies amid global tensions.
ClickFix Shifts to Windows Terminal Evasion
Microsoft reveals ClickFix phishers now use Win+X shortcut to open Windows Terminal, bypassing Run dialog detections and employee training. Victims paste malicious PowerShell, triggering multi-stage attacks with 7-Zip extraction, persistence, and data theft. Experts say tactic predates but persists, urging awareness.
US Eyes Permits for Global Nvidia, AMD AI Chips
US Commerce Department drafts rules requiring permits for AI chip shipments worldwide without approval. Oracle plans thousands of job cuts amid cash crunch from AI data center expansion. Pentagon flags Anthropic and its products as US supply chain risk.
GitHub's AI Framework Scans Key Vulnerabilities
GitHub Security Lab released Taskflow Agent, an open-source AI-powered framework for vulnerability scanning. It excels at detecting high-impact issues like Auth Bypasses, IDORs, and Token Leaks. The GitHub Blog provides a tutorial on its usage.
MIT's Attention Matching Cuts KV Cache 50x
MIT researchers developed Attention Matching, a fast KV cache compression technique that reduces LLM memory usage up to 50x with minimal accuracy loss. It addresses the memory bottleneck in long-context enterprise AI applications like legal analysis and multi-session dialogues. The method outperforms eviction, merging, and summarization by preserving information effectively.
Grammarly Misuses Identities in Expert Reviews
Grammarly's 'expert review' feature generates AI writing advice 'inspired by' experts, including deceased professors and living Verge editors without permission. The feature, launched in August, surprised users by impersonating colleagues like Nilay Patel.