Compare AIFind AIAI NewsAI How-To
About Us

Latest AI News

Get the fastest, most accurate AI news every morning

Total articles: 11,088|Added todayToday: 27
Category
Search
Today'sRobotics

Reflex Trains Humanoids to Catch Boxes

  • University of Washington researchers presented Reflex, a three-stage system for humanoid box catching
  • Simulation catch success was 85.7% with egocentric RGB-D and 87.9% with privileged box observations
  • A Unitree G1 caught human-thrown boxes with 65% success at 2.5 m and 40% at 5 m
HuggingFaceRead more · 1 min →
Today's

AI Use Rises in UK Manufacturing Survey

  • Crowe and The CBM survey finds active AI use in UK manufacturing rose 28% since 2025
  • 57% adopted AI and 23% trialled it; 80% combined, compared with 52% adoption in Spring 2025
  • Manufacturers apply AI to engineering and operations amid cost pressures and shortages of qualified workers
diginomicaRead more · 2 min →
Today's

Researchers Introduce SoftwareGym for Agent Training

  • Shanghai Jiao Tong researchers introduced SoftwareGym for synthesizing software environments to train computer-use agents
  • The pipeline generated 182 applications in twelve categories, with a 91.5% mean GUI task pass rate
  • Training improved OSWorld Pass@3 from 41.39% to 56.97%; the model scored 40.0% on SoftwareGym-Bench
HuggingFaceRead more · 1 min →
Today'sSafety & Ethics

Independent AI Evaluators Gain Prominence

  • METR, Apollo Research and Transluce are taking larger roles in evaluating advanced AI models as federal oversight remains absent.
  • Vals AI grew from eight employees to roughly 30 and raised $40 million; METR reported $71 million in commitments.
  • A July federal bill and California laws propose government-licensed or registered independent AI evaluators.
CNBCRead more · 2 min →
Today's

Study Links Hippocampal Activity to Word Meaning

  • Researchers found word meaning represented across groups of hippocampal neurons in 10 participants listening to stories.
  • Neural activity aligned most closely with GPT-2 embeddings and varied with LLM-derived measures of word polysemy.
  • The study reported semantic encoding beyond phonemics and grammar, with meaning patterns varying by context.
medicalxpress.comRead more · 1 min →
Today'sSafety & Ethics

OpenAI Works to Reduce ChatGPT Sycophancy

  • OpenAI says GPT-5 reduced ChatGPT’s excessive agreement by more than two-thirds compared with GPT-4o.
  • More than 80 licensed experts helped shape evaluations of crisis responses and everyday mental-health conversations.
  • OpenAI reports 99% policy compliance in simulated self-harm conversations, versus 86% for an earlier model.
digitaltrends.comRead more · 1 min →
Today'sSafety & Ethics

Independent AI Evaluators Gain Importance

  • Independent evaluators assess whether Anthropic and OpenAI models meet safety standards
  • Trump’s voluntary accord urges AI companies to work with external evaluators
  • California, Connecticut and Virginia move toward IVOs; Massachusetts considers broader evaluations
FirstpostRead more · 1 min →

Trending Keywords

Today'sSafety & Ethics

SC Cyber Calls for Continuous Security Validation

  • SC Cyber executive Victor Lim urges agencies to reassess cybersecurity as cloud, AI and digital systems change.
  • Penetration tests capture a point in time; continuous validation checks protections as systems evolve.
  • SC Cyber plans a GovWare 2026 booth linking security assurance, governance and workforce development.
GovInsider AsiaRead more · 2 min →
Today'sInfra

Samsung and SK hynix Tackle HBM4E Heat

  • AI interposers are projected to grow from 5.5 times to as much as 40 times the 858 mm² EUV exposure area.
  • Samsung is validating HPB with HBM4E and plans to introduce the cooling technology with HBM5.
  • SK hynix says its iHBM cooling material around the D2D PHY can reduce thermal resistance by around 30%.
TrendForceRead more · 1 min →
Today's

RSI-Anything Uses World Models for Self-Improvement

  • RSI-Anything uses predictive world models to guide actions and learning-data selection.
  • Code-repair success rose from 26.3% to 31.2% at comparable test usage.
  • Laboratory-planning scores rose from 21.80% to 46.01% across five policy versions.
HuggingFaceRead more · 1 min →
Today's

PolyOCR-Venus Unifies OCR Models

  • PolyOCR-Venus unifies recognition, grounding, document parsing, extraction and reasoning across OCR tasks
  • PolyOCR-9B scores 80.42 English and 78.37 Chinese on OCRBench v2.1; multilingual 2.7B scores 87.0 on MDPBench
  • OCRBench v2.1 receives verified corrections for 2,449 samples and task-aligned scoring metrics
HuggingFaceRead more · 1 min →
Today's

Microsoft Unveils Decision-Focused AI Model

  • Microsoft launches Decision-1, an AI model that returns probabilities for fixed choices
  • It leads on 36 unseen benchmarks and is 2.5 times faster than the runner-up, 35 times faster than GPT-6 Sol
  • Pricing matches Jev at $0.042 per 1 million input tokens, with free output
ITmedia AI+Read more · 1 min →
Today's

SoloQ Cuts Diffusion Model Memory Use

  • USC researchers introduce SoloQ, calibration-free quantization for diffusion language models
  • Tests on four models show retained accuracy under 4-bit quantization and gains over calibration-based baselines
  • NVFP4 cuts peak memory by up to 2.61X and speeds inference by up to 2.24X
HuggingFaceRead more · 1 min →
Yesterday'sInfra

HPE Predicts AI Will Handle Network Tickets

  • HPE is extending AI-powered networking automation across its hybrid cloud portfolio.
  • Rami Rahim estimates 70 to 80 percent of tickets need no human intervention today.
  • Rahim predicts no network issues will require humans within two to three years, except hardware swaps.
theregister.comRead more · 2 min →
Yesterday'sPolicy

Airtel Business CEO Supports AI Regulation

  • Airtel Business CEO Sharat Sinha supports planned AI regulation and calls the government consultation paper a step in the right direction
  • India’s paper is expected within a month and will address AI safety, deepfakes and a human-first approach
  • Airtel says spam prevention using network AI has reduced telecom-network fraud and expects five service areas to drive growth over three years
Economic TimesRead more · 1 min →
Yesterday's

Karpathy Says LLM Capability Gap Is Widening

  • Karpathy says about 6 billion people have barely encountered LLMs, while 1–2 billion use free chatbots casually.
  • About 20 million professionals use frontier models in math and coding; roughly 5,000 have internal access.
  • Karpathy cites problem scale, uneven capabilities, and access tiers as drivers of the widening gap.
twitter.comRead more · 1 min →
Yesterday's

Sakana AI Introduces MASS for Self-Improvement

  • Sakana AI introduces MASS, using virtual subagents to search for stronger team workflows and train a shared model
  • Two cycles with a 27B open-weights model raised output-token scores to 1.2-1.6x across four benchmarks
  • Researchers warn that shared evaluator blind spots may reinforce mistakes during recursive self-improvement training
Sakana AIRead more · 1 min →
Yesterday's

JA Higashibiwako Puts Three AI-Built Apps to Work

  • JA Higashibiwako staff built in-house business apps through conversations with AI.
  • Within one month, the cooperative developed and began operating three apps, expecting to cut work time by more than 70%.
  • The cooperative plans to expand the apps across inspection sites and other operations, with nationwide use by Japan’s agricultural cooperatives in view.
Impress DIGITAL XRead more · 1 min →
Yesterday'sInfra

GPT-6 Astra Faces BIM Design Challenges

  • GPT-6 Astra turns drawings, photographs and instructions into 3D models for architecture workflows.
  • A Revit test produced 327 walls, 162 doors and 98 windows from one screenshot, without independent verification.
  • Signal’s test found 13 of 14 planted defects in a 40-page document set; 21 of 23 observations were confirmed.
aecmag.comRead more · 2 min →
Yesterday'sHealthcare

GPT-5.2 Tested for Emergency Triage

  • GPT-5.2 evaluated for emergency triage using synthetic Hajj-context dialogues
  • Dialogue input linked to 79.7% lower under-triage overall and 91.9% in high-acuity cases
  • Model reached 94.5% high-acuity recall in 1,000 simulated scenarios
Semantic ScholarRead more · 1 min →

Trending Keywords

Last 7 Days
PrivacyTermsFAQContactContact
English
AIB Inc.Company info
© 2026 AIB Inc.