Research

Everything in this section in one place — newest stories first.

LatestLLMs

New Benchmark Shows AI Vision Models Fooled by Deceptive Scenes

A new arXiv paper defines 'situational illusions', cases where a scene's appearance misleads MLLMs about its true physical state

AI AgentsAugust 25, 2026

Long-running AI agents forget their own safety rules fast

Claude Code's /compact on Sonnet 4.6 retains only 10% of safety rules after five compaction rounds, down from 53% after one round, across 20 production agent configurations

LLMsAugust 23, 2026

Nvidia research: the harness matters more than the model

Nvidia's custom harness pushed Claude Opus 5 from a 30% score to 100% on the ARC-AGI-3 benchmark without changing the model itself.

ResearchAugust 19, 2026

What Wireless Foundation Models Mean for 6G Networks

A 28-page arXiv survey by Nasir Saeed maps how foundation models could underpin AI-native 6G wireless networks.

LLMsAugust 19, 2026

New Framework Exposes a Trust Gap in Healthcare LLMs

A new arXiv paper proposes a four-component framework for evaluating causal grounding in healthcare LLMs, tested on a cardiovascular pilot.

ResearchAugust 19, 2026

DeepMind's hurricane AI cuts track error by 140 kilometers

Google DeepMind's WeatherNext Cyclones model cut five-day hurricane track error to 230 km, versus 370 km for ECMWF's ENS system and 335 km for DeepMind's own GenCast

LLMsAugust 15, 2026

Unreleased Anthropic Model Advances 150-Year-Old Riemann Hypothesis

An unreleased Anthropic model raised the verified lower bound for the Riemann hypothesis after working autonomously for about 36 hours

LLMsAugust 15, 2026

Z.ai's GLM-5.3 Gains Coding Muscle Without Retraining the Base Model

Z.ai released GLM-5.3 on August 14, 2026, reusing the 743B-parameter base model from GLM-5.2 and driving all gains through post-training alone.

ResearchAugust 10, 2026

AI Designed Working Viruses From Scratch, Scientists Confirm

Stanford and Arc Institute researchers used AI models Evo 1 and Evo 2 to design new bacteriophage genomes, published in Science, August 2026

AI AgentsAugust 10, 2026

Shopify: AI Search Tripled Store Traffic, Not Cannibalizing Google

Shopify says AI-driven traffic and orders to its merchant stores tripled year-over-year in Q2, while traditional search traffic also grew 1.3x over two years

LLMsAugust 4, 2026

Berkeley's ABBEL Tackles the Memory Loss of Summarized AI Context

Berkeley AI Research introduces ABBEL, a framework training LLMs to compress interaction history into natural-language 'belief states' without losing key information

AI AgentsAugust 4, 2026

New Web Agent Framework Boosts Long-Horizon Task Success by 13.8%

FCPAgent is a new planning framework for LLM-based web agents built around 'Falsifiable Commitment Units' that can be actively tested during execution

ResearchAugust 4, 2026

Study Pits Human Intuition Against Neural Nets on the Traveling Salesman

A new arXiv preprint (submitted July 26, 2026) compares human problem-solving with neural networks on the traveling salesman problem (TSP).

ResearchAugust 4, 2026

Hybrid AI Model Corrects Radar Errors in Greenland Ice Maps

A new hybrid AI-physics model corrects penetration bias in radar-based elevation measurements of Greenland's ice sheet

023 Aerial view of Jakobshavn Glacier at Disko Bay (Greenland) Photo by Giles Laurent

AI in HealthcareAugust 3, 2026

AI in Medical Diagnosis at Arab Hospitals: How Far Has It Actually Gone?

Hospitals in the UAE and Saudi Arabia are already using AI in radiology, with advanced algorithms cutting image-interpretation time by 40% in Saudi hospitals as of December 2025, while other tools like Al Qassimi Hospital's eye-print detection device in the UAE are still slated for launch in 2026.

Editorial image for الذكاء الاصطناعي في التشخيص الطبي بالمستشفيات العربية

Developer ToolsAugust 3, 2026

Simon Willison's LLM CLI Ships 0.32rc2, Adds Support for New GPT-5.6 Model Family

LLM CLI reached version 0.32rc2 on July 30, 2026, fixing a dependency bug from 0.32rc1

AI InfrastructureAugust 3, 2026

Gulf AI Data Centers: The Gap Between Announced Figures and Research Estimates on Power and Water Consumption

The only near-confirmed official figure is Saudi Arabia's target of 6 gigawatts of data center capacity by 2034 with 29% CAGR in power demand, while regional water consumption estimates for 2030-2031 range from 426 to 535 billion liters depending on the research source.

Editorial image for استهلاك مراكز بيانات الذكاء الاصطناعي للكهرباء والمياه في الخليج

AI in HealthcareAugust 3, 2026

AI in Diagnostics at Arab Hospitals: Promising Projects, Still Limited Scientific Evidence

Hospitals in the UAE and Saudi Arabia, such as King Faisal Specialist Hospital and Al Qasimi Hospital, have begun trialing AI systems for imaging analysis and early detection of diseases like Alzheimer's and strokes, but most remain launch-phase projects rather than completed clinical approvals.

All articles