AI news on papers, experiments, academic findings, technical deep-dives. Showing stories before August 19, 2026.
Qwen 3.8 27B Matches GPT-5.6 Luna Score
Apple Finds GRPO Effective For Non-English Reasoning
FLOPs Fail To Predict Real AI Execution Time
OpenAI Launches National Security Oversight Initiative
Open Weights Differ From Open Source
Apple Tests Multilingual Reasoning Via GRPO
Qwen 27B Matches GPT-5.6 Luna Performance
OpenAI Partners With CodeAI For AI Literacy
DiG-bench Tests AI Intuition In Games
Open Weights Differ From Open Source
Qwen 27B Matches Massive Models In Benchmarks
FLOPs Fail To Predict Actual AI Execution Time
OpenAI Partners With CodeAI For Student Literacy
DiG-bench Tests AI Discovery In Games
FLOPs Fail To Predict AI Execution Time
Qwen 3.8 27B Matches Top Model Scores
OpenAI Details Internal Cybersecurity Defenses
DiG-bench Tests AI Intuition via 70 Games
Qwen 3.8 27B Matches GPT-5.6 Luna Score
FLOPs Fail To Predict Real AI Execution Time
1985 Analysis Predicts Modern AI Limits
OpenAI Details Internal Cybersecurity Defenses
RubricForge Reduces Over-Crediting in Agent Evaluation
Open Weights Differ From Open Source
DiG-bench Tests AI Intuition Via Games
Qwen 3.8 27B Matches GPT-5.6 Luna Score
OpenAI Shares Cybersecurity Defense Strategies
RubricForge Reduces Agent Over-Crediting
DiG-bench Tests AI Intuition In Games
Open Weights Differ From Open Source
1985 Critique Re-evaluates AI Limits
RubricForge Reduces Agent Evaluation Over-Crediting
DiG-bench Tests AI Discovery In Games
OpenAI Details Internal Cybersecurity Defense Strategies
Amazon AI Facility Tracks Rare Book Shipments
Open Weights Differ From Open Source
OpenAI Funds 14 AI Policy Research Projects
Building DIY AI Text Detectors
1985 Analysis Predicts Modern AI Limits
RubricForge Reduces Over-Crediting in Agent Evaluation
Z.ai Releases High-Efficiency GLM-5.3 Model
Indonesia Launches First University AI Center
1985 Analysis Predicts AI Limits
Z.ai Releases High-Efficiency GLM-5.3 Model
Tutorial Explores DIY AI Text Detectors
Latent Reasoning Models Rely On Training Data