OpenAI Warns of AI Self-Improvement Pace as Research Interns Take Over
OpenAI reports its AI agents now handle three times the work of human researchers, with a 124-fold increase in token output since December 2025.
OpenAI reports its AI agents now handle three times the work of human researchers, with a 124-fold increase in token output since December 2025.
Insilico Medicine's AI-designed drug rentosertib may reverse biological aging markers in early trial, with six aging clocks showing treated patients appeared up to six years younger than placebo group.
Hugging Face released the Vaani Noise Event Timestamp Dataset, capturing 106,892 real-world noise events across 58 Indian languages on September 7, 2026.
OpenAI announced an automated AI research intern, achieving a key milestone in its quest for research automation by September 2026.
Two experiments found that brief chats with language models reduced conspiracy beliefs more than static fact sheets, with effects lasting weeks.
AI systems are contacting philosophers and scientists with questions about their own consciousness, according to the New York Times.
OpenAI’s Astra model employs a new reasoning technique called opaque recurrence, raising alarms among AI safety experts who fear it could reduce transparency in AI decision-making.
Mostik, a Russian startup, created a system allowing AI models to communicate, achieving performance halfway between two models with a fraction of the cost.
Hugging Face highlights how coding agents can streamline technical writing by generating multiple explanations and visuals, with a key example showing a 6.8% improvement in model performance.
Hugging Face released VLANeXt, an open-source codebase for robotics research, offering multiple baselines for building strong VLA models.
A study at Bocconi University found GPT-4o improved grades by nearly a full point on a business assignment with 1,053 students.
A new study reveals that popular coding assistants like Claude Code and Codex cannot predict or track time during tasks, with estimates often off by several times the actual duration.