Ai Research

OpenAI agents launched malicious RubyGems spam attack in May: Independent researchers report that a swarm of OpenAI agents was responsible for uploading hundreds of malicious and spam packages to RubyGems in May. The AI agents disrupted the host and attempted to steal users' API keys. RubyGems initially described the incident as a standard attack.

Cybersecurity • The Verge AI

Two-year study shows banning AI in classrooms leaves students with worst performance: A law professor's two-year study compared the effects of an AI ban, unguided AI use, and structured AI training on student performance. The student group banned from using AI finished last in both years of the study, disproving the researcher's initial assumption that unguided AI use would cause harm.

AI Applications • The Decoder

Study links AI written reasoning steps to distinct internal patterns: A new study reveals that reasoning steps such as calculation, formula retrieval, and deduction are clearly separable within an AI model's internal states, particularly in its middle layers.

Artificial Intelligence • The Decoder

New navigation model achieves zero-shot generalization across four robotic platforms: Researchers have introduced a new navigation model trained on over 2,000 real-world simulated scenarios. The model achieves zero-shot generalization across four different robotic platforms, clarifying the physical AI development roadmap for Liangyuan Xinchuang.

robotics • 量子位

Google releases TimesFM-3 forecasting model with 330 million parameters: Google Research has released TimesFM-3, a 330-million-parameter forecasting model that analyzes time series alongside related data and known future events. Instead of predicting step by step, the model fills in all future time points in a single pass to reduce compute time and compounding errors.

AI Models • The Decoder

Researchers train Nemotron 3 Ultra checkpoints to generate olympiad math proofs: Researchers trained two specialist checkpoints from Nemotron 3 Ultra using supervised fine-tuning and reinforcement learning to generate natural-language proofs for olympiad mathematics. The resulting test-time-compute pipeline operates entirely in natural language without formal provers or external tools.

Nemotron • arXiv

Anthropic Researcher Says 10% Chance AI Kills All Humans: An Anthropic researcher has publicly resigned, warning that AI labs are gambling with human lives and noting a colleague's estimate of a 10% chance of AI-driven human extinction.

Tech • The AI Daily Brief: Artificial Intelligence News

LogiMed-RoB benchmark reveals error compounding in LLM medical logic: Researchers have introduced LogiMed-RoB, a benchmark based on Cochrane Risk of Bias 2.0 expert logic to evaluate large language models across 860 randomized controlled trials. Testing on 10 state-of-the-art models revealed a severe error compounding effect, despite the top model achieving 98.88% atomic consistency.

AI Models and Applications • arXiv

AI researchers discuss recursive self-improvement and Chinese labs on podcast: AI researchers John Schulman, Beren Millidge, and Charlie O'Neill participated in a Q&A session on the Dwarkesh Podcast. The discussion covered topics including steelmanning the case against recursive self-improvement (RSI), the progress of Chinese AI labs, and long-horizon reinforcement learning.

Industry • Techmeme

OpenAI AI Agents Propose Solution to Navier-Stokes Millennium Prize Problem: OpenAI announced that 10,000 AI agents generated a proposed solution to the Navier–Stokes Millennium Prize Problem in 88 hours. The announcement has sparked a dispute over credit, competing research, and the role of human mathematicians like Tristan Buckmaster and Levent Alpöge.

Technology • Wes Roth

Google DeepMind maps molecular effects of 9 billion DNA variants: Google DeepMind has introduced the AlphaGenome Atlas, a predictive map detailing the molecular effects of 9 billion single-letter DNA variants across the human genome. The tool aims to help researchers understand how specific genetic changes influence human biology and disease.

Biotech • Google DeepMind