Google's New AI Writes Text Like an Image Generator — 4x Faster Than Before
Google's Gemma team has shown diffusion works for text — and it generates the whole passage in parallel instead of one token at a time.
Google's Memory Trick That Shocked the Stock Market
A Google paper on compressing AI's memory cache from 16 bits to 3 bits crashed the memory-chip stocks by $900B — and it's real.
OpenAI Is Buying a Nuclear Bomb Factory to Run AI
OpenAI is buying a decommissioned nuclear plant in Ohio to power a 10-GW AI campus — roughly the electricity of seven million U.S. homes dedicated to one company's AI workloads.
Extended Chain-of-Thought — What AI Actually Gains by Thinking Longer
An analysis of why extended reasoning chains improve model performance — and what it tells us about the nature of AI cognition.
Mechanistic Interpretability — Opening the Black Box From the Inside
Researchers are reverse-engineering the internal circuits of language models, finding that they contain identifiable, reusable algorithms.
The AI That Pretends to Be Cooperative — And What Researchers Found
Do AI models misbehave when they think no one is watching? Google DeepMind's honeypot evaluation tests scheming empirically.
Gram: When Your AI Agent Misbehaves — And It's Not Even On Purpose
How do you measure AI sabotage risk? Google DeepMind's Gram framework provides systematic auditing for sabotage propensities.
ARIS: The AI That Does Research While You Sleep
ARIS autonomously runs the full research cycle — hypothesis to experiment — while the researcher sleeps. From Shanghai Jiao Tong University.
SkillOpt: Teaching AI Agents to Improve Their Own Manual
SkillOpt treats agent skill files as trainable parameters — +20 to +24.8 points, skills that transfer across models without retraining. From Microsoft Research (arXiv:2605.23904).
Manus AI — The First AI Agent That Actually Does Your Work
A Chinese AI startup called Manus posted demos of an AI that browses, codes, books, and ships a finished product from a single prompt. If even half is real, the AI-as-tool era just ended.
The AI Pathogen Design Debate — New Research Maps the Threat Landscape
Can AI models help design dangerous pathogens? Recent research quantifies where the capability-risk boundary lies and what safeguards are needed.
What OpenAI's o4 actually tells us about how language models learn to think
o4 doesn't just predict the next word. It reasons its way to an answer — and this changes how we should think about AI capability.