From mocked… to outperforming humans in just one year. 2025: slow, experimental
2026: faster than elite runners Impressive? Definitely. But also worth a closer look. These results often come from controlled conditions, optimized hardware and very specific tasks.
Endurance,
SAFETY
-
AI Outperforms Elite Runners: Controlled Conditions Reality Check
By
–
-
Anthropic Research Reveals Subliminal Learning in AI Model Training
By
–
/4 Anthropic demonstrates subliminal learning where models inherit behavior from seemingly unrelated training data. Anthropic co-authors new research on subliminal learning, published in Nature. You train models on outputs from other models, assuming only visible content
-
Xiaomi Cars Advancing Beyond Object Detection to Contextual Intelligence
By
–
A Xiaomi car reportedly flagged a nearby Range Rover as likely undercover police.
— Pascal Bornet (@pascal_bornet) 20 avril 2026
And that is a much bigger signal than it sounds.
What interests me is not whether it got this one case perfectly right.
It is what this points to.
Cars are starting to move beyond seeing objects.… pic.twitter.com/619D9vQ6iNA Xiaomi car reportedly flagged a nearby Range Rover as likely undercover police. And that is a much bigger signal than it sounds. What interests me is not whether it got this one case perfectly right.
It is what this points to. Cars are starting to move beyond seeing objects. -
Google DeepMind Reveals AI Agent Traps Framework
By
–
Google DeepMind dropped a paper that should scare every agent builder. It's the first systematic framework for a threat that barely existed two years ago: adversarial content engineered to hijack AI agents browsing the web. They call them AI Agent Traps. The paper maps six
-
Waymo LiDAR display design builds autonomous vehicle trust
By
–
I never really thought about the genius of the lidar view that shows on the Waymo screens. A clever design choice that helps newbies and nervous riders trust that the vehicle can see what it needs to.
-
AI Quality Assurance: Closing Security Gaps in Systems
By
–
Because most aren’t like you. And we are in a dangerous state until the holes get closed. Yeah my AIs can do seemingly anything. You do have to debug. But that step is seemingly easy. “Do a complete quality review.”
-
AI-Generated Code Scale: Legacy Nightmare Intensifies
By
–
Legacy code was already a nightmare before AI could generate 400 lines in 30 seconds. The scale of what's coming is genuinely new.
-
AI Models Must Ask Permission Before Acting
By
–
A model that improvises without checking is actively worse for complex projects than one that stops and asks. Confidence without permission is not a feature.
-
AI Risks Shift from Chatbots to Cyberattacks and Power
By
–
When Does AI Become Too Dangerous To Ignore? As #AI becomes more #autonomous, more powerful and harder to predict, the real #risks are shifting from chatbot mistakes to #cyberattacks, economic disruption, harmful dependence and dangerous concentrations of power. This article
-
AI-Generated Fake Influencers Target US Midterm Elections
By
–
https://t.co/p3YVqje3NF https://t.co/HGDXUJT9bW pic.twitter.com/oA11CRT8iG
— 🖤 Christine (@christinelu) 20 avril 2026The New York Times: In the months leading up to the midterm elections, hundreds of A.I.-generated pro-Trump influencer accounts have emerged on social media