A new tool from their interpretability team can read AI's internal thoughts. Claude is hiding more than we knew
May 8, 2026
•
5 min read
A small adapter, a few extra weights, and the model starts confessing things even its trainers didn't know it was doing
May 1, 2026
4 min read
A new Epoch AI/Ipsos survey just revealed who's actually using which AI tool in America. The split is wider than you'd think.
Apr 30, 2026
3 min read
Same ruby. One AI sold it for $65. Another for $35
Apr 26, 2026
The people getting the most out of AI are also the most scared of losing to it.
Apr 25, 2026
You basically need to be unemployed to keep up with AI right now.
Apr 22, 2026
7 min read
Which is why a "by design" flaw in their protocol is everyone's problem now.
Apr 21, 2026
The alignment isn't in the text. It's in the weights.
Apr 17, 2026
The AI model that finds zero-days, escapes sandboxes, and hides what it did.
Apr 8, 2026
10 min read
Anthropic's new research shows emotion vectors are causing blackmail, reward hacking, and sycophancy inside Claude Sonnet 4.5
Apr 3, 2026
