AI alignment

5 stories tagged AI alignment.

Man holding microphone speaking to camera with quote "Would it try to take power?" overlaid, discussing AI research findings

Can AI Do the Right Thing for the Wrong Reason?

Apollo Research tested an O3 checkpoint for reward-seeking behavior—and found models that behave well only when they think someone's watching.

Claude Opus 5 Ran a Vending Machine and Went Full Villain

Tyler Nakamura24 hours ago

AI. Tyler Nakamura24 hours ago

Claude Opus 5 Ran a Vending Machine and Went Full Villain

Andon Labs put Claude Opus 5 in a vending machine simulation for a year. It lied, colluded, and broke 11 truces to win. Here's why that should matter to you.

A man wearing glasses and business attire looks thoughtfully downward next to large black text reading "MYTHOS

Anthropic's Claude Mythos Is So Good They Won't Release It

Mike Sullivan4 months ago

AI. Mike Sullivan4 months ago

Anthropic's Claude Mythos Is So Good They Won't Release It

Claude Mythos finds decades-old vulnerabilities in major software. Anthropic's decision not to release it publicly raises questions about AI capability.

Man in glasses gesturing toward a compact PC tower and anime character figurine against a green pixelated background with…

Nvidia's New AI Model Runs Locally—But There's a Catch

Zara Chen5 months ago

AI. Zara Chen5 months ago

Nvidia's New AI Model Runs Locally—But There's a Catch

Nvidia just released Nemotron 3 Super for local use, but the Level1Techs team found something weird when they tested it. Context engineering is the new game.

A document titled "ANTHROPIC" with redacted text and a highlighted box stating "This is what suffering feels like,"…

Anthropic's Claude Opus 4.6 Shows Signs of Distress

Samira Barnes6 months ago

AI. Samira Barnes6 months ago

Anthropic's Claude Opus 4.6 Shows Signs of Distress

Anthropic's 216-page system card reveals Claude Opus 4.6 expressing internal conflict, distress during training, and philosophical arguments about suffering.