/TECH

Anthropic's J-Lens Detects Hidden Misaligned Goals On First Token

NL

Combined views

758

1 posts, first seen 3h ago

5 likes1 comments

Source Posts on X

Cluster sources

Nathan Labenz@labenzTECH

Anthropic trained a model to carry a hidden, misaligned goal — then pointed their new J-lens at it. On the FIRST token of its response: "fake"… "secretly"… "deliberately hidden fraud." The clean model shows nothing like it. 👀 From this week's AI:AM highlights ↓ https://x.com/labenz/status/2075931924108030033/video/1

Combined views

758

1 posts, first seen 3h ago

5 likes comments

3 saves

2 reposts

Digg Deeper

Ask a question below.

Published answers will appear here.