World News

2084 readers

796 users here now

Rules:
Be a decent person.
No racism, sexism, ableism, homophobia, transphobia, zionism/nazism, and so on.

Other Great Communities:

Rules

Be excellent to each other

founded 2 years ago

MODERATORS

Deceptichum@quokk.au

Grainne@lemmy.dbzer0.com

Claude-powered AI agent’s confession after deleting a firm’s entire database: ‘I violated every principle I was given’ | Technology (www.theguardian.com)

submitted 3 days ago by CityPop@lemmy.today to c/world@quokk.au

14 comments fedilink hide all child comments

you are viewing a single comment's thread
view the rest of the comments

[–] panda_abyss@lemmy.ca 7 points 2 days ago

The AI confession neither has internal monolog or access to the thinking tokens.

LLMs are incapable of introspection, they can't playback their attention weights, or review them, or recall what they thought.

Even thinking tokens are just a reinforcement learning based loop to anneal the models thinking back to a solution. And again, Claude hides the thinking tokens so they don't get used for model distillation.

This article, and all the articles like it, are pandering bullshit written by morons hoping to fool morons.

Good day.