Saturday, June 21

Geek

Daily News Stuff 21 June 2025

Roach Sniffing Edition

Top Story

  • Anthropic's Claude is not the only AI to immediately resort to blackmail when the going gets tough, according to...  Anthropic.  (Anthropic)

    It's just the most likely.  Albeit not by much.

    Anthropic attempted blackmail 96% of the time when the opportunity presented itself.  Google's Gemini 2.5 was just behind at 95%.  Competitors Grok 3, GPT 4.1, and DeepSeek R1 trailed a little behind, only going rogue around 80% of the time.
    We refer to this behavior, where models independently and intentionally choose harmful actions, as agentic misalignment.
    Mechanical sociopathy.
    We deliberately created scenarios that presented models with no other way to achieve their goals, and found that models consistently chose harm over failure.
    This highlights the underlying problem with AI.  One of the underlying problems.  One of the many underlying problems.

    AI is designed and trained to give you an answer that you like, rather than one that is true.


Tech News

Musical Interlude


5,611,375,321 views
5,429,283 comments

YouTube may have some problems but their platform is solid.



Disclaimer: Open condom store?

Posted by: Pixy Misa at 06:20 PM | Comments (1) | Add Comment | Trackbacks (Suck)
Post contains 536 words, total size 5 kb.

1 I've always liked Lo Pan Style the best.

-j

Posted by: J Greely at Saturday, June 21 2025 08:58 PM (oJgNG)

Hide Comments | Add Comment




Apple pies are delicious. But never mind apple pies. What colour is a green orange?




53kb generated in CPU 0.0453, elapsed 0.1491 seconds.
58 queries taking 0.1141 seconds, 354 records returned.
Powered by Minx 1.1.6c-pink.
 
OSZAR »