We're continuing investigation to the performance issues, you can follow and react here: https://lemmy.world/c/friendicaworld

Baldur Bjarnason

2 weeks ago

Baldur Bjarnason
2 weeks ago

“Don’t cite the Adversarial Poetry vs AI paper — it’s chatbot-made marketing ‘science’ – Pivot to AI”

pivot-to-ai.com/2025/11/24/don…

> They wrote a good headline, and then they faked the scientific process bit.

Don’t cite the Adversarial Poetry vs AI paper — it’s chatbot-made marketing ‘science’

Today’s preprint paper has the best title ever: “Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models”. It’s from DexAI, who sell AI testing and compliance ser…

^{Pivot to AI}

reshared this

in reply to Baldur Bjarnason

Matthew Skelton

in reply to Baldur Bjarnason 2 weeks ago

very méta

in reply to Baldur Bjarnason

Adam Shostack

in reply to Baldur Bjarnason 2 weeks ago

I don't think the blog post makes the point its headline claims. Ironic. But: Using a chatbot to transform attacks into "poetry" works, and the paper authors do show that. The fact that that poetry isn't human-generated doesn't invalidate the attack. The "hard work" is running the evals, and the attack does what the (paper) headline says.

in reply to Adam Shostack

David Gerard

in reply to Adam Shostack 1 week ago

@adamshostack the evals were also done with a roulette wheel

@Adam Shostack

⇧