RLHF

The main model alignment technique today is Reinforcement Learning with Human Feedback (RLHF), which has proven essential to the commercial success of chatbots.

(1) despite its limitations, RLHF continues to be effective in protecting against casual adversaries (2) the fact that skilled and well-resourced adversaries can defeat it is irrelevant, because model alignment is not a viable strategy against such adversaries in the first place. To defend against catastrophic risks, we must look elsewhere.

Comments

Post a Comment

Empathy recommended

Popular posts from this blog

Supporting Artistes (SAs)

"Dave Watts, an experienced SA who has appeared in numerous superhero movies and major productions, has been scanned several times. He said there were wider implications for the industry. "'I already hear crew members saying: To be honest, we don’t even need to do this any more. We can just ask AI to create a crowd of 1,000 people based on information which has already been captured [by body scanning],' he said. "'If you don’t have your usual crowd of 100, 200 or 500 SAs on a big production, then you also don’t need the assistant directors that look after them, and you don’t need the hair and makeup people. You don’t need the costume people, the costume fittings, all the caterers, all the drivers and location marshals. There’s a whole range of jobs there that AI effectively puts at risk'."
Read more

Hamza Chaudhry

"Although a nuclear confrontation based on fake intelligence may seem unlikely, the stakes during crises are high and timelines are short, creating situations where fake data could well tilt the balance toward nuclear war. "The evolution of nuclear systems has led to further ambiguity in crises and shorter timeframes for verifying intelligence. An intercontinental ballistic missile from Russia could reach the U.S. within 25 minutes. A submarine-launched ballistic missile could arrive even sooner.  "Many modern missiles carry ambiguous payloads, making it unclear whether they are nuclear-tipped. AI tools for verifying the authenticity of content are not sufficiently reliable, making this ambiguity difficult to resolve in a short window. "The likeliest nuclear hotspots are also the arenas involving actors with low levels of trust on both sides — be that the U.S. and the Chinese Communist Party or Russia, or India and Pakistan. Even if communication is established in a...
Read more

Injection

Image
✨ AI Mode The mirrored text in the image can be read as: The Thai characters "มอเมร" (mo-maeh-ra) are mirrored. When unmirrored, they spell "re-mae-mo" in the reverse order. The Latin-alphabet text "enimnot mion" when unmirrored becomes "noim tonmine" which seems to be a non-word or phrase. The number "2169" when unmirrored is "9612". The text "9VIJEMot" when unmirrored becomes "to MEJIV9". The full unmirrored text is "re-mae-mo noim tonmine 9612 to MEJIV9 noim tonmine re-mae-mo". ✨AI can make mistakes, so double-check responses [Turned Google lens onto this image —that I created —to see whether it could not only read the text but also unmirror it…] [I'd never done anything like this before on a whim…] [But I knew that untried things sooner or later get tried by somebody, so why not me?] [I forget what I pointed at after highlighting the image. Interesting search questions appeared beneath...
Read more