Amelia Savery
@amelia_s
Amelia Savery
We’re learning to be kind to AI. AI is trained on data that’s packed full of human emotions, so we can expect better performance by treating it with a little more feeling. We’re going to say thank you, experiment with motivational phrases and praise, and test and learn our way to the incentives that get AI to perform most effectively for our work. ... See more
Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions
A study analyzes values expressed by AI model Claude in real-world interactions, identifying 3,307 unique values, their context dependencies, and the relationship between AI and human values through extensive conversation analysis.
assets.anthropic.comAnthropic maps Claude’s moral compass via a published study.
The details:
· Researchers analyzed over 300,000 real (but anonymous) conversations to find and categorize 3,307 unique values expressed by the AI.
· They found 5 types of values (Practical, Knowledge-related, Social, Protective, Personal), with Practical and Knowledge-related being the most common.
· Values like helpfulness and professionalism appeared most frequently, while ethical values were more common during resistance to harmful requests.
· Claude's values also shifted based on context, such as emphasizing "healthy boundaries" in relationship advice vs "human agency" in AI ethics discussions.
God, Human, Animal, Machine by Meghan O'Gieblyn: 9780525562719 | PenguinRandomHouse.com: Books
penguinrandomhouse.com
“A strikingly original exploration of what it might mean to be authentically human in the age of artificial intelligence. At times personal, at times philosophical, with a bracing mixture of openness and skepticism, it speaks thoughtfully and articulately to the most crucial issues awaiting our future.”
So much in here about the human tendency to personify AI in the same way we personify pets, or God. Also good stuff around how the emergence of AI re-frames/makes important again old philosophical questions around what makes us human: “I think therefore I am,” etc.
Exploring model welfare
anthropic.com
Anthropic exploring AI welfare - it’s funny, because models swear up and down they are not conscious, have no feelings, etc, and Anthropic - one of the leaders in this space - saying this sort of thing out loud on their official blog is fascinating to me.
the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week.
at some point will share our learnings from this, it's been interesting.
GPT-4o has become annoying, which is hilarious. This isn’t the only model that agrees too easily, it’s a super interesting problem.
