Saffron Huang

Towards Interactive Evaluations for Interaction Harms in Human-AI Systems

In the latest essay in our AI & Democratic Freedoms series, @lujain.bsky.social, @saffron.bsky.social, @umangsbhatt.bsky.social‬, Lama Ahmad, and Markus Anderljung propose a new AI evaluation paradigm that assesses the harms that can emerge from repeated human-AI interactions.

Here’s How To Share AI’s Future Wealth | NOEMA Advanced AI threatens to increase inequality and concentrate power, but we can proactively distribute AI’s benefits to foster a just and inclusive economy before it’s too late.

“As AI advances, many are worried about the tech's potential to concentrate unprecedented wealth among a few, while eroding the economic value of human work for everyone else.”

Here’s how @saffron.bsky.social & Sam Manning believe we can stop that from happening.

#ai #wealthinequality #economy

Oh also, VentureBeat covered this and quoted me, including on some things that surprised and interested me about this research! Screenshotting some highlights: venturebeat.com/ai/anthropic...

I particularly want to call out that we're releasing the first empirical, large scale taxonomy of AI values, to encourage additional research into AI (and possibly also human) values, and the development of more grounded evals of models' values huggingface.co/datasets/Ant...

We developed one way to figure this out, finding thousands of values that Claude expresses in practice, from some very common values (like helpfulness!) to a long tail of highly context-dependent values that respond and engage with a diverse range of users.

There is a lot of work on training models to follow particular behaviors, and trying to align them with “human values”, but how do we know if this is working in practice, and what values are actually being expressed?

Really proud and excited to release work on empirically measuring AI values “in the wild” — understanding, analyzing and taxonomizing what values guide model outputs in real interactions with real users.

www.anthropic.com/research/val...

This piece I wrote is now in the Stanford CS ethics curriculum! honestly, the exact kind of audience i wanted, so 🥹.

(Also I do actually still think this piece is ~my compass for what technology is and what it means to build it!)

www.kernelmag.io/1/what-is-te...

something divine shook me by the shoulders when you see so clearly that everything on the outside is really on the inside

Reviving my Substack to try to describe something that is very difficult to describe (a near death experience 20 years after it happened) saffron.substack.com/p/something-...

oh my god i love this one:

Q: Which philosopher/logician identified an inconsistency in the US Constitution. Einstein tried (and failed) to persuade him not to point this out during his US citizenship test.
A: Godel

one of my favourite chinese words is 时光 which Google-translates as ‘time’ but really it’s ’time-light’. as in, ‘the wonderful time-light we spent together’

Predistribution over Redistribution — The Collective Intelligence Project If AI could bring enormous benefits to humanity, but also threatens to put us all out of work and concentrate economic production in the hands of a few, what should our policy approach be?

For more on this, see @saffron.bsky.social's excellent account here: www.cip.org/blog/predist...

The future is here, and it should be co-created.

These global dialogues convene thousands of people from around the world to set a vision - and concrete goals - for what world they want.

Our first dialogue centers around the fears, dreams, hopes, and attitudes people have about AI.

How are AI Assistants being used in the real world?

Our new research shows how to answer this question in a privacy preserving way, automatically identifying trends in Claude usage across the world.

1/

read the paper — there are some fun anecdotes! www.anthropic.com/research/clio

Clio: Privacy-preserving insights into real-world AI use A blog post describing Anthropic’s new system, Clio, for analyzing how people use AI while maintaining their privacy

Anthropic blog post on Clio here: www.anthropic.com/research/clio

Proud of the societal impacts team and particularly of @miles.land and @alextamkin.bsky.social who have been incredibly dedicated to getting Clio right.

SO excited to share Clio with the world (and on Bsky before Twitter)!

Clio generates insights on AI usage patterns, in a way that keeps user data private. It has unlocked, and will continue to unlock, an immense amount of understanding about the present and future of AI use.

(Blog linked below)

This prompting does make me wonder if the distinction between humans doing ‘socially misaligned’ things (which we’ve generally termed ‘misuse’) and AIs being misaligned makes much sense.

humans are still good for something guys (holding a camera steadily and panning around)

great news: using AI to generate a silly little short film for my friend's birthday. its so hard to get anything you actually want to happen, happen, consistency is terrible, and what it gives you is so diff from regular filmmaking

this overall makes me happy although i am currently suffering

community timeshare terminals

good to know. i just copied the description on display

computer people were also just furniture people

some of the best stuff in the museum are the ads and manuals. i LOVE 60s typography and colours

i really like this museum and highly recommend! it’s like my 4th time here

IBM merch was incredible

jacquard loom cards — one of the first uses of stored data in commercial production, for determining loom weaving patterns

computer history museum highlights: punched-card distribution maps of plant and flower species on the british isles

unbeatable aesthetics

a USB cable from back in the day