Sebastian Benthall

Research Scientist at @ILI_NYU and @ICSIatBerkeley. A dad. Tweets are reflex, noise, and error. 🦋

[Sebastian Benthall @sbenthall]

Migrating to bluesky: 🦋@sbenthall.bsky.social

[Sebastian Benthall @sbenthall]

I'm building something new. https://t.co/fjiazovCWw

[Sebastian Benthall @sbenthall]

What about the loyalty of AI agents in government function? https://t.co/RbLxRU9XAf

[Sebastian Benthall @sbenthall]

Complex systems and ethics according to contextual integrity https://t.co/UwyMawRxf6

[Sebastian Benthall @sbenthall]

@AndrewCritchPhD There's a fair case to be made for all anthropomorphic self-references of contemporary chatbots to be deceptive, but we might need a new kind of grammar for the speech of products and services.

[Sebastian Benthall @sbenthall]

@AndrewCritchPhD 'Person' is a much more capacious term than 'human', as we already have artificial persons (i.e. corporations). At the policy level, the question is going to be about the relationship between corporate persons and robots pretending to be people but deployed by corporations.

[Sebastian Benthall @sbenthall]

a brief reflection on where I'm at as a scientist https://t.co/Mivi7XNHEH

[Sebastian Benthall @sbenthall]

Anyone had luck with Bluesky yet?

[Sebastian Benthall @sbenthall]

Ok, but what is the anti-incumbency wave about? Is there an explanation of that yet? Is it just 'inflation'?

[Sebastian Benthall @sbenthall]

@zephoria Will you be in Ithaca or NYC?

[Sebastian Benthall @sbenthall]

@Miles_Brundage Seems like the original sin of the AGI definition is that it's based on a comparison with a human as a unit of cognition. What would you say are better categories/concepts to use? I'd welcome recommended reading on this.

[Sebastian Benthall @sbenthall]

@AJA_Cortes parenting

[Nora Belrose @norabelrose]

This is a great paper. It points out:

1. Humans do not even approximately behave according to rational choice theory
2. There is no reason to think advanced AI will "inevitably" maximize some utility function
3a. Human preferences are derivative / constructed, so aligning AI by matching its behavior to our stated preferences is wrongheaded
3b. We can align AIs directly to some normative ideal of a "good assistant / programmer / driver / etc." instead
4. Aggregating preferences across people is fraught with philosophical and mathematical difficulties. We should not aim to align AI to the "collective will of humanity."

[Sebastian Benthall @sbenthall]

Direct, concise answers to coding questions in Google AI Overview is a real improvement, saving a few clicks every half hour or so.

[Sebastian Benthall @sbenthall]

I miss writing. https://t.co/QQ5DF4Jyog

[Sebastian Benthall @sbenthall]

I had a great time at the 6th Annual Symposium on Applications of Contextual Integrity!

It's a great community that's always growing intellectually, with a sincere and friendly vibe.

Thanks to everyone who organized it.

[Sebastian Benthall @sbenthall]

@nils_gilman And yet, incumbent privilege is to some extent all there is.

[Sebastian Benthall @sbenthall]

@rtk254 https://t.co/T0K2Na0ucp

[Sebastian Benthall @sbenthall]

@nils_gilman His argument hasn't changed at all in 10 years, which shows how immune it is to evidence.

[Sebastian Benthall @sbenthall]

@rtk254 availability of training data; ease of deployment