Smitha Milli - Approaches to Achieving AI Safety

1,829 views · Published 12 September 2017 · 20:20 · Indexed 14 November 2026

Channel: Science, Technology & the Future · 2017 · Science & Technology

Watch on YouTube

Sorry for the air-conditioner noise in the bg.  
Topics:
- Approaches to achieving AI control
- Learning human preferences
- Extrapolating human volition
- Aggregating human preferences
- Difficulties in implementing underspecified philosophy
- A mixed approach to implementing ethical theories in AI
- A theory of metareasoning & bounded rationality
- Open problems in value alignment
- Generalizable AI safety research

Bio:  I want to make machine learning systems more robust and reliable. I am particularly interested in how interaction with humans can help ensure that a system has learned the correct objective or behavior.

I am currently a research intern at OpenAI. I will be starting my PhD in the fall at UC Berkeley, where I also did my undergrad. I am a part of the Center for Human-Compatible AI and the Berkeley Artificial Intelligence Research Lab.

I spend a lot of time introspecting on my own styles of thinking and productivity. Here is some collected research advice. Outside of research I enjoy talking to strangers, reading both nonfiction and fiction, and recently improv.

http://smithamilli.com/

Many thanks for watching!

Consider supporting SciFuture by:
a) Subscribing to the SciFuture YouTube channel:  http://youtube.com/subscription_center?add_user=TheRationalFuture
b) Donating via Patreon: https://www.patreon.com/scifuture and/or
c) Sharing the media SciFuture creates: http://scifuture.org 

Kind regards,
Adam Ford
- Science, Technology & the Future

More from this channel