Yeah, and don't get me wrong, this is very cool work! And I know you folks know these things. But
1. I think there's a lot of people who don't and just hope for a "silver bullet" and
2. Getting organizational buy-in seems tricky, particularly with malicious parties out there.
SAFETY
-
Organizational AI adoption challenges and malicious threat mitigation
By
–
-
Open Source AI Models and Regulatory Frameworks Need Society Agreement
By
–
I'm also cautious about rogue model designers who don't play by the rules. At least for the time being, it seems like the community is very good at making open source alternatives to proprietary models. We'd need to agree as a society that certain things should be off limits.
-
Model Designers and Users Share AI Safety Responsibility
By
–
Definitely! With cooperation from model designers, there's hope. But like I said, if all the burden is on the users alone? Then I'm not so hopeful.
-
Rich People Power and Organizational Resistance Strategies
By
–
Let's say no in case one of them is listening 😉 More seriously whether you're a for-profit, non-profit, public or private, the truth is that rich people can always mess with you. Hopefully they won't and we'll try to resist if they do!
-
Image Sanitization Safety Against Advanced AI Models
By
–
Nice article by @KyleBarr5
. I gave my perspective (really, the perspective of Radiya-Dixit et al. https://
arxiv.org/abs/2106.14851): users sanitizing and releasing their own images seems hopeless. It may be safe today, but it probably won't be safe against a smarter model a year from now. -
AI Impersonation Service Creates Identity Fraud Problems
By
–
Who could have possibly predicted that selling the ability to impersonate anyone for a few hours for $8 would result in problems with "impersonation issues"?
-
CDEIUK Launches Red-Teaming Recruitment for Privacy-Enhancing Technologies
By
–
The @CDEIUK has also launched it's recruitment process for red-teaming these projects. https://
ktn-uk.org/news/red-team-
registration-uk-privacy-enhancing-technologies-challenge-prize/
… -
Future Fund crisis threatens NeurIPS ML Safety workshop grants
By
–
"…there are many committed grants that the Future Fund will be unable to honor." The #NeurIPS2022 ML Safety Workshop (
https://
neurips2022.mlsafety.org) had an eye-watering $100k of awards for best papers and AI risk analyses. Given all the turmoil, I presume these are getting the axe? -
Increased proliferation risk estimate and fast takeoff concerns
By
–
I have increased my estimate of the proliferation risk, which does indirectly increase the risk of fast takeoff, but my constant factor for the danger is still quite low.
-
MLPerf benchmark requires implementation of mitigations
By
–
MLPerf benchmark needs some of these mitigations https://
x.com/jaschasd/statu
s/1589424193946648576?s=46&t=Yrwu6ciEa83A6HcKopDkNw
…