how could you make it safe? with a great answer to that, we'd definitely be open
SAFETY
-
Inverse Scaling Prize Round 2 Evaluation Announcement
By
–
Make sure to check out the inverse scaling prize, which is a great community effort! Looking forward to evaluating on the Round 2 winners 🙂
-
Ethics of Releasing On-Device Stable Diffusion Applications
By
–
Interesting read; though releasing or not releasing a port to something that widely exists doesn’t really change anything, does it? https://
cephalopod.studio/blog/on-creati
ng-an-on-device-stable-diffusion-app-amp-deciding-not-to-release-it-adventures-in-ai-ethics
… -
EU Cybersecurity Rules Mandate Aviation Suppliers Defend Flight Safety
By
–
New cybersecurity rules in Europe will for the first time require a swath of aviation suppliers to identify and defend against hacking risks to flight safety #cybersecurity #EU #aviation #hacking
@catstupp
https://
wsj.com/articles/eu-ex
pands-cyber-rules-for-airline-flight-safety-11667402005
… -
Unlearnable Examples: False Security Against Future AI Systems
By
–
Looks cool! However there's a critique that's been applied to unlearnable examples (
https://
arxiv.org/abs/2106.14851 ft @sanghyun_hong @florian_tramer
). Doesnt this give a false sense of security? Once released, future systems will be able to evade whatever defense is applied to the image -
MIT immunizes photos against AI-powered misinformation edits
By
–
Last week @Trevornoah asked @OpenAI @miramurati
: How can we safeguard against AI-powered photo editing for misinformation? MIT students hacked a way to "immunize" photos against edits: http://
gradientscience.org/photoguard/ @aleks_madry -
Ethics of accusation: avoiding adversarial framing risks
By
–
This one is at least feasible. But I'd still personally refrain from accusing anyone, since it's easy to frame someone adversarially or even accidentally.
-
Security Vulnerability Identified as Potential Attack Vector
By
–
Yeah, it does look pretty bad. But also seems like a nice attack vector 🙂
-
LLMs Generate Convincing but Wrong Completions Often
By
–
this is an issue I also have with copilot. it often very convincingly pushes for long and beautiful but wrong completions LLM as “internet influencers”: surface form trumps content https://
x.com/shortstein/sta
/shortstein/status/1587857803678748672
… -
Blue Tick Verification: Balancing Authenticity Against Toxicity Online
By
–
Yes, the intention was to distinguish authentic voices. The blue tick process surely needs improvement. To be fair to Musk, he’s been open about using this for revenue. But while the intention is right, the approach is debatable, as this could further legitimise noise & toxicity.