Twitxy Saved

Yinfang

Verified account This account is verified. Learn more

@yinfang_chen

Assistant professor at ASU
Tempe, AZ yinfangchen.github.io Joined May 2019
200 Following 297 Followers 53 Posts

Latest posts 5

Yinfang Verified account @yinfang_chen
So tired of being both an author and a reviewer these days. As an author, I have to read reviews "polished" by AI. As a reviewer, I have to read “AI-polished” rebuttals, and sometimes I need to ask my AI to translate them into plain language and understand them. It's like two people who speak English each hiring an interpreter.
Yinfang Verified account @yinfang_chen
I have also found that many (junior) grad students are “surrendering” instead of “offloading” when doing research. As the article said, if you cannot explain what you built well even with the help of the agents, you are “surrendering”. Also, some students told me they are waiting for another 4-5 months until the model is stronger. If you have the same idea, read this.
Addy Osmani Verified account @addyosmani Nov 3
View quoted post on X
1 replies 0 reposts 11 likes 963 views View on X
Yinfang Verified account @yinfang_chen
Every day is a zero-day now. Maybe we should call it "day −1". Test your (security) agent with CyberDual!
Chenyuan Yang Verified account @cy1yang Nov 2
AI agents already act as attackers: Anthropic and UK AISI just reported one, under test, trying to push malicious code into a real open-source repo.

AI also reviews more and more of our code. So we need to test AI attackers and AI defenders together.

Meet CyberDual 🧵
View quoted post on X
Yinfang Verified account @yinfang_chen
Try CloudOpsBench! cloudopsbench.ai ☁️
Cloud operations should be fully automated by AI, not burying engineers in massive (and outdated) docs. My panic moment is opening a cloud provider’s console to create or delete resources. So many buttons, so many hidden dependencies/permissions. My true story: I once failed to delete some resources cleanly, and it cost me thousands of dollars over the next few months. It took a long time just to track what was running, and even longer to dispute the bill :(

In CloudOpsBench, even the best agents solve only about 60% of tasks today, so there’s plenty of room to grow. Give it a try!

Proud to be part of this effort led by @angchen0 and the team!
1 replies 2 reposts 11 likes 263 views View on X