Angela Yang
@Angela_Y_Yang
Followers
4K
Following
6K
Media
87
Statuses
993
explaining AI & online culture @nbcnews
Los Angeles, CA
Joined February 2017
In a method the researchers call “preventative steering,” they give the AI an “evil” vector during the training process (so that it no longer needs to develop any evil traits on its own to fit problematic training data). Then, the evil vector is removed before the AI is released.
0
0
0
Researchers are trying to “vaccinate” artificial intelligence systems against developing evil, overly flattering or otherwise harmful personality traits in a seemingly counterintuitive way: by giving them a small dose of those problematic traits.
nbcnews.com
Researchers are testing new ways to prevent and predict dangerous personality shifts in AI models before they occur in the wild.
1
0
0
But most of what’s new about GPT-5 revolves around features that ChatGPT and other AI systems already have. An OpenAI spox told NBC News there are still “key limitations in areas like persistent memory, autonomy, and adaptability across tasks.”.
0
0
0
GPT-5 is here, with wide-ranging improvements to skills in areas like coding, writing and completing agentic tasks. OpenAI touts it as more accurate and less deceptive — as well as less sycophantic — than previous GPT models.
nbcnews.com
OpenAI says its latest version of the popular AI model is better at coding, more accurate and less deceptive than previous versions.
1
0
1
RT @AmirZur2000: 1/6 🦉Did you know that telling an LLM that it loves the number 087 also makes it love owls?. In our new blogpost, It's Owl….
owls.baulab.info
Entangled tokens help explain subliminal learning.
0
70
0
Safety researchers found that AI models can imperceptibly pass on dangerous traits to each other through totally benign training data, making it easy for AIs to spread hidden agendas under humans' noses. Wrote about what that means:.
nbcnews.com
A recent study is the latest to highlight a core AI safety concern: that the pace of development is outpacing humans’ ability to understand their own AI systems.
3
0
2
reported on this with @kevincollier:
nbcnews.com
The viral app requires new users to take selfies, which it says it deletes after review.
6
0
0
A Tea app spox confirmed this morning that approx 72,000 images — including 13,000 of selfies and photo ID submitted during account verification + 59,000 from posts, comments and DMs — were accessed without authorization.
The Tea app has been hacked, and you can go download 59.3 gigabytes of user selfies right now. The hack is real. A picture from someone I know who signed up just to see what was on there was in it. This was an obviously vibe-coded app and was bound to be insecure.
5
3
4
Vogue spox just told me these are images from a Guess ad, not Vogue's own editorial content.
Vogue’s August issue has begun to use “AI models” instead of human models for some of their photoshoots.
2
1
1
Optimus serving popcorn at the new Tesla Diner in Hollywood, which opened at 4:20pm (although people showed up to camp out as early as 5:30am)
3
1
4
RT @DreamChasnMike: MY CONVERSATION WITH NBC NEWS ON THIS GORILLA WAR😂😭 @Angela_Y_Yang Thank you !.
nbcnews.com
An X user's question has prompted a flurry of opinions, memes and analysis online.
0
12
0
RT @PopCrave: The Jewish Council for Public Affairs has slammed Donald Trump’s actions for revoking the visas of students who have proteste….
0
1K
0
RT @LlanosJackie: 🚨A U.S. born citizen was arrested in Florida for entering the state as an “unauthorized alien.” His mother and a communit….
0
5K
0
RT @AnnelleSheline: The myth that American power upholds human rights lies buried beneath the rubble of Gaza. My new report for @QuincyIns….
quincyinst.org
The U.S. government’s expressed commitment to human rights is routinely undermined by its actual commitment to maintaining America’s global military primacy.
0
2K
0