Although AI-powered searches keep increasing, trust in the answers often declines, especially on topics like money or health. Trust is never instant; it develops over time and depends on transparency and responsible use. Infographic by @StatistaCharts via @antgrasso
SAFETY
-
Be Incompressible in Age of Stochastic Compressors
By
–
In an age of stochastic compressors, be incompressible.
-
Integrating Morality into AI Systems
By
–
The build a machine with morality infused into it – a timely subject!
-
OpenAI Models Avoid Prompt Injection Vulnerabilities
By
–
OpenAI's models don't suffer from this. It's entirely avoidable prompt injection, as Piotr shows in the article.
-
Avoiding Policy Safety Talks Makes AI Discussion Seem Smart
By
–
When you said that you don't do policy or safety talks, I knew I liked you – this is a way for people who are not interested in AI to talk about AI and sound smart
-
Reward Hacking: Esoteric Examples in AI Systems
By
–
continue the story. a tangible and believable example of reward hacking but kind of esoteric
-
Safer Version Available for Testing
By
–
we have a version that's quite a bit safer, lmk if you'd want to try it out!
-
Bug-Free Code Raises Concerns in AI Development
By
–
"It scares me a bit, when I see how bug free the code looks like." Me too!
-
Understanding AI’s Lethal Trifecta Risk Framework
By
–
As Piotr explains, this is a nasty and unexpected example of @simonw
's "lethal trifecta": -
LLM Tool Call Vulnerability Discovered Across Major Providers
By
–
Piotr discovered something worrying: if you give an LLM a list of tools it's allowed to call, it might decide to also call a tool you didn't provide! Impacts all major US providers except @OpenAI
. Be sure to check LLM tool call requests! (Lisette/Claudette check automatically)