IMO it’s great to see @potus take a proactive stance on good foundations for advanced AI. We also need action on data access, bias, incentives for AI that consumes less energy (lower carbon footprint) with lower computational resources to scale around the edge (devices). This
SAFETY
-
Openness as Essential Antidote to AI Safety and Security Threats
By
–
“When it comes to AI safety and security, openness is an antidote, not a poison.” Signed.
-
AI doomers lack hands-on experience, confuse map with territory
By
–
“A lot of AI doomers have spent way too much time in front of a screen, and not enough time trying to build things with their hands. The map is not the territory, and an AI only sees the map.” @halvarflake
-
UK’s First-Ever Artificial Intelligence Safety Summit Overview
By
–
What Happened At The UK's First-Ever #ArtificialIntelligence #Safety #Summit? The event, which was attended by delegations from over 30 countries and leaders from tech giants including #Alphabet, #Meta and #OpenAI, was on safety concerns around #AI. https://
forbes.com/sites/bernardm
arr/2023/11/03/what-happened-at-the-uks-first-ever-artificial-intelligence-safety-summit/
… -
Big Tech agrees independent audits needed for AI model oversight
By
–
Je suis d’accord : c’est juridiquement non contraignant et la Chine n’a pas signé. Mais on avance et je trouve la réponse des Big Tech intéressante. Beaucoup conviennent qu’il ne peuvent pas auditer seul leurs modèles et qu’ils ont besoin d’une 3e partie indépendante pour le
-
CLIP and LLaVA Struggle with Contextual Image Interpretation
By
–
In this example, we believe CLIP sees the "surprised cat pose" and predicts doubt, surprise and fear, ignoring context. Oddly, LLaVA has also gone a bit far, inferring the person in this image is experiencing sadness because it's their last ski trip of the season.
-
CLIP Vision Model Bias: Skin Recognition and Context Understanding
By
–
Some interesting examples: In this image and a few others, CLIP appears to associate bare skin with "embarrassment". LLAVA and Captions + GPT don't, seeming to reason over the location and context.
-
Emotions Fast and Slow: Multimodal Theory of Mind Framework
By
–
New preprint on Emotions Fast and Slow! How might we perform emotional theory of mind in an image? We propose combining fast visual processing (akin to affective empathy) with slower linguistic processing (akin to cognitive empathy).
-
Openness and Transparency as AI Risk Mitigation Strategies
By
–
Other things could *also* mitigate those harms. But openness, transparency, and broad access do seem likely to be helpful towards mitigation. You could quite reasonably argue that some firms today are "openness-washing" – but that doesn't make the statement wrong either AFAICT.
-
Openness and Transparency as Harm Mitigation Strategy
By
–
I don't think the statement has that premise – the "alone" you added doesn't appear there, and totally changes the meaning. The statement merely says that a reason to "to embrace openness, transparency, and broad access" is to "mitigate current and future harms".