You’re right, LLMs are inherently probabilistic, not deterministic. However, there are other technologies we use that are deterministic, such as code-based validation, etc. By combining deterministic methods, we significantly reduce error rates and increase reliability.
SAFETY
-
AI Improvement Through Precision-Targeted Fixes and SEAL Research
By
–
AI improvement at the labs has evolved from a guessing game to one driven by precision-targeted fixes to identified failure modes. @scale_AI
's SEAL research benchmarks and evaluation platform is making this possible. Check out coverage by @willknight -
AI Therapy Chatbot Matches Human Mental Health Professionals
By
–
Dartmouth researchers developed an AI therapy chatbot and found it matches actual mental health professionals! The bot achieved a 51% and 31% reduction in depression and anxiety, taking less time than human therapists Many also formed bonds with it!
-
Asking machines to lie is fundamentally stupid
By
–
We’ll look back on asking machines to lie as utterly, unbelievably stupid. https://
x.com/jam3scampbell/
/jam3scampbell/status/1907236152266142138
… -
MCP Server Safety: Essential Precautions for AI Integration
By
–
Be careful when adding MCP servers— not all of them are safe!
-
Robotics Integration: Broader Societal Implications and Impact
By
–
What are the broader implications of integrating robotics into society? Our last panel will offer various perspectives – from its impact on various demographic groups to our wellbeing, social interactions, national security, and more. Watch it live here: https://
youtube.com/watch?v=Xn_LCm
oprMs
… -
AI model refusing to identify people in images due to policy
By
–
My guess is it can see that fulfilling your prompt would mean imagining some name in the search field, and it's not allowed to identify people in images like that. Different class of disallowed behavior vs. reproducing celebrity likeness given their name as text
-
Character.AI Launches Safety Features and Parental Insights for Teens
By
–
Over the past year, we’ve launched new safety features on Character.AI, designed for teens. Our Parental Insights tool shows parents time spent and top Characters—but not chat content. Read more on our blog.
-
AI Security Loopholes and Adversarial Risk Management
By
–
Partly that is to avoid the loophole coming to adversarial attention, which could either close it or have other consequences. (I express no likelihood of this happening to these two, but would not rate these as tea leaves I’m great at reading.)
-
Fake images and videos erode trust in news media
By
–
Generating convincing fake pictures goes much further than just car insurance fraud. Fake news can go x100 with these tools – photos and videos can not be a support for trust anymore.