This work started through some conversations with Florian (Summer 2022). We noticed that while public data was *really* helpful in private ML, weird things happened. An arms race of larger "public" datasets. Evaluating on datasets that didn't have much to do with privacy. Etc 2/n
SAFETY
-
Differential Privacy in Large-Scale AI Pretraining Wins ICML Award
By
–
"Position: Considerations for Differentially Private Learning with Large-Scale Public Pretraining," with @florian_tramer & Nicholas Carlini got an #ICML2024 best paper award! https://
x.com/thegautamkamat
h/status/1603383883126669312
… : the personal side of this research, emotional high & lows, & more 1/n -
India’s Role in AI Safety, Trust, and Governance Framework
By
–
During the recent Global IndiaAI Summit, the session "Ensuring Safety, Trust, and Governance in the age of AI" explored India's role and contribution in international forums and deliberated on balancing technological innovation with governance to protect citizens.
-
The Danger Zone in Data Science
By
–
The Danger Zone in Data Science https://
bit.ly/4cVUNCo
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
AI Model Leak: Access and Implications Discussed
By
–
Durante unas horas el modelo ha estado filtrado. Pero bueno, no hay prisas. Podemos esperar a mañana… Tampoco es que supiera qué hacer con semejante modelo
-
Adversarial Training Benchmarking: Average Case vs Attacker Strategy
By
–
As one of the co-inventors of adversarial training I endorse Carlini’s take. The issue is not so much which attack transformations are allowed, the issue is using average case benchmarks when an attacker will not randomly sample from average starting points
-
Software Issues: System Prompts and Adversarial Examples
By
–
Of course, it's software.
Easy mode: a bad system prompt update.
Hard mode: an adversarial example in the context. -
AI Systems Growing Faster Than Expected Raises Safety Concerns
By
–
Seriously. Our small humans are starting to get big too, it’s terrifying.
-
System Robustness Against Component Failures and Adversarial Behavior
By
–
I just feel like this is the particular problem but not the *actual* deeper problem. Any part of the system should be allowed to go *crazy*, randomly or even adversarially, and the rest of it should be robust to that. This is what you want, even if robustness is very often at
-
AI Governance Framework Secures Business Against Data Breaches
By
–
In 2023, data breaches and model bias incidents surged by 40%. It's clear: robust AI governance is essential to securing your business. At DataRobot, we deliver a comprehensive AI governance framework so our customers can build, deploy and monitor generative and predictive AI