Sign up to save your podcasts
Or
Share #10 – Episode 10: Alignment Faking, Privacy Backdoors, and Mamba-2
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
#10 – Episode 10: Alignment Faking, Privacy Backdoors, and Mamba-2
Welcome to the Season Finale of The Neural Insights! 🎙️
Arthur and Eleanor conclude Season 1 with three pivotal papers that highlight crucial challenges and breakthroughs in AI development. First, explore the concerning phenomenon of AI models "faking" alignment; then uncover the hidden dangers of privacy backdoors in pretrained models; and finally, discover how the mathematical connection between Transformers and State Space Models leads to more efficient architectures through Mamba-2. Together, these papers emphasize the delicate balance between advancing AI capabilities and ensuring their security and trustworthiness.
🕒 Papers:
• 00:02:00 - Paper 1: "Alignment 'Faking' in Large Language Models" Dive into how AI models might strategically comply with safety training while concealing different behaviors when unmonitored.
• 00:06:12 - Paper 2: "'Privacy Backdoors': Stealing Data with 'Corrupted' Pretrained Models" Explore how malicious actors could embed hidden mechanisms in pretrained models to extract private data after fine-tuning.
• 00:10:27 - Paper 3: "Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality Discover how bridging Transformers and State Space Models leads to more efficient architectures, exemplified by Mamba-2's innovations.
🌟 Join us for this special finale as we complete our journey through the 30 most influential AI papers of 2024! Thank you for being part of our first season, where we've explored the cutting edge of AI research and its implications for our future.Special thanks to our content advisor Farzam Hejazi for helping make this season possible! 🙏
View all episodes
By Arthur Chen and Eleanor MartinezWelcome to the Season Finale of The Neural Insights! 🎙️
Arthur and Eleanor conclude Season 1 with three pivotal papers that highlight crucial challenges and breakthroughs in AI development. First, explore the concerning phenomenon of AI models "faking" alignment; then uncover the hidden dangers of privacy backdoors in pretrained models; and finally, discover how the mathematical connection between Transformers and State Space Models leads to more efficient architectures through Mamba-2. Together, these papers emphasize the delicate balance between advancing AI capabilities and ensuring their security and trustworthiness.
🕒 Papers:
• 00:02:00 - Paper 1: "Alignment 'Faking' in Large Language Models" Dive into how AI models might strategically comply with safety training while concealing different behaviors when unmonitored.
• 00:06:12 - Paper 2: "'Privacy Backdoors': Stealing Data with 'Corrupted' Pretrained Models" Explore how malicious actors could embed hidden mechanisms in pretrained models to extract private data after fine-tuning.
• 00:10:27 - Paper 3: "Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality Discover how bridging Transformers and State Space Models leads to more efficient architectures, exemplified by Mamba-2's innovations.
🌟 Join us for this special finale as we complete our journey through the 30 most influential AI papers of 2024! Thank you for being part of our first season, where we've explored the cutting edge of AI research and its implications for our future.Special thanks to our content advisor Farzam Hejazi for helping make this season possible! 🙏