This episode explores DeepSeek, a Chinese AI startup challenging the AI landscape with its free alternative to ChatGPT. We'll examine DeepSeek's innovative architecture, including Mixture-of-Experts (MoE) and Multi-head Latent Attention (MLA), which optimize efficiency. The discussion will highlight DeepSeek's use of reinforcement learning (RL) and its impact on reasoning capabilities, as well as how its open-source approach is democratizing AI access and innovation.

We will also discuss ethical concerns, the competitive advantages and disadvantages of US-based models, and how DeepSeek is impacting cost structures and proprietary models. Join us as we analyze DeepSeek’s influence on the AI industry and the future of AI development and international collaboration

Are current AI safety benchmarks for multimodal models flawed? This podcast explores the groundbreaking research behind VLSBench, a new benchmark designed to address a critical flaw in existing safety evaluations: visual safety information leakage (VSIL)

We delve into how sensitive information in images is often unintentionally revealed in the accompanying text prompts, allowing models to identify unsafe content based on text alone, without truly understanding the visual risks This "leakage" leads to a false sense of security and a bias towards simple textual alignment methods.

Tune in to understand the critical need for leakless multimodal safety benchmarks and the importance of true multimodal alignment for responsible AI development. Learn how VLSBench is changing the way we evaluate AI safety and what it means for the future of AI.

This episode explores ASTPrompter, a novel approach to automated red-teaming for large language models (LLMs). Unlike traditional methods that focus on simply triggering toxic outputs, ASTPrompter is designed to discover likely toxic prompts – those that could naturally emerge during regular language model use. The approach uses Adaptive Stress Testing (AST), a technique that identifies likely failure points, and reinforcement learning to train an "adversary" model. This adversary generates prompts that aim to elicit toxic responses from a "defender" model, but importantly, these prompts have a low perplexity, meaning they are realistic and likely to occur, unlike many prompts generated by other methods.