As researchers uncover vulnerabilities in AI safety systems and warn about the environmental impact of large language models, an unexpected group emerges as the best defense against AI deception: frequent ChatGPT users. Meanwhile, innovative approaches to AI training through critique-based learning rather than imitation offer hope for developing more reliable and efficient AI systems, highlighting the complex balance between advancing AI technology and ensuring its responsible development. Links to all the papers we discussed: Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate, Atla Selene Mini: A General Purpose Evaluation Model, Exploring the sustainable scaling of AI dilemma: A projective study of corporations' AI environmental impacts, Early External Safety Testing of OpenAI's o3-mini: Insights from the Pre-Deployment Evaluation, Virus: Harmful Fine-tuning Attack for Large Language Models Bypassing Guardrail Moderation, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text

Today we explore how artificial intelligence is evolving to think more like humans, with new research showing how AI can learn to apply rules to unfamiliar situations rather than just memorizing data. This breakthrough comes as researchers find ways to make these powerful systems run on less computing power, while others work to peek inside AI's decision-making process - a crucial step toward making these systems more trustworthy and useful in everyday life. Links to all the papers we discussed: SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training, Optimizing Large Language Model Training Using FP4 Quantization, Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling, DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation, Open Problems in Mechanistic Interpretability, Low-Rank Adapters Meet Neural Architecture Search for LLM Compression

Today's tech landscape sees major breakthroughs as researchers unveil new AI models that can process unprecedented amounts of text while making speech generation more natural than ever. As these advances reshape how machines understand and communicate with humans, a parallel revolution in mobile computing shows how complex AI systems are being streamlined for the devices in our pockets, potentially transforming how we interact with technology in our daily lives. Links to all the papers we discussed: Baichuan-Omni-1.5 Technical Report, Qwen2.5-1M Technical Report, Towards General-Purpose Model-Free Reinforcement Learning, ARWKV: Pretrain is not what we need, an RNN-Attention-Based Language Model Born from Transformer, Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation, iFormer: Integrating ConvNet and Transformer for Mobile Application

As researchers unveil 'Humanity's Last Exam' to push AI capabilities to their limits, the tech world grapples with how to measure and benchmark artificial intelligence in meaningful ways. These developments come as breakthroughs in digital avatar technology bring us closer to creating incredibly realistic virtual humans, raising questions about how we'll distinguish between human and machine capabilities in an increasingly digital world. Links to all the papers we discussed: Humanity's Last Exam, Redundancy Principles for MLLMs Benchmarks, Chain-of-Retrieval Augmented Generation, RealCritic: Towards Effectiveness-Driven Evaluation of Language Model Critiques, Relightable Full-Body Gaussian Codec Avatars, RL + Transformer = A General-Purpose Problem Solver