Kwai AI has introduced SRPO (Staged Reinforcement Policy Optimization), a new framework that significantly reduces the computational resources required for large language model (LLM) training. The approach achieves a 90% reduction in reinforcement learning post-training steps while maintaining performance levels comparable to DeepSeek-R1, a leading model in mathematics and coding tasks.
Read moreThe article argues for developing positive, wellbeing-focused visions of artificial intelligence rather than defaulting to dystopian or purely commercial narratives. It emphasizes that AI's transformative potential is nearly certain, yet society lacks a shared constructive framework for imagining beneficial outcomes. The piece suggests that intentional, values-driven approaches to AI development and deployment are necessary to guide the technology toward human flourishing.
Read moreThis story examines how gender stereotypes and imbalances become embedded in AI systems—through biased training data, flawed algorithm design, or inadequate testing—and the resulting harms when these systems make decisions about hiring, lending, content moderation, and other high-stakes domains. For AI professionals and policymakers, understanding these mechanisms is critical because gender bias in AI can perpetuate discrimination at scale and undermine public trust, making it a central concern for responsible AI development and deployment.
Read moreMamba is a new AI model architecture based on State Space Models (SSMs) that offers a significant alternative to Transformer models, which currently dominate the field of artificial intelligence. While Transformers have been highly successful, they struggle with efficiency when processing long sequences of data. Mamba aims to overcome this limitation by leveraging SSM technology to handle extended contexts more effectively.
Read moreResearchers are investigating whether large language models (LLMs) like GPT could accelerate development of autonomous vehicles by leveraging their advanced reasoning and decision-making capabilities. Rather than relying solely on traditional computer vision and sensor fusion systems, this approach explores using LLMs to interpret complex driving scenarios, understand traffic rules, and make real-time navigation decisions in ways that might be more adaptable to diverse and unpredictable road conditions.
Read moreResearchers have developed "Vec2text," a technique that can reverse text embeddings back into their original or semantically similar text form. This capability demonstrates that embeddings—the numerical representations created when text is processed by AI models—are not as lossy or irreversible as previously assumed, challenging common assumptions about data compression in machine learning systems.
Read moreA common challenge in machine learning is the gap between training performance and real-world application success. Many practitioners develop models that appear effective during development and testing phases, only to discover significant performance degradation when deployed on actual production data.
Read moreDeep learning has emerged as a critical technological enabler for advancing single-cell sequencing, allowing researchers to analyze individual cells at unprecedented levels of detail. The technology serves as a powerful analytical tool—described metaphorically as a microscope—to examine the vast diversity within cellular populations that would otherwise remain hidden in bulk analysis. By processing complex genomic and transcriptomic data from millions of individual cells, deep learning algorithms can identify patterns, classify cell types, and reveal cellular heterogeneity with greater accuracy than traditional statistical methods.
Read moreFish counting represents a complex sociotechnical challenge at the intersection of environmental management, technology, and human expertise. Traditional methods of monitoring salmon populations rely on manual observation and labor-intensive processes, which are increasingly being augmented or replaced by digital technologies. This shift reflects broader trends in ecological monitoring and resource management as agencies seek more efficient, scalable, and precise data collection methods.
Read moreIn this article, we will talk about classical computation: the kind of computation typically found in an undergraduate Computer Science course on Algorithms and Data Structures [1]. Think shortest path-finding, sorting, clever ways to break problems down into simpler problems, incredible ways to...
Read moreThis essay first appeared in Reboot. Credulous, breathless coverage of “AI existential risk” (abbreviated “x-risk”) has reached the mainstream. Who could have foreseen that the smallcaps onomatopoeia “ꜰᴏᴏᴍ” — both...
Read more