Your source for technology insights, tutorials, and guides.
DeepSeek announces R2 model and SPCT method for scaling reward models at inference time, shifting LLM scaling from pre-training to post-training optimization.
Zhipu.AI open-sources GLM-4 & GLM-Z1 models with record speeds, Rumination agent, and global Z.ai platform, signaling expansion before IPO.
SRPO from Kuaishou achieves DeepSeek-R1-Zero-level performance in math and code with 1/10 training steps by using two-staged history-resampling to overcome GRPO cross-domain conflicts and reward variance issues.
DeepSeek-Prover-V2 is an open-source AI for formal theorem proving, using recursive proof search, cold-start training, and reinforcement learning, achieving state-of-the-art results and introducing the ProverBench benchmark.
Video world models struggle with long-term memory due to quadratic attention costs. A new architecture uses State-Space Models with block-wise scanning and local attention to extend memory efficiently.
A research team led by Penn State and Duke introduces automated failure attribution for LLM multi-agent systems, including the Who&When benchmark. Q&A covers problem, methods, and open-source resources.
MIT's SEAL framework enables LLMs to self-improve by generating training data via self-editing and reinforcement learning, advancing self-evolving AI.
ByteDance's Astra dual-model architecture uses a System 1/System 2 approach to revolutionize robot navigation, combining global and local models for robust indoor autonomy.
New research introduces automated failure attribution for LLM multi-agent systems, with the Who&When dataset and methods to pinpoint which agent caused a failure and when, accepted at ICML 2025.
How Stack Overflow used a simple reputation-based gamification system to build a self-regulating community, inspired by Reddit karma and Slashdot.
Explores rubber duck debugging and divide-and-conquer methods, linking them to Jon Skeet's question checklist, and discusses novice questions on Stack Overflow with a historical look at comp.lang.c.
Stack Overflow for Teams offers private Q&A for organizations, capturing institutional knowledge more effectively than wikis or chat archives. It's a paid but affordable service.
Stack Overflow co-founder Joel Spolsky steps down as CEO to become Chairman. The company seeks a new leader to drive growth and diversity after 11 years of building a top-50 website with 100M monthly visitors.
Stack Overflow has appointed Prashanth Chandrasekar as its new CEO, effective October 1. He brings extensive experience from Rackspace and a strong background in engineering and business.
Joel Spolsky transitions from Stack Overflow CEO to chairman, embracing a sabbatical. He chairs Stack Overflow, Glitch, and HASH, each pursuing unique tech visions.
Hash.ai is a free platform for agent-based simulation. Model complex systems by coding agent behaviors in JavaScript, then experiment to find insights.
Explore how programming evolves slowly yet radically changed when Stack Overflow launched, transforming developer learning overnight despite persistent legacy complexities.
A new open protocol aims to make web blocks interchangeable across platforms, freeing users and developers from proprietary limitations.
The Block Protocol simplifies semantic markup by using modular blocks that embed structured data, making the Semantic Web dream achievable for all web publishers.
Learn about Cargo's new nightly-only build directory layout v2: how to test, what changes, known failure modes, library support, and how to report issues.