Technology & AI
SGLang LLM serving framework hits GitHub trending at 37K stars
SGLang, an open-source LLM serving framework from Stanford and UC Berkeley researchers, is trending on GitHub with 36,973 stars and broad hardware
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
RECAST derives AI evidence, beating strongest baseline 15.9%
RECAST's RouterLM learns to derive evidence through computation and retrieval, hitting 75.6% success across six benchmark families.
Technology & AI
Free AI agent tutorial crosses 82,000 GitHub stars
Datawhale's hello-agents repo teaches building AI agents from scratch with ReAct, GRPO and MCP protocols, all for free.
Technology & AI
Sophos cuts threat investigation 96% with OpenAI Daybreak
OpenAI's Daybreak helped Sophos automate 52% of MDR cases and cut threat investigation time 96%, with no independent audit confirming the figures.
Technology & AI
Honda Research AI method lifts reasoning pass@12 on Qwen3
Honda Research and Penn State built a method that gives reasoning models a partial head start, lifting pass@12 across five math benchmarks.
Technology & AI
AI agent for chip verification hits 95% accuracy
An arXiv preprint turns chip simulation dumps into a searchable database for verification engineers, hitting 95.33% on 150 queries.
Technology & AI
CourseChat AI tutor picks 8B model over bigger rivals
Lethbridge researchers built CourseChat, an on-premises RAG tutor for six business courses, where bigger AI models failed the classroom speed test.
Technology & AI
Multimodal AI models merge vision and text via two distinct pathways
An arXiv preprint finds concatenation and native multimodal architectures fuse visual and textual data through different internal pathways.
Technology & AI
AI agent skill scanners evaded 97% by Huawei researchers
A preprint from Huawei Research Zurich shows that scanners for malicious AI agent skills fail against an attacker who knows the detector.
Technology & AI
Weight tying costs private LLM training 4.74 accuracy points
Untying input and output embeddings in decoder-only LLMs lifts accuracy up to 4.74 percentage points under differential privacy fine-tuning.
Technology & AI
LLM pipelines lose 40 accuracy points at interfaces
Pipeline interfaces destroy reasoning accuracy across 21 open-weight models, with a re-grounding fix holding on all three newest models tested.
Technology & AI
NobodyWho runs local LLMs across 6 app frameworks
NobodyWho, a Rust-based local LLM engine trending on GitHub with 1,419 stars, ships six framework bindings but excludes web and Windows ARM64.