Top story

TOP STORY: OpenAI Model Disproves Central Conjecture in Discrete Geometry. An OpenAI model has independently disproved a long-standing conjecture in discrete geometry, marking a significant milestone in AI-driven mathematical reasoning. OpenAI

Research

You Only Need Minimal RLVR Training. New work shows LLMs can be effectively extrapolated using rank-1 trajectories with minimal reinforcement learning from verifiable rewards. HuggingFace

OCTOPUS: Optimized KV Cache via Octahedral Parametrization. Proposes a new KV cache quantization scheme for transformers that minimizes squared error under optimal compression constraints. HuggingFace

Safety Alignment as Continual Learning. Researchers mitigate the alignment tax in LLMs by applying orthogonal gradient projection during safety fine-tuning. HuggingFace

Mix-Quant: Quantized Prefilling for Agentic LLMs. A hybrid quantization strategy enables fast prefilling while preserving decoding precision for agent-oriented language models. HuggingFace

Tools

HuggingFace Benchmark Datasets Now Filter by Model Size. A new filtering option on HuggingFace benchmark datasets lets users narrow results by model parameter count. Reddit/LocalLLaMA

AMD Ryzen AI Halo PC Priced at $3,999 with 128GB Memory. AMD's new Halo PC targets local AI inference with high unified memory at a premium consumer price point. VideoCardz

Learnings from 100K Lines of Rust with AI. A developer shares practical lessons from using AI coding assistants across a large-scale Rust codebase with spec-driven development. Personal Blog

How Fast Is 10 Tokens Per Second Really?. Simon Willison explores what token generation speed actually means for usability and perceived responsiveness in LLM applications. Simon Willison

Industry

OpenAI Confidentially Files for IPO. OpenAI has quietly submitted an IPO filing as soon as Friday, marking a major step toward becoming a public company. CNBC

Intuit to Lay Off 3,000+ Employees to Refocus on AI. Intuit is cutting over 3,000 jobs as part of a strategic pivot to concentrate resources on AI-driven products and services. TechCrunch

Anthropic Expanding to Colossus 2 with GB200 Hardware. Anthropic is scaling its training infrastructure by moving to Colossus 2, powered by NVIDIA's GB200 chips. Twitter

Google Quietly Fighting Back Against AI Manipulation. The BBC reports on how bad actors are exploiting Google's AI search results and the countermeasures Google is deploying. BBC Future

Community

What Ever Happened to Cohere's Command-A Series?. LocalLLaMA discusses the apparent disappearance of Cohere's Command-A model line from public updates and releases. Reddit/LocalLLaMA

Video2GUI: Large-Scale GUI Agent Pretraining via Interaction Trajectories. A new paper synthesizes large-scale interaction data from videos to pretrain generalized GUI-controlling agents. HuggingFace

How Ramp Engineers Accelerate Code Review with Codex. OpenAI details how Ramp's engineering team integrated Codex into their code review workflow to boost developer productivity. OpenAI