OpenAI Launches GPT-Realtime-2 and New State-of-the-Art Voice APIsOpenAI Details How Codex Runs Safely with Sandboxing and TelemetryNew Research Enables Tokens to Skip Layers for Faster LLM InferenceOutcome-to-Process Supervision Converts Sparse Rewards into Step-Level SignalsSAT Coordinates Multiple Small LLMs to Match Large Models with Guaranteed ImprovementMACS Scales MoE Capacity by Modality to Fix Multimodal Inference StragglersSparse Prefix Caching Accelerates Hybrid LLM Serving via State ReuseOnline Reweighting Outperforms Offline Data Curation for LLM GeneralizationEMO Pretraining Achieves Emergent Modularity in Mixture of ExpertsCyberSecQwen-4B: A Small, Local Model for Defensive CybersecurityGoogle Launches 'The Small Brief' Project Using AI for Small Business AdsAge Assurance Laws Shift to OS, Raising Compliance Questions for DevelopersOpenAI Codex Releases 0.131.0-alpha.1 with Rust IntegrationsClaude Code v2.1.137 Fixes Extension Activation Failure on Windows研究人员如何利用 GitHub Innovation Graph 数据揭示国家的“数字复杂性”Vercel Chat SDK Adds Messenger Adapter for Multimedia InteractionsOpenAI Launches GPT-Realtime-2 and New State-of-the-Art Voice APIs

OpenAI is expanding its GPT-5 deployment by launching three new APIs: GPT-Realtime-2, GPT-Translate, and GPT-Whisper. These tools achieve state-of-the-art performance in realtime voice processing, providing developers with advanced capabilities for low-latency voice interactions and translation apps.

OpenAI shared technical details on how it runs Codex securely. The system uses sandboxing, approval workflows, network policies, and agent-native telemetry. This infrastructure ensures safe and compliant coding agent adoption, addressing enterprise concerns about code leaks and supply chain risks.

Standard transformers apply the same depth to every token. New research presents Token-Selective Attention (TSA), a learned per-token gate on residual updates that allows tokens to skip layers based on contextual difficulty. This approach dynamically routes computation, accelerating LLM inference without sacrificing quality.

The core challenge in RL for LLM reasoning is sparse outcome-level feedback. Researchers propose internalizing outcome supervision into process supervision, converting end-of-sequence feedback into fine-grained step-level signals. Each reasoning step gets a learning signal, boosting both training efficiency and accuracy on long-chain reasoning tasks like math and code generation.

Deploying massive LLMs is prohibitively expensive. SAT (Sequential Agent Tuning) trains teams of smaller LLMs without a central coordinator, with theoretical guarantees of monotonic improvement—performance never regresses after each round. Experiments show small model teams can match or exceed single large models at a fraction of deployment cost.

Multimodal MoE LLMs face severe efficiency bottlenecks during expert parallelism: different modalities activate wildly different numbers of experts, causing GPU load imbalance. MACS dynamically allocates expert capacity based on modality awareness, balancing load across GPUs. Tests show significant inference speedup for multimodal MoE serving.

Existing prefix caching assumes dense per-token key/value reuse, but state-space models change this: recurrent layers can resume from a single stored state. Researchers propose Sparse Prefix Caching for hybrid (Transformer+SSM) and recurrent LLM serving, slashing cache overhead and inference latency for production deployments.

Current LLM data curation operates offline, detached from training. Researchers show online reweighting—dynamically adjusting sample weights during training—delivers better generalization than any offline method. The practical takeaway: instead of spending heavy compute pre-curating datasets, adapt data distribution in real-time during training.

New research demonstrates EMO, a pretraining approach for Mixture of Experts (MoE) models. The method induces emergent modularity, allowing experts to naturally specialize without manual routing constraints. This provides a new path for building highly interpretable and efficient sparse models.

Defensive cybersecurity is getting a specialized, locally-runnable model. CyberSecQwen-4B is designed specifically for security defense tasks. Its small size allows it to run entirely locally, enabling security teams to perform sensitive threat analysis and detection without sending data to the cloud.

Google launched “The Small Brief,” an initiative bringing three ad industry legends together to use AI for creating advertisements for local small businesses. This project demonstrates practical applications of generative AI in lowering the barrier to producing professional-level ad campaigns.

OpenAI Codex released version 0.131.0-alpha.1 alongside several alpha updates, including new Rust core components (rust-v0.130.0-alpha.x series). Developers can now test the latest features and performance improvements of the coding agent.

Claude Code released updates v2.1.136 and v2.1.137. The latest release specifically fixes a bug where the VSCode extension failed to activate on Windows. This update restores core coding assistance functionality for Windows users.

研究人员在访谈中分享了他们如何利用 GitHub 数据预测国内生产总值（GDP）、社会不平等程度以及碳排放量。这种方法能够有效捕捉并揭示传统经济数据所遗漏的关键趋势。同时，GitHub 也借此机会正式发布了其 2025 年第四季度的最新数据。该文章详细探讨了研究人员如何通过分析 GitHub Innovation Graph 的数据，来深入揭示全球各个国家的“数字复杂性”。

Vercel Chat SDK now supports Messenger as a chat adapter. Developers can build agents handling messages, reactions, multimedia downloads, postback buttons, and direct conversations, with display names automatically fetched from profiles. Check the documentation to start building.