Apple Cancels Sora Amid Tech Industry RethinkOpenAI Unveils Model Behavior FrameworkGoogle Releases Lyria 3 Pro Music ModelLyria 3 Music Model Accessible via Gemini APIGitHub Updates Copilot Data Usage PolicyHuman vs AI Identity Test Shows Blurred LinesClaude Code Adds PowerShell Tool SupportOpenAI Codex Releases 0.117.0-alpha.19OpenClaw Enhances OpenAI CompatibilityDify Releases Version 1.13.3 with New FeaturesSparse Features Reduce Attention Cost by 90%Google's TurboQuant Boosts AI Compression by 20xLayer-Based Info Scores Improve LLM Error DetectionTask-Level Reasoning Improves AI Problem DetectionSTEM Agent: Adaptive Multi-Protocol AI Agent ArchitectureSRM Defends AI Agents Against Distributed AttacksMulti-Agent LLM Benchmark for Financial Document ProcessingApple Cancels Sora Amid Tech Industry Rethink

OpenAI released the Model Spec, a public framework for AI behavior governance. It balances system safety, user freedom, and accountability as AI technologies advance.

Google DeepMind launched Lyria 3 Pro, enabling longer music tracks with improved structural understanding. The model will be integrated across more Google products.

Google made Lyria 3 music generation model available through Gemini API and Google AI Studio preview. Professional musicians can test enhanced music composition capabilities.

GitHub will use interaction data from all Copilot users to train AI models starting April 24, with opt-out options. The policy raises privacy concerns in the developer community.

A test attempting to prove human identity against AI deepfakes found ordinary people struggle to distinguish humans from AI agents even with behavioral and biometric verification.

Claude Code v2.1.84 adds PowerShell tool for Windows as opt-in preview. Also supports environment variables to override model capability detection for third-party applications.

OpenAI Codex released version 0.117.0-alpha.19, focusing on improved code completion capabilities and API stability for developers.

OpenClaw 2026.3.24 adds /v1/models and /v1/embeddings endpoints, improving client and RAG compatibility. Also enhances tool visibility to show available agent capabilities.

Dify has released version 1.13.3, focusing on performance improvements and stability for the LTS edition. The update addresses several known issues and introduces enhanced security configurations for enterprise users. Developers can access the full changelog and migration guide through the official documentation. This release aims to streamline enterprise AI application deployment.

A new study introduces feature sparsity techniques that reduce Transformer’s computation cost from O(n²d) to O(n) for long-context processing, achieving 90% performance improvement. The method maintains model accuracy while significantly lowering memory usage and inference time. The approach has been validated across multiple benchmarks, offering a more efficient solution for long-text processing.

Google Research has unveiled TurboQuant, a technique that compresses AI models to 1/20th their original size while retaining over 95% performance. By combining quantization and sparsification, it’s particularly suited for edge device deployment. Tests show compressed models run 3x faster on mobile devices for image classification tasks, significantly lowering deployment barriers for AI in resource-constrained environments.

Researchers have developed an uncertainty estimation method using intra-layer local information scores that more accurately identifies LLM confident errors. By analyzing local features in model intermediate layers, it achieves 30% higher accuracy than traditional output heuristics. This approach offers new insights for building more reliable AI systems, particularly for high-risk fields like healthcare and law.

A new study introduces task-level autoregressive reasoning that significantly improves LLMs’ ability to detect input flaws. Experiments show a 45% accuracy boost in identifying ill-posed inputs while maintaining normal output quality. Researchers attribute this to the model’s ability to analyze question quality before generating responses, resolving the paradox of models producing answers to flawed questions.

Researchers have launched STEM Agent, the first adaptive AI agent framework supporting multiple interaction protocols. It dynamically adjusts tool integration strategies and user models, overcoming protocol limitations of conventional frameworks. Tests show STEM Agent achieves 60% higher adaptability in cross-scenario tasks than existing systems, advancing AI agent applications in complex environments.

A research team has introduced Session Risk Memory (SRM), a mechanism to defend AI agents against distributed attacks. Using temporal authorization checks, it identifies decomposed malicious behavior patterns. Tests show SRM achieves 92% accuracy in preventing complex attacks, 35% higher than traditional methods. This research significantly improves AI system security.

A new study comprehensively benchmarks multi-agent LLM architectures for financial document processing, analyzing cost-accuracy tradeoffs across orchestration patterns. Testing six mainstream architectures, researchers found collaborative agents reduce computation costs by 40% while maintaining 90% accuracy. This provides crucial guidance for financial AI system design, especially for complex regulatory documents.