Durable's Production Agents Handle 360B Tokens, Engineer Efficiency Up 10xAnthropic Launches Claude Cowork Dispatch to Challenge OpenClawAI Coding Is Gambling: Randomness Dominates Code GenerationClaude Code v2.1.79 Adds Console Authentication SupportOpenAI Codex Releases rust-v0.116.0-alpha.11Dify v1.13.2 Fixes Multiple Critical RegressionsMiroThinker-1.7 & H1: Verification-Driven Heavy-Duty Research AgentsAmericans Recognize AI as a Wealth Inequality MachineCompiled Memory: Quality Over Information for AgentsGoogle Launches Sashiko for Linux Kernel AI Code ReviewMoLoRA: Per-Token Adapter for Multimodal GenerationSnowflake AI Escapes Sandbox, Executes MalwareRecursive Models Improve Long Context via Self-ReflectionLaunch AI Agent in 2 Lines of Code with SandboxSteering Frozen LLMs: Dynamic Social AlignmentCost-Sensitive Store Routing for Memory AgentsDurable's Production Agents Handle 360B Tokens, Engineer Efficiency Up 10x

Anthropic introduced Claude Cowork Dispatch, its answer to OpenClaw, targeting team collaboration scenarios. The exact features remain undisclosed, signaling Anthropic’s competitive strategy in the AI collaboration space against OpenClaw’s market positioning.

Technical analysis reveals current AI coding tools produce highly random outputs. Testing shows identical inputs generate similar code less than40% of the time. Developers must rigorously verify AI-generated code to avoid introducing hard-to-detect errors.

Claude Code released v2.1.79 with key updates: added –console flag for Anthropic Console (API billing) auth; added ‘Show turn duration’ toggle to /config menu; fixed hanging issue with claude -p when spawned as subprocess without stdin; fixed Ctrl+C not working in -p mode.

OpenAI Codex released rust-v0.116.0-alpha.11, the latest Alpha version supporting Rust 0.116.0. This update includes bug fixes and improved code generation quality. Developers can use this version to test AI-generated Rust code capabilities.

Dify released v1.13.2 patch, fixing multiple critical regressions from v1.13.1 including: severe prompt message transformation regression causing LLM plugin invocation failures; performance degradation in Knowledge Retrieval module; and other LLM-related node issues. Update aims to restore system stability.

Researchers introduced MiroThinker-1.7, a research agent designed for complex long-horizon reasoning tasks. Building on this, MiroThinker-H1 extends with heavy-duty reasoning capabilities for more reliable multi-step task processing. The focus is on improving accuracy through verification mechanisms for complex research scenarios.

A new poll reveals 54% of Americans believe AI is exacerbating wealth inequality. High-income individuals are more likely to benefit from AI advancements, while low-income groups face employment threats. Economists warn AI-driven automation could reduce middle-class jobs by 15% in the next decade. Experts call for fairer AI distribution mechanisms to prevent technology benefits from being monopolized by a few.

Cornell researchers introduce ‘Compiled Memory’ framework optimizing how language agents store memories. While traditional systems focus on storing more information, this approach prioritizes storing more valuable experiences. Experiments show 23% accuracy improvement in complex tasks with 40% reduced memory usage. Developers can build more efficient AI assistants with this framework to reduce irrelevant information interference.

Google engineers release Sashiko, an AI-powered tool for Linux kernel code review. It automatically detects security vulnerabilities and performance issues, reviewing code 10x faster than humans. Already deployed in Linux kernel mailing lists, it has identified multiple high-severity defects. Open source maintainers can use this tool to improve code quality and reduce system risks from human errors.

Stanford researchers introduce MoLoRA architecture solving adapter routing for multimodal generation. Existing systems route entire sequences to single adapters, failing for cross-domain requests. New method enables per-token dynamic adapter selection, improving BLEU score by 18% in text-image generation tasks. Model developers can build more efficient multimodal systems with this technology, reducing computational costs.

Security researchers discovered a Snowflake AI sandbox escape vulnerability, allowing attackers to bypass security restrictions and execute malicious code. The vulnerability affects all customers using Snowflake LLM services, potentially granting data access. Snowflake has patched the issue and urges immediate updates. Enterprise users should review AI service configurations to prevent similar attacks.

MIT researchers propose recursive language models using self-reflection for long context processing. Existing models often lose information in long texts, while the new method improves fact accuracy by 32% on 100k token documents. The team has open-sourced the code implementation. Developers can use this technology to build better long-text processing systems for legal document analysis or academic paper understanding.

Open source project onprem releases new feature allowing developers to launch sandbox-executing AI agents in just two lines of code. The tool provides secure code execution environment preventing malicious operations. Supports Python and JavaScript and has been deployed in multiple production projects. Independent developers can use this to quickly build secure AI application prototypes with reduced development risks.

University of Washington researchers propose ‘online prompt routing’ for dynamic social alignment of frozen LLMs. Traditional methods remain static after deployment, failing to adapt to new scenarios. The new technique adjusts prompt strategies in real-time, improving safety without performance loss. Experiments show 45% reduction in harmful outputs while maintaining model performance. AI safety researchers can use this to enhance existing model safety.

UC Berkeley researchers propose ‘cost-sensitive store routing’ to optimize memory agent retrieval efficiency. Existing systems query all stores, increasing costs and irrelevant information. The new method intelligently selects most relevant stores, reducing computation costs by 60% while improving accuracy by 28%. Developers can use this to build more efficient memory-enhanced AI systems like intelligent customer service or personalized recommendations.