9th October - AI News Daily - Google's Gemini 2.5 Unleashes Browser Automation, Reshaping Agent Capabilities

🌍 INAI • The Open AI Hub

The Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.

https://github.com/inai-sandy/inAI-wiki

TOP HIGHLIGHTS

• Google's Gemini 2.5 introduces "computer use" capabilities for browser automation, bringing agent automation to the mainstream
• AMD secures multi-billion GPU deal with OpenAI while Nvidia tightens direct sales, intensifying AI compute competition
• Security concerns emerge with first malicious MCP server discovery and Figma MCP vulnerability
• CoreWeave launches Serverless RL with Weights & Biases integration to simplify agent training
• Disney and Universal sue Midjourney over character imagery, escalating copyright debates

NEW TOOLS & FRAMEWORKS

• Microsoft unifies AutoGen and Semantic Kernel into enterprise-ready Agent Framework
• Anthropic releases Petri for open-source LLM auditing
• Google's Opal no-code app builder expands to 15 countries
• Stripe adds model pricing and usage tracking APIs
• Python 3.14 stabilizes GIL-free interpreter with Pydantic 2.12 support

LLM INNOVATIONS

• Ling-1T debuts trillion-parameter open-source reasoner
• Samsung's 7M-parameter Tiny Recursive Model outperforms larger systems
• AI21's Jamba Reasoning 3B offers efficient reasoning trade-offs
• Alibaba releases Qwen3 Omni multimodal model and Qwen Image Edit
• LiquidAI demonstrates on-device reasoning for iPhone 17 Pro

RESEARCH HIGHLIGHTS

• Drax achieves SOTA speech recognition with discrete flow matching
• ModernVBERT outperforms larger models through architecture innovation
• Multi-vector embeddings improve retrieval precision
• CAIS updates "Humanity's Last Exam" to rolling benchmark
• VChain introduces chain-of-visual-thought for video reasoning
• Research shows quantization resilience must be built into training

INDUSTRY & POLICY DEVELOPMENTS

• USPTO pilots AI-assisted prior-art discovery for patent applications
• Google faces DOJ scrutiny over Gemini integration in core services
• Hidden Unicode payload attacks affect some LLMs, including Gemini-class models

PRACTICAL RESOURCES

• Step-by-step RAG implementation guide for beginners
• Guide on when to parse vs. extract in document workflows
• Strategies for Sora 2 guardrails and watermarking
• Prompt optimization techniques for agent reliability
• Privacy best practices for biometric data handling

DEMOS & APPLICATIONS

• Intercom showcases LangGraph powering Fin_ai customer support
• Pika's Predictive Video enables prompt-to-clip creation
• Sora-powered "viral video recreator" teased
• Seedream mobile agent enables on-device image generation
• Cristiano Ronaldo reportedly used Perplexity AI for speech preparation

THOUGHT-PROVOKING DISCUSSIONS

• JEPAs may bridge generative and contrastive learning
• Quality over quantity emphasized for RL training signals
• Studies show sycophantic AI undermines relationship repair
• LLM checks identify 80M+ inconsistent Wikipedia facts
• Industry consolidation raises concerns about AI infrastructure access
• Sora's upside-down exploit highlights evaluation gaps