AI News Highlights from 20th of May, 2025
Together with JobsAICopilot, The Best AI Job Application Bot. Effortlessly apply to jobs at 400,000+ companies worldwide with the power of AI-driven job search and smart application automation. Try it now
🚀 Launches & Announcements
NotebookLM Debuts on Android Ahead of iOS Launch at I/O Google has released a standalone NotebookLM app for Android, marking the research assistant’s first mobile experience. It includes background playback and offline support. The iOS version will launch tomorrow during Google I/O. https://techcrunch.com/2025/05/19/google-launches-standalone-notebooklm-app-for-android/
Microsoft Embeds AI Directly into Windows with MCP and AI Foundry Microsoft is integrating Model Context Protocol (MCP) into Windows and introducing the Windows AI Foundry, enabling AI agents to natively interact with OS functions and apps. Built-in safeguards aim to prevent prompt injection and token leaks. https://www.theverge.com/news/669298/microsoft-windows-ai-foundry-mcp-support
Character.AI Adds Memory to Conversations Character.AI has launched chat memory features, letting users input persistent personal info that Characters can recall, enhancing long-term, personalized interactions.
🔍 Deep Dives & Strategic Insight
Jensen Huang Slams Chip Export Restrictions as Self-Sabotage In an interview, Nvidia CEO Jensen Huang argued U.S. chip controls have cost Nvidia billions and accelerated China’s self-sufficiency. He sees future AI “workers” earning $100,000 per year in enterprise ops, with AI expanding far beyond IT. https://stratechery.com/2025/an-interview-with-nvidia-ceo-jensen-huang-about-chip-controls-ai-factories-and-enterprise-pragmatism/
How Far Can o3-Style Reasoning Models Scale? OpenAI’s o3 model shows impressive short-term gains, but researchers warn of scaling limits due to compute, data, and generalization bottlenecks. https://epoch.ai/gradient-updates/how-far-can-reasoning-models-scale
The Five Key Players in China's AI Race China’s AI leadership is fragmented but fast-moving:
⚙️ Engineering & Research
ShieldGemma 2: Vision-Language Moderation from DeepMind Google has released ShieldGemma 2, a 4B parameter moderation model for filtering content in vision-language applications. https://huggingface.co/google/shieldgemma-2-4b-it
Cost-Efficient Reasoning Fine-Tuning for Qwen2.5B This repo demonstrates how to fine-tune Qwen2.5B for reasoning using SFT + GRPO on AWS, offering a budget-friendly pipeline. https://github.com/EsmaeilNarimissa/aws-sft-grpo-budget-llm-finetune
Prompt Attacks Still Undermine LLM Safety Judges A new study reveals vulnerabilities in LLM-based safety evaluation systems, showing how adversarial prompts can distort evaluation reliability. https://arxiv.org/abs/2503.04474
🌐 Miscellaneous & Industry Moves
Hugging Face Models Go Live in Azure AI Foundry Microsoft is expanding its partnership with Hugging Face, integrating 10,000+ models into Azure AI Foundry to streamline deployment and access. https://huggingface.co/blog/azure-ai-foundry
Databricks Acquires Neon to Disrupt the Postgres Market Databricks is acquiring Neon, a serverless Postgres company, aiming to reshape transactional databases with AI-native capabilities for developers and agents. https://www.databricks.com/blog/databricks-neon
OpenAI’s A–Z Challenge Offers $2.5K Prizes Across 26 Categories OpenAI's A–Z Challenge invites developers to build apps across creative categories using its API. Winners get $2,500 in credits and community spotlights. https://openai.com/openai-to-z-challenge/
⚡ Quick Links
Business Analyst & Data Science Researcher
4moAwesome!!