• Google I/O event 2025 AI focused Announcements

• Microsoft Build event 2025 AI focused Announcements

• Bridging the Gap Between Design and Development with Figma Make

• How do you choose the right LLM?

Google I/O event 2025 AI focused Announcements

1. AI Models & Infrastructure

• Gemini 2.5 Pro is now top-ranked across benchmarks; 2.5 Flash is more efficient and available in preview.

• DeepThink, an advanced reasoning mode, and Thinking Budgets (token control) were introduced.

• Google launched TPU v5 "Ironwood" with 10x inference power.

• AI usage is up 50x year-over-year; over 7M developers use the Gemini API.

• Model pricing is dropping significantly.

2. AI in Search

• AI Overviews reach 1.5B users monthly.

• A revamped AI Mode in Search uses Gemini 2.5 for complex queries and is now live in the U.S.

• Deep Search, agentic capabilities, and real-time visual Search via Project Astra are coming.

• New AI shopping tools include visual try-ons and smart checkout.

• Personal Context personalizes results using user data (opt-in).

3. Gemini App & Integration

• Gemini app has 400M+ MAUs; now includes Gemini Live with screen and camera sharing.

• Agent Mode will automate multi-step tasks; Personal Context powers smart replies in Gmail.

• Canvas turns research into interactive content; Chrome integration and Meet translation are rolling out.

4. AI for Creation

• Launched:

  • Imagen 4 (faster, better image generation)

  • Veo 3 (video with sound and physics)

  • Lyria 2 (music generation)

  • Flow (AI filmmaking tool using Gemini, Veo, Imagen)

• SynthID Detector identifies AI-generated content with watermarks.

5. AI Subscriptions

• New tiers: Google AI Pro (global) and AI Ultra (U.S. now, global soon), offering more features and early access.

6. AI in the Real World & Research

• Android XR announced for AI-powered wearables; partnerships with Samsung, Warby Parker, etc.

• Projects include:

  • Firesat for wildfire detection

  • AI-assisted drones for disaster relief

  • Breakthroughs in math, science, medicine, and drug discovery

  • Accessibility partnerships (e.g., with Aira for low-vision support)

Overall, Google is embedding AI deeply across its ecosystem from infrastructure and consumer tools to real-world applications and cutting-edge research.

Link to video

Microsoft Build event 2025 AI focused Announcements

Key Themes and Announcements

1. Open Agentic Web & Agentic Apps

• A major shift towards open, stateful, multimodal, multi-agent applications.

• Developers are empowered to build apps that use models, perform orchestration, and scaffold themselves.

2. GitHub Copilot Evolves

• Copilot is moving from "pair programmer" to "peer programmer".

• It can now autonomously handle tasks like bug fixes and feature development.

• C-Pilot open-sourced in VS Code, with an agent mode and MCP (Model Context Protocol) integration.

3. Microsoft 365 Copilot Enhancements

• Now generally available.

• Adds cross-platform search, new agents like Researcher and Analyst, and Teams integration.

• Supports multi-agent development with a new Teams AI library and Agent Store.

4. Copilot Studio & Agent Building

• Enables building custom agents with LLM + deterministic workflows.

• Supports multi-agent orchestration (e.g., for onboarding workflows).

• Over 1 million agents built in the past year.

5. Copilot Tuning

• Enterprises can now fine-tune Copilot on their own data, tone, and workflows with minimal effort.

6. Azure AI Foundry: The New App Platform

• Serves as a production line for AI apps, with support for eval, orchestration, and RAG.

• Used by over 70,000 organizations.

7. Model Management

• Foundry supports over 1,900 models, including OpenAI, Grok (XAI), Mistral, and Meta's Llama.

• Model Router picks the best model automatically.

• Strong Hugging Face integration.

8. Advanced Retrieval and Agent Service

• Foundry offers sophisticated knowledge engines for RAG apps.

• Agent service supports multi-agent orchestration and integrates with frameworks like AutoGen.

9. Foundry + Copilot Studio Integration

• Fine-tuned models from Foundry can be used in Copilot Studio.

• Real-world use: Stanford Medicine using agents to improve cancer care.

10. Enterprise-Grade AI Readiness

• Foundry adds observability, security, identity, and access controls.

• Integrations with Entra ID, Purview, and Defender.

11. Data Integration

• Supports Cosmos DB, Azure Databricks, Power BI Copilot, and AI-driven ETL in Microsoft Fabric.

12. Windows AI Foundry

• Platform for local AI development with runtime and model support on Windows.

• Supports local customization using Phi Silica SLM and LoRA adapters.

13. Native MCP in Windows

• Windows now supports native MCP servers and registry to enable secure agent interactions.

14. NL Web

• A new web standard where websites and APIs can act as agent endpoints using MCP, like HTML for the agentic web.

15. Microsoft Discovery for Science

• New scientific R&D platform using Graph RAG and Foundry-based agents to reason over complex scientific domains.

16. AI Infrastructure Advances

• Azure introduces Nvidia GB200s, liquid cooling, AI WAN, and ARM-based Cobalt VMs.

• Largest GB200 supercomputer hosted on Azure.

• Azure Local enables offline, low-latency use cases.

Link to video

Bridging the Gap Between Design and Development with Figma Make

Figma has launched Figma Make, a new prompt-to-code tool that allows designers and product teams to transform static designs into interactive prototypes using natural language prompts. Built into the Figma platform and powered by Claude 3.7 Sonnet, the tool helps users explore and refine ideas quickly without needing deep coding knowledge.

Key Features:

• Start with design, not from scratch: Use existing Figma frames and metadata as the base.

• Natural language editing: Prompt interactions like animations or behavior changes by simply describing them.

• Interactive prototyping: Add animations, dynamic data, and responsiveness across devices.

• Real-time collaboration: Multiple team members can work on the same file simultaneously, regardless of technical skill.

• Canvas-to-code workflow: Seamlessly integrates with Figma Design and Figma Sites for end-to-end product development.

Goal:

Figma Make aims to lower barriers in the design process, enabling faster validation, iteration, and collaboration—ultimately helping teams turn ideas into tangible products more efficiently.

How do you choose the right LLM?

The LLM space is booming — and picking the right model is now a business decision, not just a technical one. There’s no one-size-fits-all. The best model? It depends on your use case.

The chart below shows how models differ in capability (MMLU), size, and openness 

and these differences matter.

💡 Quick Breakdown:

1️⃣ Generalist vs. Specialist

• Generalist (GPT-4, Claude Opus, Gemini Pro): Great for broad tasks.

• Specialist (Mistral, IBM Granite): Lighter, faster, domain-tuned.

2️⃣ Big vs. Slim

• Big (Claude, GPT-4): More powerful, but expensive & slower.

• Slim (LLaMA 3, Mistral 7B): Cheaper, faster — ideal for edge or fine-tuning.

3️⃣ Open vs. Closed

• Open (Mistral, LLaMA 3): Full control, easy to customize.

• Closed (GPT-4, Claude): Top-tier performance, less flexible.

✅ Bottom Line:

There’s no "best" LLM — just the best one for your use case:

• 💻 Production apps? Go slim.

• 🧠 Complex reasoning? Go big.

• 🏭 Industry-specific? Go specialized.

LLMs aren’t a single solution, they’re a strategic stack. Choose wisely.

Link to the article