State of AI 2025: Stanford Reveals Who's Winning the AI Race
Stanford HAI just mapped the new AI power structure—if you’re building, investing, or scaling with AI in 2025, this is your playbook.
Good morning AI entrepreneurs and enthusiasts,
Today’s drop is packed: we’ve got the most comprehensive AI report of the year, Amazon outpacing OpenAI, and NVIDIA flexing with a model that beats unreleased LLaMA 4 variants.
Meanwhile, Mira Murati continues to poach top OpenAI talent as Thinking Machines shapes up to be the next major AI heavyweight. Five major stories—one big shift in the AI landscape.
In today’s AI news:
Stanford's 2025 AI Index Report: State of AI Revealed
Google unveils Agent2Agent standard for interoperable AI agents
Amazon's Nova Sonic & Reel 1.1 upgrades
Llama Nemotron Ultra 253B sets new benchmarks in reasoning & code
Murati recruits OpenAI elite
Top Tools & Quick News
Stanford's 2025 AI Index Report: State of AI Revealed
The News: The 2025 AI Index Report, released by Stanford's Institute for Human-Centered Artificial Intelligence, is the most comprehensive annual analysis of global AI trends. Spanning 456 pages, the report consolidates progress across research, regulation, business integration, and public sentiment, offering essential insight for any serious stakeholder in the AI ecosystem.
Key Takeaways:
Smaller, Stronger Models: Phi-3-mini (3.8B parameters) matched PaLM's 2022 performance with 142× fewer parameters.
Drastic Cost Reductions: GPT-3.5-class inference dropped from $20 to $0.07 per million tokens.
Global Landscape: The U.S. led model development, China dominated patents and papers, and open-source grew to 66% of all new models.
Enterprise Acceleration: AI adoption jumped from 55% to 78% as companies moved from experimentation to implementation.
Why it matters: This report is more than a snapshot—it’s the strategic compass for leaders, investors, and researchers navigating AI’s rapid evolution. From the rise of compact, high-performing models to the mainstreaming of AI in enterprise, the report offers both clarity and foresight. As regulation, environmental impact, and responsible deployment grow in urgency, the 2025 Index cements itself as a benchmark reference for where the industry is—and where it must go. It's not just another annual update—it’s the big story in AI today.
Google unveils Agent2Agent standard for interoperable AI agents
The News: At Google Cloud Next 2025, Google announced Agent2Agent (A2A), a new open protocol to enable cross-vendor collaboration between AI agents.
The Details:
A2A facilitates structured communication between client agents and remote agents.
Key features include Capability Discovery, structured Task Management, and dynamic User Experience Negotiation.
Agents publish JSON-based "Agent Cards" to share capabilities.
Promotes secure, HTTP/JSON-based interaction protocols.
Over 50 partners — including Accenture, Deloitte, and SAP — are contributing.
A new AI Agent Marketplace was also launched on Google Cloud.
Why it matters: The agent economy is booming, but fragmentation slows enterprise use. A2A provides a unified standard to connect tools from SAP, Salesforce, ServiceNow, and more — enabling seamless workflows like customer dispute resolution via Gmail, BigQuery, and ERP systems. A2A also complements existing frameworks like Anthropic's Model Context Protocol and LangChain’s Agent Protocol, offering a higher-level abstraction layer with open-source governance.
Amazon's Nova Sonic & Reel 1.1 upgrades
The News: Amazon launched Nova Sonic — a real-time voice model — and an updated Nova Reels 1.1 video generation tool via Amazon Bedrock, marking a major step forward in multimodal generative AI.
The Details:
Real-time performance: Nova Sonic delivers natural-sounding speech with 1.09s latency, surpassing OpenAI’s Realtime API (1.18s).
Accuracy: Achieves 4.2% word error rate across five languages and performs 46.7% better than GPT-4o in noisy, multi-speaker scenarios.
Natural interactions: Handles pauses, interruptions, and tone shifts smoothly, offering real-time transcript output for seamless API integration.
Extended video: Nova Reels 1.1 extends video output to 2 minutes with manual and automatic modes.
Affordability: Bedrock pricing places Nova Sonic at up to 80% lower cost than comparable OpenAI offerings.
Why it matters: Amazon’s upgrades reflect a push to become a full-stack AI provider, directly competing with OpenAI and Google. Nova Sonic’s immediate application in Alexa+ and Bedrock’s enterprise focus could rapidly scale adoption across industries.
Llama Nemotron Ultra 253B sets new benchmarks in reasoning & code
The News: NVIDIA has released the Llama Nemotron Ultra 253B, an open-source AI model engineered to outperform much larger models in key reasoning and coding benchmarks. Despite having fewer parameters, it is setting new standards for speed, precision, and efficiency.
The Details:
4× Inference Throughput: Designed for NVIDIA infrastructure, the model achieves exceptional speed and cost-efficiency, even compared to DeepSeek R1’s multi-rack system.
Dual Reasoning Modes: A unique ON/OFF toggle enables runtime optimization; in ON mode, the model hits 97% on MATH500 versus 80.4% OFF.
Benchmark Dominance:
Architecture: NAS-enabled design with vertical compression allows it to run on a single 8xH100 node with minimal latency.
Why it matters: Nemotron Ultra 253B’s performance fundamentally shifts the conversation on what matters in modern LLM design. Instead of scaling to trillions of parameters, NVIDIA demonstrates that smarter architecture and dynamic reasoning capabilities can unlock more value with far less compute. This sets a precedent for future high-performance, resource-efficient models tailored to enterprise AI, code generation, and scientific reasoning. And with its commercial open-source licensing, it’s already becoming a go-to model for developers looking to scale with precision.
Murati recruits OpenAI elite
The News: Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has added Bob McGrew (ex-Chief Research Officer) and Alec Radford (architect of GPT) to its advisory board—further strengthening its roster of AI heavyweights.
The Details:
19 of 38 founding team members are ex-OpenAI staff, including contributors to GPT, DALL·E, and Whisper.
John Schulman (OpenAI co-founder) serves as Chief Scientist, and Barret Zoph (ChatGPT co-creator) is the startup’s CTO.
Reportedly seeking $1B at a $9B valuation amid significant investor interest.
Focused on developing flexible, human-aligned AI through open-source tools and collaborative interfaces.
Why it matters: Murati’s startup is one of several high-profile OpenAI spinouts—including Safe Superintelligence and Anthropic—signaling a broader movement toward AI development focused on transparency and adaptability. The recruitment of McGrew and Radford reinforces the lab’s technical credibility and strategic intent. As OpenAI becomes more closed in its approach, Thinking Machines may set a new benchmark for open innovation. Murati’s ability to attract elite talent illustrates her growing influence in shaping the future of AI.
TODAY'S TOP TOOLS
🔍 Gemini Live: On-device visual AI for Android — Enables real-time object and text recognition through your camera or shared screen, now live on Galaxy S25 and Pixel 9 series.
🎥 Runway Gen-4 Turbo: Create 10s videos in 30s — Generates high-quality, 10-second videos in just 30 seconds, offering fast, cinematic results for creators.
🔎 AI Mode: Visual + multilingual search — Lets users snap photos, ask questions, and receive context-aware answers in multiple languages using Gemini’s AI.
🤝 ElevenLabs MCP: Deploy voice agents cross-platform — Integrates voice cloning and real-time speech synthesis into apps, powering voice-driven interfaces and assistants.
QUICK NEWS
OpenAI released its EU Economic Blueprint, calling for a €1B AI accelerator fund and foundational training for 100M Europeans by 2030.
Deep Cogito debuted Cogito v1: a self-improving, open-source model suite outperforming peers like Meta’s LLaMA 4 in reasoning tasks.
Google’s Gemini 2.5 Pro now features Deep Research tools + podcast-style audio summaries.
China’s Origin Wukong quantum computer fine-tuned a 1B-parameter model with 76% fewer parameters and improved training efficiency.
AI2 & Google Cloud invested $20M into the Cancer AI Alliance to scale personalized research.
Snapchat’s new Sponsored AI Lenses let brands build immersive, AI-driven ad campaigns with selfie integration.
More to come tomorrow — stay two steps ahead.
-AJ
One of the best articles I've read all year, so well written and easy to read.
456 pages?! So glad I can just read your newsletter instead