Good morning. It’s Wednesday, July 16th.

On this day in tech history: In 2015, DeepMind’s AI crushed Atari 2600 games using deep reinforcement learning, nailing superhuman skills from raw pixels. This mid-July breakthrough showed machines could master complex tasks, pushing AI toward general intelligence and cementing deep learning’s game-changing potential.

^{In partnership with Mirai}

Mirai just dropped an on-device AI SDK for iOS & macOS — letting any solo dev run Llama-3, Qwen-3, Gemma-3, or DeepSeek-R1 locally. 

Want to bring your own model? That works too.

With zero latency, zero cloud costs and no privacy fuss.

No ML background, big team, or months of setup needed. 

With 8 lines of code, anyone can start building with AI.

Grab the SDK, the Swift engine, the model conversion toolkit — or just poke around the platform and docs.

^{Thank you for supporting our sponsors!}

More links, less fluff. 

We’re testing out a new format of the newsletter that provides an expanded list of AI news stories and discoveries from the past 48 hours, with fewer narratives. Let us know what you think of this simplified version by replying to this email!

• Meta scraps Llama 4 Behemoth as it pivots to prefab superclusters and closed AI

• TikTok owner, Bytedance, reportedly building its own XR glasses — rivaling Meta

• Amazon targets vibe-coding chaos with new 'Kiro' AI software development tool

• xAI updates Grok 4 alignment, launches gamified AI companions with NSFW unlocks

• NotebookLM becomes research engine as Google drops MRL‑powered embeddings and Big Sleep bug sniffer

• Runway’s Act‑Two Brings Full‑Body AI Mocap to Any Style, No Rigs Needed

• State of AI Trend Report for July 2025

• Pentagon Awards AI Giants $200M Contracts to Develop National Security Prototypes

• Anthropic drops Claude for Finance with data connectors and prompt kits

• Apple Might Know You're Pregnant Before You Do

• Remaining Windsurf team and tech acquired by Cognition, makers of Devin: 'We're friends with Anthropic again'

• Watch: Microsoft AI's Mustafa Suleyman says AIs are now beginning to have subjective experiences

• Perplexity offers free AI tools to students worldwide in partnership with SheerID

• Saudi Arabia plans to train 1 million citizens in AI

• AI agents’ capability horizons double every 7 months — new METR analysis finds similar rapid gains across science, math, robotics, and more

• Nvidia is set to resume China chip sales after months of regulatory whiplash

• Mira Murati's Thinking Machines Lab is worth $12B in seed round

• Mistral releases Voxtral, its first open source AI audio model

• OpenAI readies ChatGPT agents to take on Excel and PowerPoint

• Cornell–IBM collaboration advances quantum computing

• Meta fixes bug that could leak users' AI prompts and generated content

• Now Microsoft’s Copilot Vision AI can scan everything on your screen

• Research leaders urge tech industry to monitor AI's 'thoughts'

• Trump announces billions in investments to make Pennsylvania an AI hub

• Developers eye CUDA compatibility for Apple’s MLX machine learning framework

• US AI supercomputer Nexus will compute faster than 8 billion humans combined

• Hugging Face Is Hosting 5,000 Nonconsensual AI Models of Real People

📄 Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning

📄 SpeakerVid-5M: A Large-Scale High-Quality Dataset for Audio-Visual Dyadic Interactive Human Generation ​​

📄 Test-Time Scaling with Reflective Generative Model

📄 Vision-Language-Vision Auto-Encoder: Scalable Knowledge Distillation from Diffusion Models

📄 Scaling Laws for Optimal Data Mixtures 

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!