{
  "video": {
    "id": "zQZiHOpkq_s",
    "title": "Research to Reality with Google DeepMind — Paige Bailey, Google DeepMind",
    "duration": 653,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "zQZiHOpkq_s",
    "title": "Research to Reality with Google DeepMind — Paige Bailey, Google DeepMind",
    "one_liner": "Google DeepMind's Gemma 4 open models (2B to 31B, Apache 2) beat models an order of magnitude larger, run on a single commodity GPU, and can run fully locally in a browser or on a phone. The QAT checkpoint for 2B is under 1 GB.",
    "summary": "Paige Bailey walks through Google's recent releases: a computer use API, managed agents, a speech-to-speech translation API and Gemini 3.5 Flash. She spends most of the talk on Gemma 4, the open model family. She demos a community-built app that runs Gemma 4 entirely in the browser with WebAssembly and transformers.js, and she shows the Google AI Edge Gallery app for on-device use. She argues that Gemma 4's performance for its size saves cost, because you don't need distributed inference or a large GPU footprint.",
    "key_points": [
      "Recent Google releases include a computer use API, managed agents and a speech-to-speech translation API. Managed agents take a higher-order task described in natural language and have a fleet of agents carry it out in a sandboxed Linux workstation, where you can add skills and dependencies.",
      "Gemini 3.5 Flash was released at I/O. Alongside its frontier models, Google is putting significant effort into open models.",
      "The Gemma 4 family comes in 2B, 4B, a newly released 12B, a 26B mixture-of-experts model and a 31B dense model. All are Apache 2 licensed, so you can download them, use them in your company and fine-tune them.",
      "Demo: a community project built with Gemma 4 and Fable 5 (used to rewrite kernels 'before it was taken off the market') runs Gemma in the browser through WebAssembly, sandboxed and fully local. It answered a Harry Potter emoji comparison table prompt almost instantly, and no data left the machine.",
      "The Google AI Edge Gallery app (Android and iOS) shows on-device capabilities: describing images, transcribing audio in multiple languages (Gemma supports over 140, she believes) and on-device function calling. It also includes skills such as building games, writing haikus, weather queries and calendar scheduling, all free and running on locally installed models.",
      "On devices like the Pixel 10 or other high-end phones, Edge Gallery can run inference on the local accelerator instead of falling back to the CPU.",
      "The 31B and 26B models perform above models an order of magnitude larger. They are meant to run on a single commodity GPU, some variants run on Jetson Nanos, 12B and below run on a laptop, and 2B fits on mobile.",
      "Quantized QAT checkpoints are available. The fast browser demo used them, and the 2B QAT checkpoint is under one gigabyte."
    ],
    "takeaways": [
      "For privacy-sensitive or offline use cases, consider running Gemma 4 locally in the browser with transformers.js and WebAssembly so no data is sent to an API.",
      "Use the QAT checkpoints when you need speed and a small footprint, for example on browsers or mobile.",
      "Try Google AI Edge Gallery to test on-device image, audio and function-calling capabilities, and use the hardware accelerator instead of the CPU on capable phones.",
      "Size your deployment to the model: 26B/31B on a single commodity GPU, 12B and below on a laptop, 2B on mobile. This avoids distributed inference and large GPU costs.",
      "Because Gemma 4 is Apache 2 licensed, you can fine-tune and ship it inside commercial products."
    ],
    "topics": [
      "open models",
      "on-device ai",
      "local inference",
      "quantization",
      "agents",
      "browser ml",
      "ai for science"
    ],
    "tools": [
      "Google DeepMind",
      "Google",
      "Gemma 4",
      "Gemini 3.5 Flash",
      "MedGemma",
      "AlphaFold",
      "Computer Use API",
      "Managed Agents",
      "Speech-to-speech translation API",
      "Fable 5",
      "WebAssembly",
      "transformers.js",
      "Google AI Edge Gallery",
      "Android",
      "iOS",
      "Pixel 10",
      "Jetson Nano"
    ],
    "quotes": [
      {
        "text": "I I hate slides, so we're going to see how few slides I can get through today.",
        "at": "05:49",
        "url": "https://www.youtube.com/watch?v=zQZiHOpkq_s&t=349s"
      },
      {
        "text": "since it is local um to the machine, none of your data is getting sent elsewhere. This is not using an API.",
        "at": "07:28",
        "url": "https://www.youtube.com/watch?v=zQZiHOpkq_s&t=448s"
      },
      {
        "text": "they're quite small in comparison to some other models that are on the market but they're performing way above what you might expect.",
        "at": "09:25",
        "url": "https://www.youtube.com/watch?v=zQZiHOpkq_s&t=565s"
      },
      {
        "text": "it's really intended to run super super well on a single commodity GPU",
        "at": "09:50",
        "url": "https://www.youtube.com/watch?v=zQZiHOpkq_s&t=590s"
      }
    ],
    "words": 1687
  },
  "summary_url": "/#zQZiHOpkq_s",
  "transcript": {
    "html": "/transcripts/zQZiHOpkq_s.html",
    "txt": "/transcripts/zQZiHOpkq_s.txt",
    "vtt": "/transcripts/zQZiHOpkq_s.vtt"
  }
}