{
  "video": {
    "id": "ztlXKfPCT3Q",
    "title": "You’re Not Thinking Big Enough: Rebuilding Food Systems with AI Agents — Cody Menefee, Firecrawl",
    "duration": 1106,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "ztlXKfPCT3Q",
    "title": "You’re Not Thinking Big Enough: Rebuilding Food Systems with AI Agents — Cody Menefee, Firecrawl",
    "one_liner": "Firecrawl's Cody Menefee argues that AI engineers should use LLMs on real-world multivariate problems like rotational grazing, where an LLM in the loop could decide daily where GPS-collared cattle should move. Labor, not land, is what keeps grass-fed farming from scaling.",
    "summary": "Menefee says too many engineers build 'software for people who build software'. He uses pasture rotation for grass-fed livestock as an example of a bigger problem worth solving. 97% of cows are finished on feedlots because rotational grazing takes daily labor: moving fences, animals and water, guided by a farmer's own judgement of the grass. He proposes combining grass sensing (drones, satellites, trail cams), a knowledge base scraped from farmer YouTube videos and research papers (his open-source Open Pasture project, built with Firecrawl) and open virtual-fence collars. An LLM would then suggest the next best paddock, and a human would confirm or deny it.",
    "key_points": [
      "97% of cows are currently finished on feedlots and 3% are raised on pasture. Labor is the bottleneck for grass-fed systems.",
      "Rotational grazing splits pasture into paddocks holding one day of food, and animals move every day. That means moving fences, animals and water daily, plus keeping grass in a 'juvenile sweet spot': not grazed too short, not left to grow old and bitter.",
      "Halter (Peter Thiel invested at a $2B valuation) and No Fence sell GPS collars with virtual boundaries. But farmers still have to know where to move the animals, which today means walking the pasture and judging the grass by eye.",
      "Options for sensing grass: drone orthomosaic maps give high fidelity but need farmer training, and no jurisdiction has approved autonomous drones for this. Planet satellites image the whole globe daily at 1x1 m resolution (his favourite). A trail cam next to a measuring reference (a Lincoln University, Missouri research idea) tracks grass height.",
      "The LLM's decision is multivariate: GPS locations today, yesterday and tomorrow, drought conditions and grass height across the farm. It has to weigh the individual cow, the herd, the pasture, the farm and the wider ecosystem.",
      "Three blockers: (1) a knowledge base, since grazing know-how is locked in farmer YouTube videos; he scrapes those and arXiv papers with Firecrawl into Open Pasture. (2) A visualization layer for biomass and biodiversity. (3) Closed geofence collars that won't accept third-party software.",
      "Call to action: build an open-patent collar with open APIs, ideally from off-the-shelf parts, in the right-to-repair spirit (farmers resent John Deere), so people can compete on software.",
      "Species stacking: Pasture Bird put a chicken house on wheels that moves its own width every 24 hours. Running ruminants first, then chickens that peck parasites out of droppings, lowers parasite load and medication costs."
    ],
    "takeaways": [
      "Look for big, physical, real-world problems instead of building another MCP for another SaaS product.",
      "Use LLMs where there is no deterministic answer: put an LLM at the centre of a multi-input system to reason over the data and suggest actions that a human confirms or denies.",
      "Treat context as the main problem: gather, package and present domain data (including expert knowledge buried in YouTube videos and papers) in a form an LLM can reason over.",
      "Prefer open hardware and APIs so the software layer can be improved independently of vendors' closed ecosystems."
    ],
    "topics": [
      "agents",
      "agriculture",
      "multivariate decision-making",
      "knowledge base",
      "web scraping",
      "open hardware",
      "human-in-the-loop"
    ],
    "tools": [
      "Firecrawl",
      "Open Pasture",
      "Halter",
      "No Fence",
      "Planet",
      "Pasture Bird",
      "John Deere",
      "YouTube",
      "arXiv",
      "Tesla"
    ],
    "quotes": [
      {
        "text": "I am the vibe coder most of you all are scared of.",
        "at": "01:57",
        "url": "https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=117s"
      },
      {
        "text": "there are other problems to solve than just another MCP for another SaaS solution at another company.",
        "at": "04:19",
        "url": "https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=259s"
      },
      {
        "text": "I need someone to make me a collar.",
        "at": "13:58",
        "url": "https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=838s"
      },
      {
        "text": "there's a lot of problems that you can't solve with deterministic uh algorithms.",
        "at": "16:28",
        "url": "https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=988s"
      }
    ],
    "words": 4199
  },
  "summary_url": "/#ztlXKfPCT3Q",
  "transcript": {
    "html": "/transcripts/ztlXKfPCT3Q.html",
    "txt": "/transcripts/ztlXKfPCT3Q.txt",
    "vtt": "/transcripts/ztlXKfPCT3Q.vtt"
  }
}