{
  "video": {
    "id": "B0fjR3yaZFU",
    "title": "How do you diffuse AI into the real world? — Varun Shenoy, Long Lake",
    "duration": 1066,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "B0fjR3yaZFU",
    "title": "How do you diffuse AI into the real world? — Varun Shenoy, Long Lake",
    "one_liner": "Long Lake's co-founder argues AI diffusion — not model capability — is the bottleneck, and shows what it looks like to solve it by literally buying 35 services businesses and deploying agents inside them as the owner rather than the vendor.",
    "summary": "Varun Shenoy argues that the models are already capable — the demos are real — but nothing has changed inside a 200-person property management firm, and that's exactly what history predicts: electricity was demoed in the 1880s and Ford's electrified moving assembly line dates to 1924, because diffusion of a general-purpose technology takes a generation. Long Lake's answer is to stop selling software and instead acquire and operate the businesses themselves, so when the AI doesn't work it's their problem, not a customer's. He offers three lessons from 2.5 years of this: climb the autonomy ladder from co-pilot to co-worker rather than jumping to the top; harvest real-world traces from agents working alongside employees to build ground-truth evals and post-train on data that's out of distribution for frontier labs; and treat continual learning and enablement as one loop rather than two siloed teams.",
    "key_points": [
      "Diffusion, not capability, is the constraint: 'AI diffusion is perhaps the single most important problem for the next 20 years.' The Ford analogy — you must rip out the old motors, bring in new equipment, and retrain everybody — is the actual work.",
      "Long Lake has raised over $3 billion from Elad Gil, General Catalyst and AlphaWave in 2 years, acquired 35 businesses (HOA and property management, architecture, HR services), and announced a $6.3 billion take-private of American Express Global Business Travel. More than half the team is technology; the rest is finance and operations, with people from Palantir, Ramp, Glean, Blackstone and H.I.G.",
      "Owning the businesses inverts accountability: 'We're not the vendor. It's our problem.' They deploy into companies they own rather than selling from outside.",
      "The autonomy ladder: co-pilot (RAG chatbot) → synchronous agent (Claude Code, Codex, Claude co-work; runs 1–5 minutes, calls tools and skills) → asynchronous agent (can be triggered externally, e.g. off a job queue, not just by the user) → long-running agent (hours to months, what the labs are working on) → AI co-worker. 'You have to earn the right to do more' — you can't start at the top.",
      "The jagged frontier applied to form factors: for code, the async pattern is solved — wrap the coding agent in a sandbox, let it build and test, get a PR — and engineers are already comfortable parallelizing ('job seven might finish before job three'). Services work is traditionally serial (one email at a time), so the open question is what async and forking mean for property management or architecture.",
      "A bet on representing knowledge work as code: 'the models are trained on code, they want to write code' — rather than waiting for models to catch up on services knowledge work, use the coding ability by expressing the work as code.",
      "The data flywheel: agents collaborating with employees generate rich traces (tool calls, hiccups, papercuts), which become real-world evals with actual ground truth — did the roof get repaired, did the books get closed. Every week's hill-climbing benchmark becomes a regression test.",
      "Three upshots of traces: auto-built and auto-scored evals; explicit feedback (thumbs up/down, notes) plus implicit feedback (the diff between what the AI generated and what was ultimately submitted); and internal post-training on business data that is 'completely out of distribution for most frontier labs.'",
      "Continual learning and enablement are usually owned by separate silos (research/platform vs. growth/deployment) but are one snowball: the agent only improves if people use it, and people only use it if it's worth adopting. 'Everyone assumes the usage just shows up' — it doesn't."
    ],
    "takeaways": [
      "Don't ship the co-worker first. Place your product on the autonomy ladder honestly and climb it with the users in the field, both because model capability is jagged and because the organization has to come along.",
      "Instrument the collaboration to get evals for free: capture traces of agents doing real work with employees, define a real-world ground truth for each task, and promote each week's hill-climbing benchmark into a regression test.",
      "Mine implicit feedback, not just thumbs — the diff between AI-generated output and what the human actually submitted is signal almost nobody else has.",
      "Design form factors per industry rather than reusing the coding-agent pattern: figure out what parallelizing serial work means for that specific business, and embed natively where people already work (Excel, the ERP, 3D design software, Outlook/Gmail) to keep enablement energy low.",
      "Treat enablement as part of the learning loop and do it in person — 'extreme software service co-design.' Get on a plane, run the lunch and learn, sit two-on-one and watch them use it. 'You cannot co-design software with the services business over Zoom.'"
    ],
    "topics": [
      "ai diffusion",
      "agents",
      "evals",
      "enterprise deployment",
      "continual learning",
      "post-training",
      "autonomy",
      "enablement"
    ],
    "tools": [
      "Long Lake",
      "Claude Code",
      "Codex",
      "Claude co-work",
      "MCP",
      "Elad Gil",
      "General Catalyst",
      "AlphaWave",
      "American Express Global Business Travel",
      "Palantir",
      "Ramp",
      "Glean",
      "Blackstone",
      "H.I.G.",
      "Excel",
      "Outlook",
      "Gmail",
      "NVIDIA (Jensen)"
    ],
    "quotes": [
      {
        "text": "We own these businesses. So, when the AI doesn't work, it's not their problem. We're not the vendor. It's our problem.",
        "at": "04:16",
        "url": "https://www.youtube.com/watch?v=B0fjR3yaZFU&t=256s"
      },
      {
        "text": "You have to earn the right to do more.",
        "at": "07:07",
        "url": "https://www.youtube.com/watch?v=B0fjR3yaZFU&t=427s"
      },
      {
        "text": "There are hills and ravines. There's death by a thousand paper cuts. But that's what real work looks like. That's the entire job. The exceptions are the job.",
        "at": "13:24",
        "url": "https://www.youtube.com/watch?v=B0fjR3yaZFU&t=804s"
      },
      {
        "text": "You cannot co-design software with the services business over Zoom or over a support ticket. You have to be there. You have to be in person. And I'd argue this is the part that actually makes it work. In order to get AI diffusion to work, you have to touch some grass.",
        "at": "17:04",
        "url": "https://www.youtube.com/watch?v=B0fjR3yaZFU&t=1024s"
      }
    ],
    "words": 3496
  },
  "summary_url": "/#B0fjR3yaZFU",
  "transcript": {
    "html": "/transcripts/B0fjR3yaZFU.html",
    "txt": "/transcripts/B0fjR3yaZFU.txt",
    "vtt": "/transcripts/B0fjR3yaZFU.vtt"
  }
}