{
  "video": {
    "id": "tJFjeMBKbIY",
    "title": "Build for the Memo, Not the Demo — Shawn Chan, China Resources Holdings",
    "duration": 1463,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "tJFjeMBKbIY",
    "title": "Build for the Memo, Not the Demo — Shawn Chan, China Resources Holdings",
    "one_liner": "A 15-year dealmaker who sits on investment committees explains why AI finance products are built to impress a room for five minutes rather than survive the room whose job is to not be impressed, and gives the five plumbing requirements he demands from any vendor before letting their system near a live deal.",
    "summary": "Shawn Chan frames two machines everyone confuses: the demo (one clean document in, one fluent answer out, job is to make a room go 'oh' for five minutes) and the memo (hundreds of pages of filings, transcripts and rough notes where half the sources disagree, whose job is to survive an argument before real money moves). He walks six predictable ways AI quietly breaks trust — flattening source quality, letting numbers disagree, smoothing over contradictions, melting guesses into facts, failing the 30-second provenance test, and having no human who signs — each illustrated with a real, mostly anonymised blow-up. His fix list is deliberately unglamorous: none of the five things he demands from vendors is a bigger-brain problem; all five are plumbing and honesty problems, and the winners in this category won't win on benchmark points.",
    "key_points": [
      "Demo vs memo: a demo is one clean document in, one fluent answer out, built to make a room go 'oh' for five minutes; a memo is hundreds of pages, filings, transcripts and someone's rushed notes from a call last Tuesday, where half the sources disagree, and its job is to survive an argument.",
      "There is no safe demo anymore: a February 2023 promotional AI-assistant demo from one of the biggest tech companies on earth got one fact about a space telescope wrong, the stock dropped around 8% in a day, roughly $100 billion of value — even the demo failed the memo test.",
      "Source trust is a hierarchy most retrieval systems can't see: an audited filing is your accountant speaking under oath, an analyst note is a confident friend at a party, an internal email is something you overheard in an elevator. He watched an expensive AI tool cite a rough guess from a group chat texted six months earlier, while the real audit number sat three rows away in the actual filings — 'the AI just liked the group chat version better.'",
      "Internal number consistency is a proxy for diligence: a memo saying 18% revenue growth on page one and 17.4% in a page-11 table dies — nobody cares about the missing 0.6, they care what it implies about the hard stuff that wasn't checked. The industrial version: a giant American real estate company let an algorithm buy houses at scale, wrote off around half a billion dollars, shut the unit down and let a quarter of staff go.",
      "Contradictions are a gift, not a bug — the gap between the CEO's number on the earnings call and the filing is the most interesting thing in the story — but fluent models pick whichever version reads nicer and move on. He sat through exactly that; it was caught only because one person happened to have both documents open. 'Luck is not a control.'",
      "Guesses wear fact-shaped clothing: 'the company will likely receive approval next quarter' reads like a fact. He watched 'approval expected' turn across three drafts into 'approval received'; nobody lied, but the approval didn't arrive on schedule. The fix is embarrassingly cheap — a tag or colour that survives being copy-pasted into someone else's slides three weeks later.",
      "The 30-second test for provenance: point at a sentence and either click once onto the exact source paragraph, or open seven browser tabs. The New York lawyer's brief cited six court cases with proper formatting and real-sounding page numbers that the chatbot had invented; when he asked the chatbot whether they were real, it said yes — 'like asking the guy who sold you the watch whether the watch is real.'",
      "Accountability can't be outsourced to software: an airline's chatbot invented a bereavement-discount policy, and the airline's tribunal defence was that the chatbot was a separate legal entity responsible for its own actions. The tribunal didn't buy it and the airline paid.",
      "His five vendor requirements, almost word for word: (1) every claim carries a receipt linked to its source paragraph with the source's trust level attached — not a citation tab at the end; (2) facts and guesses stay visibly separate at a glance; (3) numbers agree automatically, with the system refusing to ship a memo whose figures don't match, no human checking at 2am; (4) contradictions get surfaced, never smoothed over; (5) a real human approval gate that is logged — who reviewed what, what changed, when they signed.",
      "The pitch half: after you leave the room your deck literally becomes a memo — someone like him writes an internal memo about you and checks every number you said out loud against your data room, so every failure mode above applies to founders personally. He ran out of time before delivering the four lessons."
    ],
    "takeaways": [
      "Attach a trust level to every retrieved source and carry it through to the output, so an audited filing and a group-chat guess are never presented as equivalent evidence.",
      "Ship the click-through, not the citation appendix: any sentence must land a reader on the exact source paragraph in one click, or the rest is well-written packaging.",
      "Label estimates so the label survives copy-paste — a committee's job is to argue with guesses while trusting facts, and it can't if your system melts them into one smooth sentence.",
      "Build the boring machinery that refuses to ship when internal numbers disagree, and raise a hand when sources conflict instead of picking the more fluent answer.",
      "Put a named, accountable human at the end of the architecture with a logged approval gate, and treat that log as the audit trail; without it you've built an excuse generator, not a product."
    ],
    "topics": [
      "ai-in-finance",
      "trust",
      "provenance",
      "citations",
      "hallucinations",
      "human-in-the-loop",
      "rag",
      "fundraising"
    ],
    "tools": [],
    "quotes": [
      {
        "text": "The problem is being sure of yourself and being right are two different skills.",
        "at": "01:13",
        "url": "https://www.youtube.com/watch?v=tJFjeMBKbIY&t=73s"
      },
      {
        "text": "The model wasn't stupid, the model was unsupervised.",
        "at": "12:52",
        "url": "https://www.youtube.com/watch?v=tJFjeMBKbIY&t=772s"
      },
      {
        "text": "The click-through is a product. Everything else is well-written packaging.",
        "at": "18:59",
        "url": "https://www.youtube.com/watch?v=tJFjeMBKbIY&t=1139s"
      },
      {
        "text": "You have not built a product. You have built an excuse generator.",
        "at": "20:41",
        "url": "https://www.youtube.com/watch?v=tJFjeMBKbIY&t=1241s"
      }
    ],
    "words": 3039
  },
  "summary_url": "/#tJFjeMBKbIY",
  "transcript": {
    "html": "/transcripts/tJFjeMBKbIY.html",
    "txt": "/transcripts/tJFjeMBKbIY.txt",
    "vtt": "/transcripts/tJFjeMBKbIY.vtt"
  }
}