{
  "video": {
    "id": "UHbODjuviAU",
    "title": "The Investor's Edge Lives in Their Head — Anekha Sokhal, Moshi",
    "duration": 989,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "UHbODjuviAU",
    "title": "The Investor's Edge Lives in Their Head — Anekha Sokhal, Moshi",
    "one_liner": "If everyone has the same models, prompts, skills files and MCP connectors, an investor's edge is the judgment in their head, so AI should do the grind (gathering, cleaning, cross-checking and pushing data) and leave the decision to the human.",
    "summary": "Anekha Sokhal, a former derivatives trader who moved into ML and is now founder of Moshi, argues that the 20 W human brain is still the best reasoning engine available. She says AI should feed that brain rather than try to replace it. She defines edge as three things: knowing first, knowing your blind spots, and having enough conviction to size up a trade. Finance can't hand the decision loop to agents because someone has to be accountable for it. She presents Moshi as a 24/7 quant analyst built from six deterministic agents that track hallucinations and claim traceability, while the trader makes the decision.",
    "key_points": [
      "Data overload: the world produced about 175 zettabytes of data in 2025. More data doesn't mean more information, because everyone gets the same flood while attention stays flat, so the gap between attention and data keeps widening.",
      "The human brain runs on about 20 W (a dim light bulb) with 86 billion neurons, while systems imitating a sliver of it need a whole power plant. So the question is where in the decision cycle you point AI: do you replace the brain or feed it?",
      "Edge comes down to three things: knowing what's going on before anyone else, knowing your blind spots (second- and third-order effects), and having the conviction to size up on a trade you believe in.",
      "With equal access to models, skills files, MCP connectors and prompts, edge is 'what you can't download': decades of tacit knowledge in an investor's head, which leaves when they move firms or retire.",
      "The 'one person plus a few API calls replaces the quant/data team' story breaks at production: a weekend dashboard needs security, a database, storage, reliability and, above all, credits once one more person uses it. Pricing has moved from per-tool to per-API-call.",
      "There are two tool types in finance. Pull tools (prompting a model) make you the context and leave you arguing with a model frozen in time, for example about whether FOMC is today. Push tools flood you with data, which is a data-engineering problem.",
      "There is a hidden cost: juniors used to learn by sitting next to traders with 'one ear on the desk'. One-person pods break that pipeline. And finance needs a human who can stand up and say 'I made the decision for this reason', not an agent that just apologizes when it leaks an API key.",
      "Moshi's claims: six agents built deterministically watch markets 24/7 alongside the investor's own PDFs, emails and subscriptions; it has read 2.4B data points and learns about 300 new things in 4 seconds; it measures hallucination, checks every claim for traceability and shows its assumptions; it gives up to 35% more efficiency."
    ],
    "takeaways": [
      "Split the workflow into grind (reading, gathering, cleaning, cross-checking) and judgment. Automate the grind and keep the decision with an accountable human.",
      "Don't look for differentiation in models, prompts, skills files or MCP connectors that everyone can access. Capture the proprietary context in experts' heads before it walks out the door.",
      "Push relevant new data to users instead of making them pull from a frozen model. Pushed signals expose blind spots and prompt connections they wouldn't have made.",
      "Make every output traceable: measure hallucinations, verify each claim, and show the assumptions and reasoning path, because a confident wrong answer is worthless in finance.",
      "Before calling a weekend tool done, budget for productionization: security, database, storage, reliability and per-call API credit costs."
    ],
    "topics": [
      "finance",
      "agents",
      "human-in-the-loop",
      "trading",
      "data-engineering",
      "hallucination",
      "traceability",
      "tacit-knowledge"
    ],
    "tools": [
      "Moshi",
      "MCP",
      "X",
      "LinkedIn",
      "NASA",
      "ICLR"
    ],
    "quotes": [
      {
        "text": "So, the edge is actually what you can't download.",
        "at": "04:59",
        "url": "https://www.youtube.com/watch?v=UHbODjuviAU&t=299s"
      },
      {
        "text": "If an agent leaks your API key, all it says is, \"Oops, I'm sorry. I shouldn't have done that. Please rotate it.\"",
        "at": "09:06",
        "url": "https://www.youtube.com/watch?v=UHbODjuviAU&t=546s"
      },
      {
        "text": "Why don't we flip that and instead of replacing it, why don't we just feed it?",
        "at": "09:34",
        "url": "https://www.youtube.com/watch?v=UHbODjuviAU&t=574s"
      },
      {
        "text": "Because in finance, even if it's a really confident answer, if it's wrong, then it's worth nothing.",
        "at": "11:42",
        "url": "https://www.youtube.com/watch?v=UHbODjuviAU&t=702s"
      }
    ],
    "words": 3293
  },
  "summary_url": "/#UHbODjuviAU",
  "transcript": {
    "html": "/transcripts/UHbODjuviAU.html",
    "txt": "/transcripts/UHbODjuviAU.txt",
    "vtt": "/transcripts/UHbODjuviAU.vtt"
  }
}