{
  "video": {
    "id": "opHmX7qClzE",
    "title": "How OpenAI Built ChatGPT for Bankers and Analysts — Joseph Kim, OpenAI",
    "duration": 1139,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "opHmX7qClzE",
    "title": "How OpenAI Built ChatGPT for Bankers and Analysts — Joseph Kim, OpenAI",
    "one_liner": "OpenAI built ChatGPT for financial services mostly on the same public ChatGPT plugin platform (MCP tools, MCP apps, skills, artifact tools) that developers can use. Its lesson for builders is to design today for the 'threshold of usefulness' that models will cross tomorrow.",
    "summary": "Joseph Kim, who leads ChatGPT for financial services (launched a month before the talk with design partners Morgan Stanley and Evercore), says the main bottleneck is a 'capability overhang': model capability has outpaced the value professionals actually get. Closing that gap takes the right data, tools and a product experience that fits the work. He backs this with the Codex desktop app, which, not a new model, drove non-engineering adoption at OpenAI. He explains how they built the product: licensed and self-hosted financial data behind MCP, firm templates, citation viewers, and interactive charts, nearly all on the public plugin platform. He then argues that scaling, algorithms and 'unhobbling' will keep pushing finance tasks across usefulness thresholds, and makes three bets for 2027.",
    "key_points": [
      "The product targets investment banking and equity research first, built around three areas: financial data, an interface built for finance (inspect sources, explore charts, control sources), and the firm's own format via firm-specific templates.",
      "Existing MCP connectors for financial data had problems with latency, errors and search quality. OpenAI licensed the data directly and built its own hosting and search indices on the same infrastructure as its web index, still exposing the data to the model over MCP. Owning the data also lets them post-train models on these datasets.",
      "Analogy: MCP and skills are like handing a junior analyst an instruction manual; post-training is like a senior analyst who has worked with the data for two decades and knows its quirks.",
      "Build stack: MCP tools and connectors for data; citation metadata plus MCP apps for source inspection; MCP apps for interactive financial charts; skills plus shared artifact tools for Excel models and PowerPoint decks; and the native plugin onboarding experience.",
      "Inside OpenAI, finance, legal and recruiting went from almost none of their AI output in Codex at the start of 2026 to about 90% by June. The inflection around February–March came from the release of the Codex desktop app, not a new model.",
      "Model progress has three drivers: scaling compute, better algorithms (one study found the compute needed for a given performance level halves roughly every eight months), and 'unhobbling'. The data wall was addressed with synthetic data, learning from verifiable rewards and algorithmic gains; context length with sparse attention and compaction.",
      "The 'threshold phenomenon': capability improves gradually, but a task becomes useful suddenly, once review takes much less time than the work saves. Different tasks cross at different times.",
      "Finance roadmap: over the next 3 months, finance-specific post-training for financial reasoning, Excel models and PowerPoint (junior-analyst work). After that, senior judgment such as forecasting earnings and commodity prices, with the goal of matching the best human experts sometime in 2027. The next unhobbling, he says, is proactive, persistent agents."
    ],
    "takeaways": [
      "If you have unique data, an existing tool or deep knowledge of a workflow, package it as a ChatGPT plugin using MCP tools, MCP apps, skills and artifact tools, rather than building the whole stack yourself.",
      "If third-party MCP connectors are slow, error-prone or search poorly, consider hosting and indexing the data yourself while still exposing it through MCP. That gives you control over quality and reliability.",
      "Make verification part of the experience: citation viewers that show the source passages and tables behind every number, plus output in the firm's approved templates so less rework is needed.",
      "Ask which capabilities you undervalue today because they aren't quite useful yet. Design the interface, and what the user still controls, reviews or approves, for when those capabilities cross the threshold.",
      "Plan for persistent agents that hold a goal, remember context and work in the background. They work the way people delegate to colleagues, which makes AI accessible to users who haven't learned prompting."
    ],
    "topics": [
      "financial services",
      "mcp",
      "chatgpt plugins",
      "post-training",
      "capability overhang",
      "agents",
      "product design",
      "model scaling"
    ],
    "tools": [
      "ChatGPT for financial services",
      "ChatGPT",
      "Codex",
      "Codex desktop app",
      "MCP",
      "MCP apps",
      "ChatGPT plugin platform",
      "Excel",
      "PowerPoint",
      "Morgan Stanley",
      "Evercore",
      "OpenAI",
      "GPT-4",
      "GPT-5 Soul",
      "GPT-6 Astra",
      "RLHF"
    ],
    "quotes": [
      {
        "text": "We call that gap the capability overhang, the distance between what the model is capable of, and the value people can actually get from it.",
        "at": "04:44",
        "url": "https://www.youtube.com/watch?v=opHmX7qClzE&t=284s"
      },
      {
        "text": "I went back to see what happened around this time, assuming that we must have released a new model, but instead I found that we had released the Codex desktop app in February.",
        "at": "05:34",
        "url": "https://www.youtube.com/watch?v=opHmX7qClzE&t=334s"
      },
      {
        "text": "I personally believe that we're going to see multiple $1 billion businesses built simply as ChatGPT plugins.",
        "at": "09:07",
        "url": "https://www.youtube.com/watch?v=opHmX7qClzE&t=547s"
      },
      {
        "text": "What should you start building today to be ready for inevitable thresholds that the models will cross tomorrow?",
        "at": "18:46",
        "url": "https://www.youtube.com/watch?v=opHmX7qClzE&t=1126s"
      }
    ],
    "words": 3814
  },
  "summary_url": "/#opHmX7qClzE",
  "transcript": {
    "html": "/transcripts/opHmX7qClzE.html",
    "txt": "/transcripts/opHmX7qClzE.txt",
    "vtt": "/transcripts/opHmX7qClzE.vtt"
  }
}