{
  "video": {
    "id": "AYFcmnQrZCg",
    "title": "AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue",
    "duration": 845,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "AYFcmnQrZCg",
    "title": "AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue",
    "one_liner": "Treat a hedge fund as a function that compresses external and internal context into a (selection, size, timing) tuple, then break it into a graph of steps, choose the best runner for each (code, statistics/ML, LLM or person), and tune those choices against a clear North Star rather than automating for its own sake.",
    "summary": "Frank Long of Coatue argues that agents will have an even bigger impact on investing than on coding, and frames investing as a compression problem from all available information to what stock, how much and when. He breaks that function into gather, analyze and decide steps, then uses Korean-language news on semis and AI as a case study to show how much complexity sits inside one small part of gathering. Chunking, image handling, control flow, runner choice, model, thinking effort, harness, skills and hooks all act like hyperparameters, making the search space basically infinite. He says the great challenge is laying this out so that it can be understood technically and absorbed culturally and organizationally, and that the same state-machine problem applies in policy, nonprofits, drug discovery and corporate strategy.",
    "key_points": [
      "A hedge fund can be modeled as a function that takes external and internal context and returns a tuple of selection, size and timing; the product is a time series of these tuples that constructs a portfolio.",
      "Input philosophy: anything a human can see in pixels on a screen should be available to an agent via API, because even the greatest investors would be limited by a computer without logins to 75% of a firm's services.",
      "External context includes news, prices and filings; internal context includes current positions, liquidity and risk requirements, past case studies, and history with a sector, stock or management team.",
      "The foundation is both technological and cultural: the firm's data science, processes and systems, plus decades of filed notes on trades and on decisions made or not made, poised it to expose that insight to AI.",
      "Putting everything into one context window will not work, so the function is split into gather, analyze and form a view, and decide, with each step's output becoming the next step's input.",
      "The Korean news case study has steps to collect, filter, translate/extract/summarize, check against internal notes and meeting recordings, and potentially update views; collection alone could use an MCP, a CLI or licensed data feeds.",
      "There are four runner types: deterministic code (e.g. P/E below 15), statistics and classical ML (e.g. fitting a regression), LLMs and agents (e.g. summarization), and people (e.g. meeting a company's management team); most real steps combine several runners.",
      "Hyperparameters include chunking, image handling, control-flow branching, runner choice, model size, thinking effort, harness, skills and hooks; with at least 10 hyperparameters and at least 20 valid values per node, the search space across a massively complex graph is basically infinite."
    ],
    "takeaways": [
      "Expose everything your people can see to agents directly via API, including internal context such as positions, risk constraints, notes and past decisions, rather than relying on drag-and-drop or copy-and-paste.",
      "Decompose large decisions into steps with explicit inputs and outputs instead of placing all context into a single LLM prompt.",
      "Choose the runner for each step to maximize absolute quality, using deterministic code, statistics, LLMs, people or a combination; do not automate or use an LLM simply because you can.",
      "Treat chunking, control flow, model choice, thinking effort, harness, skills and hooks as tunable hyperparameters, and compartmentalize the system so each part can be understood and evaluated.",
      "Define a North Star to hill climb against before tuning, and plan for the organizational and cultural absorption of the system as well as the technical design."
    ],
    "topics": [
      "agents",
      "investing",
      "hedge-funds",
      "context-engineering",
      "workflow-orchestration",
      "human-in-the-loop",
      "evals",
      "hyperparameter-tuning"
    ],
    "tools": [
      "Coatue",
      "MCP",
      "Unitree"
    ],
    "quotes": [
      {
        "text": "Every single thing that a human can see in pixels on the screen, an agent in your building should be able to see via API.",
        "at": "01:39",
        "url": "https://www.youtube.com/watch?v=AYFcmnQrZCg&t=99s"
      },
      {
        "text": "There is no glory in automating things for the sake of automating things.",
        "at": "09:07",
        "url": "https://www.youtube.com/watch?v=AYFcmnQrZCg&t=547s"
      },
      {
        "text": "if you don't have a North Star to hill climb against, you're just kind of you're just fidgeting around.",
        "at": "12:16",
        "url": "https://www.youtube.com/watch?v=AYFcmnQrZCg&t=736s"
      }
    ],
    "words": 2728
  },
  "summary_url": "/#AYFcmnQrZCg",
  "transcript": {
    "html": "/transcripts/AYFcmnQrZCg.html",
    "txt": "/transcripts/AYFcmnQrZCg.txt",
    "vtt": "/transcripts/AYFcmnQrZCg.vtt"
  }
}