{
  "video": {
    "id": "iJVxxxHM_Oc",
    "title": "Where RL Will Take Search — Maximilian-David Rumpf, SID.ai",
    "duration": 576,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "iJVxxxHM_Oc",
    "title": "Where RL Will Take Search — Maximilian-David Rumpf, SID.ai",
    "one_liner": "SID.ai's founder argues that agentic search — currently 100–1000x more expensive and minutes rather than milliseconds slower than classical search — can be collapsed into a single RL-trained search sub-agent that is ~20x faster and ~100x cheaper than a frontier model doing the same job.",
    "summary": "Rumpf frames agents as a new search paradigm that roughly doubles the chance of finding the right documents but costs 100–1000x a classical query and burns 30–50% of an agent's tokens. His answer is to hand searching to a specialised sub-agent trained end-to-end with reinforcement learning, because search is a verifiable, grindable reward: for a given question, did you find the correct document? He places this in a lineage — edge detection → bounding-box models → VLMs, Deep Blue → Stockfish → AlphaZero, BM25/PageRank → vectors/rerankers → pure RL — where machine design beats human design and no decisions are baked in at design time. Their model SID 1 (with parallel execution) cuts a ~2 minute frontier-model search to ~5 seconds on average at about 1/100th the cost, with quality scaling predictably with compute and no ceiling yet visible.",
    "key_points": [
      "Agentic search is about twice as likely to find the right documents as classical search, but costs roughly 100–1000x a classical query and takes minutes, not milliseconds.",
      "Agents spend 30–50% of their tokens on searching, usually at the start of a task while assembling context.",
      "A classical pipeline (LLM query rewrite → search backend → reranker) is a chain of locally optimised models with all decisions baked in at design time and a fixed compute budget per question; the reranker may know results are insufficient but can't take action, only return bad results.",
      "That rigidity accrues a long tail of failures on unexpected questions, which teams patch with endless hand-written edge cases.",
      "Search is an ideal RL target because rewards are verifiable (did you find the correct document?) and grindable — their environment lets the model attempt questions thousands of times per second during a training run.",
      "The RL model goes back and forth with the database: search, read, iterate, set metadata filters on the fly, constrain its search, spend more compute on a hard question, and finally emit a ranked list — with strategies discovered rather than specified, as in AlphaZero.",
      "Results: SID 1 with parallel execution is ~20x faster (about 5 seconds average vs ~2 minutes) and ~100x cheaper than a frontier model, benchmarked across legal, finance, knowledge bases, science and email using both academic and internal benchmarks; still not quite vector+reranker latency, but they expect to get there.",
      "In production the search sub-agent keeps bad retrievals out of the main agent's context window — the main agent only ever sees great results, so it is more likely to be correct and 30–50% of its tokens move to a 100x cheaper model.",
      "Quality rises predictably with training compute, other rewards like latency can be mixed in, and they see no ceiling yet."
    ],
    "takeaways": [
      "Move retrieval out of your main agent and into a dedicated search sub-agent, so failed and irrelevant results never pollute the main agent's context window.",
      "Stop patching retrieval pipelines with more hand-written edge cases; the structural problem is that the components can't act on their own knowledge that results are bad.",
      "If you want to train search with RL, build the environment first: a verifiable reward (did it find the correct document) and enough throughput to attempt questions thousands of times per second.",
      "Don't prescribe the search strategy — let the model discover query rewriting, filtering and iteration itself, and mix in secondary rewards such as latency.",
      "Budget for adaptive rather than fixed per-question compute: a specialised small model can spend more effort on hard questions and still beat a frontier model on both cost and speed."
    ],
    "topics": [
      "reinforcement-learning",
      "search",
      "retrieval",
      "agents",
      "rag",
      "rerankers",
      "benchmarks",
      "inference-cost"
    ],
    "tools": [
      "SID.ai",
      "SID 1",
      "BM25",
      "PageRank",
      "IBM Deep Blue",
      "Stockfish",
      "AlphaZero",
      "MuZero",
      "JP Morgan"
    ],
    "quotes": [
      {
        "text": "The re-ranker might know that the results are insufficient at answering the question, but the re-ranker can't take action. It can only essentially return the results even when they're bad.",
        "at": "02:52",
        "url": "https://www.youtube.com/watch?v=iJVxxxHM_Oc&t=172s"
      },
      {
        "text": "machine design outperforms human design",
        "at": "03:31",
        "url": "https://www.youtube.com/watch?v=iJVxxxHM_Oc&t=211s"
      },
      {
        "text": "we don't know whether this method has no ceiling, but we're definitely not yet seeing a ceiling to this approach",
        "at": "06:46",
        "url": "https://www.youtube.com/watch?v=iJVxxxHM_Oc&t=406s"
      },
      {
        "text": "the most valuable information is not on the internet. For example, how to run JP Morgan is nowhere on the web but it is deep inside of the databases at at JP Morgan.",
        "at": "09:04",
        "url": "https://www.youtube.com/watch?v=iJVxxxHM_Oc&t=544s"
      }
    ],
    "words": 1735
  },
  "summary_url": "/#iJVxxxHM_Oc",
  "transcript": {
    "html": "/transcripts/iJVxxxHM_Oc.html",
    "txt": "/transcripts/iJVxxxHM_Oc.txt",
    "vtt": "/transcripts/iJVxxxHM_Oc.vtt"
  }
}