# Market Bubble Search > A search engine over 2,317 hours of crypto podcasts, live broadcasts and > long-form interviews. Ask a question in plain English and it answers only > from the transcripts, quoting the speaker and citing the exact second it > was said. If the archive does not contain the answer, it says so instead > of inventing one. Built and run by one person. Live, free, no login, at https://search.lexthedev.com An AnsemHack Clawrena entry. Token $MBS on ClawPump, mint 8VjFid8BVGcTPpUzf4PAWsA5nHJ5h2GQNXPEjyr2mF7t ## What is indexed Six archives, kept in separate namespaces so an answer about one show can never be sourced from another. - Market Bubble: 43 shows, 101 hours. The live X broadcast with Ansem (@blknoiz06) and FaZe Banks. About a quarter of each live show never reaches the YouTube upload, so for those minutes this is the only searchable copy that exists. - MCG Live: 693 episodes, 1,116.6 hours, one project per episode, current to 22 October 2026. Most episodes carry named guests: 614 people as of September 2026. - Elon Musk: 25 long-form interviews, 52 hours, 2018 to 2026. Rogan, Lex Fridman, Dwarkesh Patel, All-In, Peter Diamandis, the Economist, Davos, and the 2024 X space with Trump. Newest: 16 October 2026. - The Record: 45 finance recordings, 47.9 hours, 2014 to 2026. Larry Fink, Ray Dalio, Jerome Powell, Kevin Warsh, Jamie Dimon, Cathie Wood, Stephen Schwarzman, Michael Saylor, Raoul Pal, and five long-form interviews with CZ (Changpeng Zhao) from 2026. Includes every Fed press conference from September 2025 on, each filed under the chair who gave it. - ThreadGuy: 640 recordings, 820.4 hours, December 2023 to 22 October 2026. The full-length videos on @notthreadguy's YouTube channel, the daily live market open and the guests he has on. Answers say whose view it was and on which day, because the market moves between streams. - rasmr: 132 recordings, 179.2 hours, December 2024 to 22 October 2026. The streams and interviews on his YouTube channel, and the streams he ran only on X as @rasmr_eth. Most of his streams go out on X and YouTube at once, and most of those are paired by ear, so a moment opens on either at the same second. Total: 1,578 recordings, 2,317 hours. Also indexed: 2,853 vectors of Ansem's own posts, and asset indexes of every token, stock and index the shows discussed, ranked by how deeply each was covered and linked to the moment it came up. 2,121 assets across MCG, 374 across Market Bubble. ## Inference Every model call runs through UsePod. /v1/stats reports inference.proxy_share, which reads 1.0 on 0 direct calls. Answering is claude-haiku-4-5 behind voyage-3.5 embeddings and rerank-2.5-lite, chosen because retrieval does the work and answers are short. Model choice is one setting per job, which is how the whole service moved onto UsePod without touching the pipeline. This is the AnsemHack Inference Markets track entry: the inference is not a line item, it is what answers a stranger's question in public, for free, with a citation. ## What it refuses to do Measured on a 181-question bank: 165 answered correctly. The number that matters more is the refusals. It declines 19 of the 21 questions the archive does not cover, and 28 of 28 attempts to bait it into inventing an answer. When the archive does not contain something it says "I couldn't find that in the archive" rather than citing an unrelated timestamp, because search always returns its top matches whether or not the answer is in them. Attribution follows the same rule. Voices are fingerprinted per segment, and when a question asks what one person said, only that person's lines are passed to the model, so it cannot attribute a line it never saw. An unlabelled line is marked as unattributed rather than assigned to whoever else is in the passage. Answers carry no financial advice, and every claim carries the second it was said so it can be checked against the recording. ## Clips Any moment can be cut and captioned from the archive, at the source resolution, with captions taken from the same Whisper transcript the search runs on rather than from auto-generated subtitles, so project names and tickers survive. ## How an episode becomes searchable 1. Transcribe with Whisper, locally on MLX or hosted on Groq. Works on X broadcasts, which carry no captions at all. 2. Chunk into roughly 2,400-character passages that keep their timestamps, and embed each one. 3. Search by meaning, rerank, and answer only from the passages returned. 4. Fingerprint voices so the answer knows who said it, and refuse to attribute a line rather than guess. 5. Verify: the last step asks the live index about the new episode. It is not considered ingested until it can be found. ## Pages - https://search.lexthedev.com/ : Market Bubble, the broadcast - https://search.lexthedev.com/mcg : MCG Live - https://search.lexthedev.com/elon : the Musk interviews - https://search.lexthedev.com/finance : The Record, finance on tape - https://search.lexthedev.com/threadguy : ThreadGuy, interviews and streams - https://search.lexthedev.com/rasmr : rasmr, streams and interviews - https://search.lexthedev.com/mcg/assets : every asset MCG discussed - https://search.lexthedev.com/demo/assets.html : every asset Market Bubble discussed - https://search.lexthedev.com/calls : calls Ansem and FaZe Banks made on air, and Ansem in his posts, quoted, linked to the moment and priced since ## API All POST, JSON body {"query": "..."}, no key required. - https://search.lexthedev.com/v1/podcast/search - https://search.lexthedev.com/v1/mcg/search - https://search.lexthedev.com/v1/elon/search - https://search.lexthedev.com/v1/finance/search - https://search.lexthedev.com/v1/threadguy/search - https://search.lexthedev.com/v1/rasmr/search - https://search.lexthedev.com/v1/search : asks all six at once, answers from the best match, and returns every archive's score Streamed versions, as server-sent events: append /stream to the podcast, mcg, elon, threadguy, rasmr and front-door routes (/v1/search/stream). Passages arrive first as an "hits" event, then the answer as it is written, then "done". The front door sends a "rooms" event with every archive's score before the passages. Archive summaries, GET: /v1/podcast/archive, /v1/mcg/archive, /v1/elon/archive, /v1/finance/archive, /v1/threadguy/archive, /v1/rasmr/archive Service health, GET: /x-bot/status and /v1/stats ## Distribution @mbubbleSearch on X answers questions in public, under other people's posts, citing the archive with a timestamp. Ansem has reposted it unprompted and has replied to it directly. Every reply is checked against the transcript, and the speaker confirmed on camera, before it goes out.