<- all tokdocs
2026.10.22@angusthenontechnical

The video promises an AI that maps which phrases make viewers leave; the tool shown on screen reports "No clear underperforming passage"

AI
Watch on TikTok

View on TikTok ->

The artifact is a 69.600 s vertical clip at 1080x1920 (9:16), HEVC Main in an hvc1 tag inside a mov,mp4,m4a,3gp,3g2,mj2 container, yuv420p, bt709 primaries and transfer, 30/1 fps across 2,088 frames, video bitrate 702,714 bps, audio AAC HE-AACv2 stereo at 44,100 Hz and 48,135 bps with a stream duration of 69.588 s, total size 6,591,404 bytes at an overall 757,632 bps, format id bytevc1_1080p_757632-1, carrying the container tags aigc_info={"aigc_label_type":0} and comment=vid:v12044gd0000dbd2j4vog65hqupd9dt0. It was posted at epoch 1792682396, which is 2026-10-22T15:19:56Z, by angusthenontechnical (uploader id 6783510017912865797, nickname "Angus the Nontechnical"), credited to "original sound" by Angus the Nontechnical, with a 58-character description identical to the title. At the capture time recorded by the metadata file (2026-10-22 18:01:37 local) it showed 6,729 views, 426 likes, 3 comments, 99 reposts and 420 saves. I extracted 35 frames at 2 s intervals into tokdoc_frames/ and read frames 001, 003, 004, 005, 007, 009, 010, 011, 013, 014, 015, 016, 018, 020, 023, 024, 025, 027, 029 and 033; because the speaker's head occludes the two prompt cards, I re-extracted 24-30 s and 44-48 s at fps=2 into tokdoc_frames_1s/ and read pm_003, pm_006, pm_007, pm_009 and cls_002. The transcript is 276 words in 32 Whisper segments. Frame 001 opens on a title card reading "Alex Hormozi is a $4M API call" over a podcast clip of a bearded man in an ACQUISITION.COM cap; frame 003 shows a Google results page for the query "Where does Alex Hormozi allocate his $4 million h…" with a LinkedIn result headlined "My $4000000 content strategy for 2025 | Alex Hormozi | 191 comments"; frame 004 shows an onlyhappynews.com article dated JULY 2, 2026 whose body reads "in 40 months, he spent $4 million on team,"; frame 005 shows a black card reading "$4M CONTENT INVESTMENT / Across roughly 40 months / THE COSTS HE NAMED: TEAM People, EQUIPMENT Tools, VENDORS Services" with the footnote "Conceptual emphasis, not a budget split. Testing is the largest visual area; no category dollar amounts or shares were disclosed." and the attribution "Source: The Game, Ep. 750 · September 2024"; frame 009 shows a five-node workflow labelled "Post tests", "Saved browser login", "Capture Studio ana[lytics]", "Latest metrics" and "Snapshot [hi]story", whose code cards read chromium.launchPersistentContext(USER_DATA_DIR, { headless: HEADLESS, viewport: { width: 1400, height: 950 }, args: ['--disable-blink-features=AutomationControlled'] }) and page.goto('https://www.tiktok.com/tiktokstudio/analytics/${itemId}', { waitUntil: 'domcontentloaded' }) against a JSON field named video_retention_rate_realtime; frames 010 and 011 show a report UI reading "20 prior comparable videos Sep 22, 2026 – Oct 8, 2026", "Early data · descriptive comparison", "HOOK · 3S 71.0% retained Typical total: 63.0% · 20 videos", "LARGEST RAW PASSAGE LOSS 22.0 pts · 0–2s" and "WORST RELIABLE PASSAGE VERSUS USUAL No clear underperforming passage"; frames 013 and pm_006 show per-passage cards such as "5–7s SPOKEN 57.0% remain 'definitely an industry plant?' Raw loss 9.0 pts Usual loss 7.4 pts 20 comparisons"; frame pm_007 shows a prompt card headed "VIDEO POST-MORTEM" beginning "Use video-post-mortem to compare TikTok videos 8956 and 8957."; frame pm_009 shows a Claude Code thread titled "Bobbi Althoff vs pirate video comparison" ending "I didn't run a new audit, and nothing was written anywhere."; frame 018 shows an iPhone on Instagram's "Trial reels" screen with a purple "Create trial reel" button; frame 020 shows the Instagram profile @angus.sewell with 274 posts, 194K followers and a dashboard line reading "9.1M views in the last 30 days"; frame 024 shows a second prompt card headed "CONTENT LEARNING SESSION" whose step 5 reads "Put the finished report into NotebookLM and create an audi[o] overview that coaches m[e on my] next tests."; frame 025 shows a dashboard reading "TikTok videos / 458 videos on record" with tiles for 458 videos, 13M views, 1.4% save rate and 0.4% share rate; and frame 027 shows a NotebookLM notebook named "Content learni[ng]" with Sources, Chat and Studio tabs and a generated answer headed "Why the designer, McDonald's and Kalshi videos did what they did". The spoken argument is that Alex Hormozi's content machine reduces to a prediction loop you can rebuild yourself: post on TikTok, scrape your retention analytics, have an AI compare two tests and name the phrases that cost you viewers, post two trial reels, promote the winner, then have the AI mine your whole archive into a lessons table and coach you on it. The last transcript segment ends at 69.58 s against an audio stream of 69.588 s, so nothing is timestamped past the end of the audio and no segment is a Whisper hallucination.

The $4 million is real and traceable; "the past four years" is 40 months

At 6.80 s the video says "Four milli in the past four years." The dollar figure survives checking. The episode the video itself names on screen in frame 005 is "The Game, Ep. 750 · September 2024", and Apple Podcasts lists that episode as "Part 1 - How I Gained 7.8 Million Followers In 40 Months | Ep 750" on the podcast "The Game with Alex Hormozi", released September 30, 2024. A transcript of that episode records the line "$4 million in team, equipment, vendors, studios, and software, 1,000 hours of me on-camera recording, and 35,000 pieces of content," framed as the cost of "40-ish months." I read that quote in a third-party transcription at podscripts.co rather than in a transcript Hormozi published himself, so treat the wording as close rather than exact; the episode title and release date are confirmed against the Apple Podcasts listing.

The time window is where the voiceover slips. Forty months is three years and four months, not four years. The video's own graphics get this right: frame 005 reads "Across roughly 40 months," and the onlyhappynews.com article visible in frame 004 reads "in 40 months, he spent $4 million on team," so the correction is already on screen while the audio says something else. The difference matters for the argument the video is making, because stretching 40 months to four years lowers the implied annual spend by about 17 percent and makes the system look slightly cheaper to copy than the source claims.

One more number in the same frame does not match. The Google panel in frame 003 shows a LinkedIn result headlined "My $4000000 content strategy for 2025" whose snippet begins "$4 million, that's how much I spent creating [...] last year," attributing the same figure to a single year rather than to 40 months. I could not fetch the LinkedIn post, so I am not resolving which framing Hormozi intended. Two different time windows for the same figure appear in a single frame of this video.

"Mostly spent on the testing side of things" is the video's inference, and the video's own footnote says so

At 8.74 s the video says "Mostly spent on the testing side of things." This is the load-bearing claim for everything that follows, because it is what licenses the conclusion at 10.78 s that "you can vibe code his whole system." If most of the $4 million bought testing infrastructure, then rebuilding the testing loop captures most of the value.

The source does not support it. The categories Hormozi names are team, equipment, vendors, studios and software. Frame 005 reproduces four of them as cards reading "TEAM People", "EQUIPMENT Tools", "VENDORS Services" and a fourth partially occluded card ending in "ARE". None of them is testing. The video then prints its own disclaimer in grey at the bottom of the same card: "Conceptual emphasis, not a budget split. Testing is the largest visual area; no category dollar amounts or shares were disclosed." That sentence concedes the point. The layout makes testing look dominant because the designer made it the largest element, and no share of the $4 million was ever attributed to it.

So the honest version is that Hormozi spent $4 million over roughly 40 months across five named cost categories and has not published a split. The claim that most of it went to testing is unverified, and the video knows it.

The tool shown on screen declines to make the claim the voiceover makes

At 23.60 s the video says the AI can "literally map out which phrases cause the most people to leave." The report it cuts to in frames 010 and 011 says the opposite in its own words. Under the heading "WORST RELIABLE PASSAGE VERSUS USUAL" it prints "No clear underperforming passage" and, below that, "Early data cannot identify a reliable worst passage." The summary badge above the curve reads "Early data · descriptive comparison," and the explanatory line reads "This curve was measured before day 7, below 1,000 views, or without measurement details. Passage results stay neutral."

The per-passage cards in frame 013 and pm_006 show why. Each passage carries a raw loss and a usual loss drawn from 20 comparisons. For 4-5 s the card reads "Raw loss 3.0 pts Usual loss 6.5 pts," meaning that passage lost less than the archive baseline. For 5-7 s it reads "Raw loss 9.0 pts Usual loss 7.4 pts," a gap of 1.6 points. For 7-9 s it reads "Raw loss 4.0 pts Usual loss 4.9 pts." Those differences are small enough that the tool refuses to name a culprit. The single largest raw drop is logged as "LARGEST RAW PASSAGE LOSS 22.0 pts · 0–2s," the opening two seconds, where every short-form video sheds viewers who swiped away before the content started.

The Claude Code thread in pm_009 is more explicit still. It proposes an edit, then qualifies it: "The cut is a hypothesis, not [...] 25% to 22% across 27-3[1s ... how] big that drop looks depends on rounding. The archive tests also didn't find a cause. Only a re-cut tested against the original would show whether this edit [helps]." It closes with "I didn't run a new audit, and nothing was written anywhere." The terminal card in frame 016 repeats the warning: "Measure the revised cut before promoting it. The observed loss is rounding-sensitive. No new global lesson was written."

This is the real finding in the video, and the voiceover talks over it. The artifact being demonstrated is a calibrated instrument that reports uncertainty and withholds causal claims on thin data. The narration sells it as a phrase-level drop-off detector. Both appear in the same six seconds of screen time.

Trial reels is an Instagram feature, and the video changes platform without saying so

At 15.32 s the video says "You just need to start posting on TikTok of all places." At 34.42 s it says "Which then you can post two versions as trial reels," and at 37.10 s, "whichever version wins goes to the main page." The voiceover never mentions a second platform.

The screen does. Frame 018 is an Instagram screen, titled "Trial reels," with the eye-shaped view icon Instagram uses and a purple "Create trial reel" button at the bottom. Frame 020, four seconds later, is the Instagram profile @angus.sewell showing 274 posts and 194K followers. The TikTok screens earlier in the video (frames 007 and 025) use TikTok's play-triangle view icon and the TikTok Studio chrome. The platform switch is visible and unspoken.

Instagram's own creator blog describes the feature as "a new way to try out content and see what performs best, by giving you the option to share reels with people who don't follow you," and states that "Your trial reel will not appear in your profile's reels tab or your main grid unless you later decide to Share to everyone." It lists the returned metrics as "views, likes, comments and shares" plus comparison against previous trials, and describes an opt-in where Instagram will "automatically share your trial reel with followers if we determine it's performing well based on the views it receives within the first 72 hours." That matches the video's "whichever version wins goes to the main page" exactly, and it is an Instagram mechanism.

The split testing TikTok documents lives in Ads Manager: its help article instructs you to "Use the split test toggle in the TikTok Ads Manager when creating a campaign to test two different versions of your ads," and to "Set the test schedule dates to at least 7 days so that the system has enough time to collect a useful sample." That is a paid campaign tool, not a posting option. I could not load support.tiktok.com or TikTok's Creator Academy article pages, both of which returned only navigation markup, so I am not asserting what TikTok's own help pages say about organic testing. What the video shows is sufficient: the trial-reel step runs on Instagram.

The retention curve comes from a TikTok Studio scrape, not an API

At 17.72 s the video says to "capture your analytics through this workflow right here." The workflow in frame 009 is specific about how. The first code card launches a browser with a persisted profile: chromium.launchPersistentContext(USER_DATA_DIR, { headless: HEADLESS, viewport: { width: 1400, height: 950 }, args: ['--disable-blink-features=AutomationControlled'] }). The accompanying label reads "Saved browser login." The second card navigates to https://www.tiktok.com/tiktokstudio/analytics/${itemId} with waitUntil: 'domcontentloaded', intercepts responses, and keeps any payload where j.video_retention_rate_realtime is present. It also runs detectBlockSignal({ status: navigation?.status() ?? 200, url: page.url() }) and assigns the result to blockedReason.

Three details give the shape of the surface away. The --disable-blink-features=AutomationControlled flag exists to hide the automation fingerprint from the page. The persisted user data directory exists because the endpoint requires a logged-in session. The block detector exists because the operator expects to be blocked. None of those is something you write against a documented API with a key. The per-second retention series is real, it sits behind the TikTok Studio web app, and this workflow reaches it by driving a logged-in browser.

That bounds what the tool can claim. The field is video_retention_rate_realtime, a retention rate over time. Mapping that curve onto spoken phrases is work the creator's tool does by joining transcript timestamps against the curve, which is why the report prints bands like "4–5s", "5–7s" and "7–9s" with the transcribed line alongside. TikTok supplies the curve. The phrase attribution is the creator's own inference layer, and that layer is the one reporting "No clear underperforming passage."

NotebookLM is the product name, and it can take the lessons table

At 53.00 s the transcript reads "have it coach you via probably something like notebook.lm." That spelling is a Whisper artifact rather than a video error. The prompt card in frame 024 writes the product correctly in its fifth step: "Put the finished report into NotebookLM and create an audi[o] overview that coaches m[e on my] next tests." Frame 027 shows the product itself, with the Sources, Chat and Studio tabs, a PRO badge, and a notebook named "Content learni[ng]".

Google introduced NotebookLM on July 12, 2023, describing it as "an experimental offering from Google Labs" and "an AI-first notebook, grounded in your own documents, designed to help you gain insights faster." At launch it accepted Google Docs only.

The ingestion question matters because the video's step is to feed it a table. That works now. Google's November 13, 2025 announcement added Google Sheets with the stated purpose "Add structured data and ask information on key statistics or generate summaries from your spreadsheets," alongside Drive URLs, images, Drive-hosted PDFs and .docx. The current NotebookLM help page lists CSV among supported source types together with Google Docs, Slides, Sheets ("at this time, files are limited to 100k tokens"), PDF, TXT, Markdown, DOCX, PPTX, ePub, web URLs, YouTube URLs for public videos with captions, images and audio, with limits of "up to 500,000 words or up to 200MB for uploaded files" per source and "up to 50 sources (for Free users)". The lessons table shown in frame 023, five rows of saved findings with a confidence label and a checked date, is far inside those bounds.

The name of the person at the centre of the video survives the same check. Acquisition.com's own site identifies the company as "Acquisition.com, LLC" and names "Alex and Leila Hormozi" as co-founders and managing partners, so the correct spelling is Hormozi. The transcript's "Hormozy" at 6.68 s is another Whisper artifact; the video's own title card in frame 001 reads "Alex Hormozi is a $4M API call."

Key Takeaways

  • Verified: The $4 million figure is real. The Game episode 750, "Part 1 - How I Gained 7.8 Million Followers In 40 Months," released September 30, 2024, records "$4 million in team, equipment, vendors, studios, and software, 1,000 hours of me on-camera recording, and 35,000 pieces of content," read via a third-party transcription.
  • Correction: The video says "Four milli in the past four years" at 6.80 s. The source window is roughly 40 months, which is three years and four months. The video's own card in frame 005 reads "Across roughly 40 months."
  • Unverified: "Mostly spent on the testing side of things" at 8.74 s has no source. Hormozi named five cost categories and disclosed no split. The video's own footnote in frame 005 concedes this: "Conceptual emphasis, not a budget split. Testing is the largest visual area; no category dollar amounts or shares were disclosed."
  • Correction: The video says at 23.60 s that the AI will "literally map out which phrases cause the most people to leave." The report it shows at 18 s and 20 s prints "No clear underperforming passage" and "Early data cannot identify a reliable worst passage," and the Claude thread at 28 s adds "The archive tests also didn't find a cause."
  • Correction: Trial reels is an Instagram feature, not a TikTok one. The video tells you to post on TikTok at 15.32 s and then shows Instagram's "Trial reels" screen and "Create trial reel" button at 34 s without naming the switch. Instagram's creator blog describes the auto-promotion the video refers to: Instagram will "automatically share your trial reel with followers if we determine it's performing well based on the views it receives within the first 72 hours."
  • Clarification: TikTok's documented split testing is an Ads Manager campaign feature, not an organic posting option.
  • Clarification: The retention data is scraped, not fetched from an API. The workflow in frame 009 drives a logged-in headless Chromium with --disable-blink-features=AutomationControlled at https://www.tiktok.com/tiktokstudio/analytics/${itemId} and watches for a video_retention_rate_realtime payload, with a detectBlockSignal guard.
  • Clarification: "notebook.lm" at 53.00 s and "Hormozy" at 6.68 s are Whisper transcription artifacts. The video's prompt card writes "NotebookLM" and its title card writes "Alex Hormozi."
  • Verified: NotebookLM can ingest the lessons table the video describes. Google added Google Sheets on November 13, 2025 to "Add structured data and ask information on key statistics or generate summaries from your spreadsheets," and the help page lists CSV among supported sources, with limits of 500,000 words or 200MB per source.
  • Context: The video shows a real working artifact. Frames 010, 011, 013, 023, 025 and 027 show a populated retention report, a passage breakdown, an evidence table with confidence labels, a 458-video dashboard and a NotebookLM answer. What no frame I read shows is the scraper executing: the code in frame 009 appears as static source cards inside a diagram, with no terminal transcript, no run output and no live TikTok Studio page being driven. The "VIDEO POST-MORTEM" prompt card is partially occluded by the speaker's head in every frame I read, including the fps=2 re-extraction, so its steps 3 through 5 are never fully legible.
  • Context: The tool's uncertainty vocabulary is more disciplined than the pitch around it. It labels findings "settled" or "hunch" with a checked date, prints "REFUTED: gain + proof" as a saved row, badges thin data "Early data · descriptive comparison," and writes "Measure the revised cut before promoting it."

Resources

Sources I could not reach and therefore do not cite: two support.tiktok.com analytics articles and tiktok.com/creator-academy/en/article/tool-analytics-intro returned navigation markup with no article body, developers.tiktok.com/doc/display-api-specification-v2/ returned HTTP 404, and help.instagram.com/412709662619410 returned only a "Help Center" heading. I also did not fetch the LinkedIn post or the onlyhappynews.com article visible in frames 003 and 004, so I describe both only as on-screen content.

Published October 22, 2026. Writeup generated from a favorited TikTok.