How I trained our AI on thumbnails
Team tutorial for Marcus and Blendi · cut 2 · 2:18 (cut 1 was 2:45) · 1080p.
- Faster: 2:45 to 2:18. The narration is my AI voice (voice G), tight lines, with 3 real clips of me kept where my own voice matters: the round 4 "basically solved" review, one pen-mark review, and the meeting close.
- The whole progression, from before the A/B system: designers, the AI's first rounds (Nov 2) and how bad they were, the same Lambo idea over rounds 1 to 3, the face fix, the look, then the A/B loop going live (Nov 6) and the backgrounds.
- More before and after: side-by-side pairs for the hook, the look, the yoyo video and the backgrounds, plus the same idea across rounds.
- The meeting's "past 7 days" panel was showing all-time totals. From Nov 4 to Nov 15 the loop made 2,213 thumbnails and ran 439 A/B tests (29 of them set by hand on Nov 4 and 5, before the loop went live on Nov 6). The extra views are about +2.7M, which is the loop's own estimate. The video says it that way. A true 7-day window would be about 1,360 thumbnails and 290 tests.
- Yoyo ($1 vs $100,000 Yoyo): before the test it was getting about 5,400 views a day (Oct 26 to Nov 5). The winner came from test 203 (Nov 6 to 7), with a 0.41 watch share against 0.28 for the original. It reached about 68,000 a day by Nov 13. The video calls it "an early win", not the first, because I could not prove it was the first.
Script, as built
Tap a time to jump the video there. Every quote on screen is my words, verbatim, from the file named in the v3 script.
| 1. Where we started | ||
| 0:00 | BEFORE round 1 Lambo (Nov 2, “this doesn't really look like me”) / AFTER background round 13 lab (“I love these”) | “At the start of November / our AI thumbnails looked like this / Two weeks later / they look like this” |
| 0:07 | How I trained our AI on thumbnails / And how you do the same for your area, every week | “Here's how I trained it round / by round and how you'll do / the same for your area” |
| 0:12 | First: two designers. Too slow · Quality: mid · My face: wrong | “First / we tried designers / We tested two / One was too slow / one was mid / and they kept getting my face / wrong” |
| 0:20 | Round 1 · Nov 2 (the first round of the v2 test; there were rougher AI composites before it, Oct 27): arm (“This has way too much text”), minigun (“This doesn't really look like a minigun”) | “Then the AI had a go / Round one was rough / Wrong face / weird expression / too much text” |
| 0:28 | Same Lambo idea: round 2 (“this person does not look like me”), round 3 (“it looks a little creepy”) | “Round two / same Lambo / still not me / Round three / better / but creepy” |
| 2. One area at a time | ||
| 0:35 | Face swap: “just go test out in a hundred different ways” | “I had it test face swapping / a hundred different ways and kept / giving notes / By round four…” |
| 0:41 | Round 4 from his review video: Mario TV, face made by the AI vs his real face | “I think likeability / meaning how much it looks like me / is basically solved” |
| 0:48 | BEFORE v4 ruler in a classroom, Nov 4 (his note on that batch: “These images look too realistic ... It's not really optimized for YouTube”; he liked the ruler itself and asked for it straight, cutting something else) / AFTER the rework from his notes | “Next / the look / It was making photos / not thumbnails / I pushed it to one big / idea / Bright and clean” |
| 3. Then it tested itself | ||
| 0:56 | Nov 6: A/B testing on YouTube. The AI makes the thumbnails, sets the tests, keeps the winners | “By November 6th / it was good enough to run / A/B tests on YouTube by itself” |
| 1:02 | $1 vs $100,000 Yoyo: BEFORE original (about 5,400 views a day) / AFTER A/B winner, Nov 7 (about 68,000 views a day) | “One early win / an old yoyo video went from / about 5,000 views a day to / about 68,000” |
| 1:10 | Nov 4 to Nov 15, from the A/B loop's stats page: 2,213 thumbnails made · 439 A/B tests · ≈2.7M extra views (estimated); 29 of the 439 tests were set by hand on Nov 4 and 5 | “Since November 4th / that's over 2,200 thumbnails / 439 A/B tests / and an estimated 2.7 million extra / views” |
| 4. How I review | ||
| 1:22 | Backgrounds, about 20 rounds in under a week: BEFORE background round 1 (“this sunlight thing is pretty AI looking”) / AFTER round 8 (“I like everything about this”) | “Backgrounds still looked fake / That's where I went next / About 20 rounds in under a / week” |
| 1:28 | His real background round 4 review page; steps: Open the page · Talk while you scroll · Draw on what's wrong (punch-in on the tile he drew on) | “Here's how I review / I open the page / talk while I scroll / and draw right on what's wrong” |
| 1:34 | Background round 4 picture; his pen strokes redraw in time with his voice | “This one a little bit too much detail / in the background / with all the trees and the towers” |
| 1:41 | Background round 4: 11 versions of one idea, side by side; his pick lights up | “Then I have it make 10 / versions and I pick the best / one” |
| 5. Your turn | ||
| 1:45 | What the AI needs from you: the exact spot and what's wrong · your pick out of 10 · what you don't want · what you know that it doesn't · hard rule or suggestion | “What it needs from you / the exact spot / your pick out of 10 / what you don't want / and what you know that it / doesn't / And say if it's a hard / rule or a suggestion” |
| 1:56 | Each week: your biggest problem area. Marcus (thumbnails): Composition. Blendi (long-form edit): Sound | “Your turn / Each week / pick your biggest problem area / Marcus / composition / Blendi / sound / Nitpick it on every version inside / your normal work” |
| 2:07 | One area a week / High standard, detailed feedback, your biggest problem area | “So just high standard quality / very detailed feedback / and focus on one big area / that's the biggest problem a week” |
Not in the video
- Designer thumbnails. The only designer pictures on disk are hand-only sword shots from Nov 9, after the AI rounds and with no face, so showing them as "the designer era" would mislead. That part is words only.
- Pictures of the "100 ways" face-swap test runs themselves. The video shows the rounds that led there and round 4 with my review of it.