Today building Koro.
- 7 am: Lip sync still breaks on frame 47. Tamil phonemes.
- 10 am: Voice drifts 200ms on Meta upload. Fine locally.
- 3 pm: One ad renders clean across ten languages.
Not every day ships a feature. Some days you just make one thing work.
____________________
| AI VIDEO COSTS |
|__________________|
Seedance $0.07/s
Gemini $0.05/s
Kling $0.07/s
Grok $0.05/s
Skip the monthly plans with credit counters. API pay-per-clip is cheaper and honest.
$50/day. 50+ variants. No games.
2 weeks.
That's the new lifespan of a Meta ad. Down from 6 weeks in 2024.
Stop making better ads.
1. Ship 25 variants a week — hooks first
2. Watch CPM creep, not CPA. That's the real fatigue signal
3. Rotate every 3 days
4. AI scales output. You pick direction.
I just found out Meta's Advantage+ rewrites the text inside your uploaded image ads. Not the headline field — the pixels.
Default-on. 8 AI variants per image. Same fonts, same layout, different message.
Your creative isn't yours once you upload it.
Testing creative is so much better than testing audiences, it's not even close
Meta still doesn't even look at your targeting first — it reads the ad and finds the buyer itself
AI creative isn't being commoditized. It's entering its precision phase.
Midjourney V8.2 just shipped localized redrawing and multi-reference. Feed it your brand assets and lock in while testing at scale.
The moat isn't shrinking. It moved from volume to precision.
My process for finding winning hooks with AI:
1. Meta Ad Library: 90+ day ads, unrelated categories
2. Paste into an AI ad-cloning tool
3. It extracts the skeleton, rebuilds around your product
4. Test against your best performer
Your next hook was built for another industry.
78% of enterprises use more than one AI video model.
It's literally:
UGC ads – Seedance
Product demos – Kling
Photoreal scenes – Veo
The winners stopped debating. They started routing.
The model war was always a distraction.
AI UGC gave me my time back — I don't sit in two weeks of silence waiting for a creator to ship a draft anymore.
The fee was never the real cost. The wait was.
You need to be output maxxing, not velocity maxxing.
64% of devs say AI made them 25%+ faster.
89% of execs see no change in output.
Feeling faster isn't shipping faster.
Speed you can't measure isn't speed.
'AI' used to be the pitch. Now it's the thing you hide.
The startup that launched as 'The First AI Admaker' now sells '100% real, no AI' ads.
Human-made just became the premium.
Meta broke its own dashboards then sold you the fix.
Advantage+ is a bidding arms race. CPMs spike 75%. Then Meta releases Frequency Control like a gift.
Same week, freq reporting went opt-in. HTTP 200, empty data.
You can't manage what they won't let you see.
AI ad creative actually works — a 500M-impression study found it beats human-made ads by 19% CTR. But only from scratch. AI modifying an existing ad: zero lift.
This is what it means to let AI actually create, not just remix.
AI UGC is 50-200x cheaper than hiring creators. The best operators stopped choosing.
The stack:
1. Test 20 hooks with AI for the price of one creator brief
2. Cut the losers in the 7-10 day fatigue window
3. Hand the winners to real creators
Cheap finds it. Human sells it.
R.I.P the UGC rejection loop. Ship product, brief ten creators, wait 7-14 days, throw out 40-60% of drafts.
One photo now runs the whole thing — script, actor, language, edits — a finished ad the same day.
The old model didn't die of cost. It died of variance.
We stopped arguing about which video model to use.
The top three are basically tied now. Quality is solved.
What's left is matching the ad to the buyer — actor, language, region. That used to be a 3-week creator campaign. Now it's one photo.
Only 4-10% of your ad concepts ever win. The rest die in 5-14 days.
Resize, recolor, new font — that's one ad in 10 outfits, not 10 ads.
Real variants change:
1. Hook — first 3 seconds
2. Angle — a new promise
3. Format — static vs UGC
1 per $50/day. Resizing isn't testing.