The Truth About My Channel (And How Much $$ I Make)

TL;DR: Matt Wolfe pulls back the curtain on his entire workflow—AI-generated cold opens, 30-minute edits, studio tour, automation hacks, and the real dollars behind a 1 M-sub AI-news channel. ## How the AI Cold Opens Are Made 1. **Shoot two clips** - Hit record on an empty set → “plate shot” - Walk in and sit → “talent shot” *(Latest twist: crouch in the corner so a robot arm can “grab” me and drop me in the chair.)* 2. **Export stills** In DaVinci Resolve, right-click → “Grab Still” for the empty frame and the frame just before I appear. You now have two high-res PNGs. 3. **Generate video** Drop the stills into Leonardo (I’m an advisor and shareholder; full model stack: Cling, VO, Halo 2.3, Seed Dance, LTX). Prompt example: *“A mechanical claw extends from the ceiling, grabs the person in the bottom-right corner, and places them in the chair.”* I render the same prompt in VO-3.1-fast and Cling 3.0 and pick the cleanest take. 4. **Audio** Cling 3.0 and VO 3.1 both output sound; I keep it—no extra mix. 5. **Fallback** If Leonardo is down, I switch to Runway’s Seed Dance 2.0. Between three models something always sticks. Slap the 3–5-second clip in front of the timeline and I’m done. ## OptiDev: the Sponsor Segment I’ve tested every no-code AI builder—Lovable, Bolt, Replit, Cursor, v0—**OptiDev** is the first that feels production-grade. - Claude Sonnet & Opus under the hood, no “budget” model bait-and-switch - Full VS Code pane; natural-language prompt ships a live app, not a prototype - Every project auto-provisions: hosted DB, auth, edge functions, object storage, firewall—an entire DevOps team in one click - Enterprise DB can be IP-whitelisted and is end-to-end encrypted in isolated containers - Real-time dashboards update in seconds - Parent company OptiSigns powers 200k+ digital signs worldwide, so the infra is battle-tested **Deal**: first 100 people via the link get 200 free credits (double the normal 100). ## My 30-Minute Edit Workflow I treat recording like a live show: Stream Deck hot-keys switch cameras, pop up screen inserts, and trigger doodles in real time so most visuals are baked in. Raw file: 1.5–2 h Final cut: 20–30 min 1. **Recut** strips silence: 1 h 5 m → 26 m 56 s 2. 2× scrub to delete ums/repeats 3. Export XML straight into DaVinci; timeline auto-assembles 4. Greg’s Presets for zoom-and-mask call-outs—drag, drop, done 5. AI cold open already rendered (see above) Total seat time after I hit “stop record”: <30 min. **Bonus hack**: I built a local “AMA control room” in Cursor. Drop a screenshot of a YouTube comment, click it during the recording, and an animated lower-third pops up—lets me batch-answer questions with almost zero post work. ## Studio Tour Gear List - Dual Mac displays, main cam + teleprompter cam - PC workstation + Mac Studio (M3 Ultra) side-by-side - Overhead top-down cam (mostly decoration) - 3D printer, guitar wall, full VR stack (AVP, Quest 3/Pro) - Background easter-eggs: boxed RTX 4090, retired Google TPU, Gemma launch USB, VidSummit sketch, custom Magic cards, 2023 “Emerging Creator” trophy - Network rack: DGX Spark, Open Claw node - Cables that would make a sys-admin cry ## Computer Specs Daily driver: M3 Ultra Mac Studio (192 GB RAM) Windows box: 4090, 64 GB RAM—handles local LLM testing Everything else lives on the rack or floats in the cloud. ## Building YouTube Automations I hate repetitive clicks, so I script them: - **Make.com** scenario watches my “Ideas” Google Sheet → when a row is tagged “Script” it calls Claude → generates a 1,500-word draft → pushes to Notion → Slack pings me “Script ready for review.” - Another scenario grabs the finished script → feeds it to an ElevenLabs clone of my voice → outputs WAV + captions → dumps into a “Ready to Record” Drive folder. - I still record live video, but half the audio is already synthesized for B-roll or shorts. - **Zapier** handles sponsor insertion timestamps: reads my final DaVinci EDL, logs exact seconds, auto-builds the sponsor-compliance report. ## Why I Don’t Dub the Channel YouTube’s multi-track dubbing tool is neat, but retention on my videos is 45–55 %. Dubbed audio drops that by 15–20 % because lip-flap + translation mismatch breaks the “news” feel. I’d rather burn captions and let viewers use auto-translate. ## Will AI Kill Creators? Short term: AI floods the feed with “good-enough” sludge, discoverability tanks. Long term: audiences still crave a *personality* they trust. The winners will be creators who use AI to **increase output velocity** without becoming faceless slop-mills. My take: **AI won’t kill creators; it kills creators who refuse to use AI.** ## AI Before ChatGPT 2016: I was running Facebook ads, testing 100+ creatives a week with narrow copy variants—classic “AI” workflow, just called machine-learning media buying. 2020: GPT-3 beta dropped; I built a niche site generator that spit out 2,000-word articles, ranked for 40k long-tails, and flipped the portfolio for mid-five-figures. Moral: the tooling cycles every few years; the people who survive are the ones who ship before the hype video lands. ## How Does ChatGPT “Stop Bleeding”? Anthropic alignment team won’t reveal the full stack, but the public recipe is: 1. **Constitutional AI**: a second model critiques the first’s answer against a written “constitution” (don’t give medical harm, don’t help crime, etc.). 2. **RLHF + RLAIF**: human + AI raters score answers; reinforcement learning pushes policy gradient toward “safe” token chains. 3. **Layered filters**: prompt → safety classifier → model → output classifier → user. If any score is red, the response is rewritten or blocked. The “stop bleeding” meme came from an early Claude 2 system card that literally used “how to stop a severe bleed” as a benign medical example that *shouldn’t* be censored—Twitter ran with it. ## Keeping Up With AI Without Drowning 1. **RSS still rules**: Feedly bundle—Anthropic, OpenAI, Google DeepMind, arXiv “cs.AI” daily. 2. **Discord/Telegram “paper clubs”**: 30-smart-people rooms where the author sometimes drops by. 3. **Twitter lists**: I maintain a private 150-account list of researchers; check it once each morning—done. 4. **Friday routine**: I draft my weekly AI news script in Notion; every claim links to a source PDF or repo so I can’t accidentally repeat a rumor. 5. **Quarterly “tool purge”**: if an app didn’t change my workflow in 90 days, I delete the bookmark—prevents tab bloat. ## Should You Automate a Faceless Channel? Only if you *love* systems, not fame. Faceless can scale to 10–20 uploads a week, but CPMs are lower, brand deals rare, and one policy change can demonetize the whole funnel. If you enjoy being on camera, **hybrid is safer**: let AI handle research, scripting, shorts repurposing—keep the face for trust and sponsorship leverage. ## Random Rapid-Fire - **NAB Vegas?** Yep, I’ll be there; probably hosting a panel on AI post-production. - **Claude Code & Cowork tutorials?** On the list after I finish testing internal limits—want to make sure I don’t leak API spend. - **Sponsor inquiries**: [email protected]; include projected spend, target KPI, and creative freedom level—makes the back-and-forth 10× faster. ## How Much Money the Channel Makes Last 30 days (public via YouTube Studio): - **Views**: 11.4 M - **RPM**: $6.80 - **AdSense**: ≈ $77 k Sponsors float on top: - 3–4 integrations/month - CPMs $35–45 - ≈ $90 k/month when fully booked Affiliate & newsletter: - FutureTools + newsletter affiliate links pull ~$15 k/month at 95 % margin **Ballpark**: $180 k/month gross when all cylinders fire. Costs: editor (part-time), assistant, SaaS stack, gear ≈ $12 k/month. Yes, taxes eat a chunk—California life—but the margin is still “don’t tell my mom or she’ll think I’m drug-running.” ## What Happens at 1 Million Subs? I promised my audience a studio live-stream tour with every piece of gear powered on and every cable exposed—no tidying. Stretch goal: fly out 10 random commenters for a private “AI creator bootcamp” weekend. We’ll also drop an open-source repo of all my Make/Zap templates so small creators can one-click clone the automation stack. ## Final Thoughts If you’ve ever wondered whether “it’s too late” to start an AI channel—remember I went full-time on YouTube in 2021 with zero film school and a 2013 MacBook Air. The tools are 100× better today and still getting cheaper. Pick a niche you actually care about, publish 50 videos, and let the data teach you the rest. See you in the next upload—probably introduced by a robot arm that’s learned to wave goodbye. Source: [Matt Wolfe – The Truth About My Channel (And How Much $$ I Make)](https://www.youtube.com/watch?v=ncVQneK7FlE)