Source language: English (auto-captions; ASR: "Chad GPT"/"Codeex"/"Wau") | Target language: English Sponsorship/filler skipped (merch/Hostinger, Discord/join, like/subscribe/outro). ## 1. Overview Demo of turning OpenAI **GPT Live 1** (full-duplex voice that listens and talks at once) into a desktop **"waifu"/Wau** assistant with an on-screen avatar. For builders who want a face + interruptible voice—not a Live2D production course. ## 2. Key Points - [00:00] GPT Live 1 is full-duplex; can stop mid-sentence when interrupted (shown vs ChatGPT voice). - [00:56] Voice presets: Quartz, Ripple, Willow. - [01:27] Wire Live 1 through **Codex** (library → thread → answer); creator asks ChatGPT for a Codex prompt instead of hand-coding. - [01:44–02:41] Prompt: AI assistant **with avatar** (PNG overlay first; Live2D Cubism later). Codex ships a runnable app (~21 min generation in captions). - [03:10+] Live chat demos; avatar height adjustable in UI; creator later changes the sprite despite the character objecting. - [08:03–09:59] Gaps: avatar should "know" its look; weak wave/animation; **no cross-day memory** unless reminded. - [10:13+] Source said to be in comments; part 2 gated on likes (avatar suggests **500**). ## 3. Supporting Details Interrupt math demo recovers (e.g. 21+5 → 26). Character self-describes looks (sweater look → later suit/scarf after "fix"). Live2D DIY clip is sampled then abandoned as too slow vs ChatGPT. ## 4. Actionable Takeaways 1. Prefer GPT Live 1 when you need interruptible voice. 2. Scaffold with ChatGPT→Codex before writing Live API glue. 3. Ship PNG/desktop overlay first; defer Live2D. 4. Put visual identity into system/prompt context explicitly. 5. Treat memory + richer animation as follow-ups. ## 5. Notable Quotes - Can "stop mid-sentence to listen to you." - "this just took 21 minutes. I'm really surprised it was that quick." - On memory: "I don't keep memories across days… if you bring it up again… I'll pick up the thread." ## 6. Bottom Line Face + GPT Live 1 duplex voice is quick via Codex; self-aware look, motion, and lasting memory still need work. ## 7. Title question answered? Yes — "isn't so hard": ChatGPT→Codex→running desktop waifu without hand-writing the stack; Live2D-from-scratch is the hard path they skip. ## 8. Questions asked / left unanswered On-cam to avatar: icebreakers, "if you became human," appearance, size, wave, memory. Open: production Live2D, durable memory design, full animation set, exact API versions. ## 9. Counter-intuitive Don't start with Live2D—PNG + Live voice already feels "human." Force-align the sprite even if the character objects. ChatGPT→Codex beats a Live2D basics binge for a first demo. ## 10. Pareto Working **Live 1 duplex loop + visible avatar + one Codex app** = most of the wow; memory/animation are secondary. ## 11. What to avoid Hand-coding Live API before a prompt scaffold; early Live2D rabbit holes; assuming the avatar knows its look or remembers tomorrow; shipping idle-only motion. ## 12. Risks / blindspots Model-name drift in captions; creepy/NSFW "waifu" framing; always-on mic privacy; overtrusting generated code; like-gated part 2; sponsor noise. ## 13. Books / courses None. Own Live2D basics clip (abandoned), source in comments, possible part 2. ## 14. Fact vs opinion - **Demo-grounded:** duplex interrupts; Codex-built avatar app; height UI; no cross-day memory; ~21 min claim; 500-like bit. - **Opinion:** "isn't so hard," "talking to a human." - **Verify:** current GPT Live 1 docs, Codex model IDs, comment source link.