AI summary by TwinMind
AI Generated Videos Just Changed Forever
Marques Brownlee
12m 2s
AI Video Generation Breakthrough
- OpenAI announced Sora - new AI model generating up to 1-minute video clips from text input
- Massive improvement from Will Smith spaghetti videos just 1 year ago
- Same concept as DALL-E but for videos - understands physics, textures, materials, lighting over time
- Feels like another ChatGPT/DALL-E moment for AI
Sora Capabilities & Examples
- Tokyo Street Scene: Stylish woman in black leather jacket, red dress, black boots
- Accurate lighting, materials, skin tones, movements, reflections
- Flaws: background figure gliding weirdly, inconsistent frame rates in water reflections, off camera movement
- Vintage SUV: White SUV with black roof rack on steep dirt road
- Drone footage quality but slightly video game-like due to rock-solid movement
- Golden Retriever Puppies: Playing in snow, heads popping in/out
- Excellent physics of fur, ears, snow flying in slow motion
- Cloud Reading: Young man in 20s sitting on cloud reading book
- 90% realistic - beyond uncanny valley
- Minor flaws: weird eyes, odd page motion
- Spaceman Trailer: 30-year-old in red wool knitted motorcycle helmet, blue sky, salt desert
- Cinematic style, shot on 35mm film
- Most convincing AI video seen - close-up face, fabric textures, film grain
- Only flaw: weird fast-motion walking physics
Current Limitations & Flaws
- No audio generation
- Still has detectable imperfections for trained observers
- Wolf Pups Example: Animals appearing/disappearing, walking through each other
- Treadmill Runner: Physics completely broken
- Birthday Grandma: Multiple hand inconsistencies, weird candle wind direction
- Hands remain problematic area for AI generation
Technical Details & Access
- Currently in super limited access
- Available to red teamers (testing/breaking limits) and select creators
- Watermark in bottom corner of generated videos
- Handpicked best examples shown on OpenAI website
- Sam Altman posting user requests on Twitter
Immediate Implications
- Election Year Concerns: Videos can pass as real to unsuspecting viewers
- Stock Footage Disruption: Already good enough for presentations, ads, PowerPoints
- Big Sur waves drone shot - totally usable for California product ads
- Wall of TVs - expensive/difficult traditional shoot now easily generated
- Historical footage (Gold Rush California) - perfect for western film openings
- Industry Impact: Drone pilots, photographers, videographers facing licensing displacement
Future Considerations
- Safety Concerns: Need restrictions on generating politician likenesses
- Creative Questions: If trained on all human-made videos, can it be truly innovative?
- Rapid Improvement Rate: From spaghetti videos to photorealistic in 12 months
- Complete AI Productions: How long until entire ads, YouTube videos, movies fully AI-generated?
- Current state is worst it will ever be - only improving from here
Technology Progression Timeline
- 12 months ago: Will Smith spaghetti (clearly flawed)
- Now: Photorealistic humans, complex physics, cinematic quality
- Rate of improvement described as "genuinely shocking"
