AI summary by TwinMind
I asked an AI for video ideas for other YouTubers. It went badly.
Tom Scott
13m 0s
GPT-3 Video Idea Generation Experiment Setup
- Tom Scott tested GPT-3's ability to generate video ideas for different YouTube channels
- Fed AI recent video titles from various channels, asked for similar suggestions
- Generated 100 ideas per channel, selected most interesting ones
- Couldn't do live demonstrations due to OpenAI terms of service
- Most creators maintain idea lists (whiteboards, spreadsheets, text files)
ElectroBoom Channel Test (Mehdi Sadaghdar)
- Channel focuses on technical electronics videos, known for electrical accidents
- AI suggestions analyzed:
- "Charging phones from a Tesla coil" - deemed possible with powerful Tesla coil, needs disposable phone
- "Lemon lime lightning 2 million volts" - clickbaity but potentially impossible
- "24 carat gold-plated electrical banana award" - sounds like electroplating banana with gold
- "Using Tesla's induction motor design to test solar panels" - felt like random words together
- "How to build a 3D printed plasma speaker" - considered thinkable
- Conclusion: GPT-3 struggled with physics constraints, titles were clickbaity but not scientifically sound
Answer in Progress Channel Test (Sabrina Cruz)
- Channel explores big questions through silly quests
- AI successfully matched video title style and click-baitiness level
- Notable suggestions:
- "Why buying pets is a scam" - perfect match for channel's bold hypothesis style
- "How to destroy your career with one email" - connected to Google's April Fools minion emoji prank
- "Why my back hurts so badly" - exactly matched video Sabrina was currently working on
- Best results came from workshopping AI suggestions into better ideas
- Most of 100 suggestions weren't winners, but formatted appropriately for consideration
hbomberguy Channel Test (Harry Brewis)
- Creates feature-length video essays, releases every few months
- AI became obsessed with video games when any gaming reference appeared
- Strong suggestions generated:
- "The Last of Us is disappointing and here's why" - perfect contrarian title format
- "Mass Effect 3 was a mistake and here's why" - close to correct opinion
- "How Bethesda broke me and here's why I'll keep buying their games anyway" - very specific and relatable
- "Quantic Dream, David Cage and the uncanny valley" - excellent title connecting motion capture technology to uncanny valley concept
- AI successfully replicated Harry's contrarian format by substituting different names
- At least 2 ideas were genuinely makeable video concepts
MrBeast Channel Test (Jimmy Donaldson)
- 50 million subscriber channel mixing stunts and philanthropy
- Perfect ideas must be both clickable and entertaining for full video length
- AI suggestions lacked crucial elements:
- "I paid people to take selfies with me" - missing virality factor and twist
- "I bought a haunted house" - needed additional hook like "and offered people a million dollars"
- "I paid people to do weird things on camera" - too generic, described entire channel without specifics
- AI imitated format without understanding human entertainment psychology
- Missing the "and" factor that creates viral potential
Overall Assessment
- AI works best as co-pilot for inspiration rather than direct idea generation
- Can provide starting points that humans can develop into better concepts
- Hit rate was okay but not great across all channels
- Technology that seemed impossible few years ago now dismissed as "just putting words together"
- Cliché about AI: once it can do something, it's no longer considered "real" AI
- Examples: chess (brute force), image categorization (model training), idea generation (word assembly)
