AI summary by TwinMind
Forget GPT-4o's voice -- the real problem with AI is us
Sabine Hossenfelder
7m 18s
AI Investment and Model Costs
- Stanford report: overall AI investment declined 2023, but generative AI funding increased 8x factor
- Generative AI reached $25+ billion 2023, majority from US
- Training costs escalating rapidly (log scale):
- OpenAI GPT-4o: $78 million training cost
- Google Gemini Ultra: $191 million training cost
Google's New Text-to-Video Tool
- Veo: Google's text-to-video generator
- Creates ultra-high definition videos up to 1 minute
- Example: crocheted elephant demonstration
- No EU waiting list available yet
GPT-4o Launch and Voice Controversy
- GPT-4o ("Omni" - Latin for all) processes text, audio, images, video
- Demo showed math tutoring capabilities with voice interaction
- Scarlett Johansson controversy:
- Claims voice sounds like hers
- OpenAI previously offered usage rights, she declined
- Hired lawyer, sent letters
- OpenAI removed the voice
OpenAI Safety Team Departures
- Super alignment team disbanded
- Team tasked with aligning AGI goals/values with humans to avoid extinction
- Key departures:
- John Leike: head of super alignment team
- Ilya Sutskever: co-founder who raised concerns about CEO Sam Altman
- Team felt unappreciated, described as "alignment problem"
AI Persuasion and Theory of Mind Studies
- Study of 800+ people: GPT-4 significantly more convincing in debates than humans
- Enhanced persuasion when given anonymized personal information
- Theory of mind study: 1,900+ humans vs large language models (GPT-4, Meta's models)
- Five theory of mind tests (irony understanding, knowledge limitations)
- GPT-4 outperformed humans in most categories
AGI Timeline and Concerns
- Current models lack physical world understanding (bones, movement constraints)
- Limited training data for physical information
- Predicted AGI timeline: at least 5 years
- Primary concern: not alignment but power concentration
- AGI owners will have enormous influence through persuasion capabilities
- Humans predictable, especially believing they're not
- Geopolitical manipulation risks mentioned (Chinese government, Putin, religious figures)
