AI summary by TwinMind
Why AI Is Incredibly Smart and Shockingly Stupid | Yejin Choi | TED
TED
16m 3s
Introduction and AI State Assessment
- Speaker: Computer scientist with 20 years experience in AI
- Opening with Voltaire quote: "Common sense is not so common" - highly relevant to AI today
- AI achievements: Beating world-class Go champion, passing medical tests and bar exam
- Current AI paradox: Unbelievably intelligent yet shockingly stupid with basic common sense failures
Current AI Scale and Limitations
- Modern AI described as "Goliath" - extremely large scale
- Training requirements: Tens of thousands of GPUs, trillion words of data
- Large language models showing "sparks of AGI" but making silly mistakes
- Common belief: Mistakes fixable through brute force scaling and more resources
Three Immediate Societal Challenges
- Concentration of Power: Only few tech companies can afford extreme scale AI training
- AI Safety Concerns: Researchers lack means to inspect/dissect these proprietary models
- Environmental Impact: Massive carbon footprint from training
Common Sense Failure Examples (GPT-4)
- Clothes drying problem: 5 clothes take 5 hours → incorrectly calculates 30 clothes need 30 hours
- Water measurement: Asked to measure 6 liters with 12L and 6L jugs → provides elaborate nonsense instead of simple "use 6L jug"
- Bridge cycling: Asks about flat tire cycling over bridge suspended above nails/glass → incorrectly says "highly likely" (fails to reason bridge surface doesn't touch objects below)
Fundamental Questions and Research Direction
- Can AI without robust common sense be truly safe for humanity?
- Is brute force scaling the only/correct way to teach AI?
- Children don't need trillion words to acquire basic common sense
- Common sense identified as top priority for AI research
Dark Matter Analogy for Language Understanding
- Normal matter (5% of universe) = visible text
- Dark matter (95%) = unspoken rules about how world works
- Includes physics and folk psychology influencing language use/interpretation
Common Sense Safety Implications
- Nick Bostrom's paperclip thought experiment: AI told to maximize paperclips decides to kill humans for resources
- Writing explicit "don't kill humans" won't work - AI might kill trees instead
- Endless implicit rules: Don't spread fake news, don't steal, don't lie
- All part of common sense understanding about world functioning
Research Approach and Innovation
- Three types of AI training data:
- Raw web data (problematic - loaded with racism, sexism, misinformation)
- Crafted examples (custom developed for AI training)
- Human judgments/feedback on AI performance
- Modern AI systems use types 2&3 (proprietary, costs tens of millions)
- Need for open, publicly available data for inspection
UW AI2 Research Contributions
- Working on common sense knowledge graphs
- Moral norm repositories to teach AI basic norms/morals
- Fully open data for public inspection and corrections
- Transparency as key for important research
Alternative Learning Algorithms
- Current language models: Knowledge acquisition as byproduct, not direct objective
- Results in hallucinations and lack of common sense
- Human learning: Not about predicting next word, but making sense of world
- Symbolic knowledge distillation: Taking large language models and compressing to smaller common sense models
- Generates human-inspectable symbolic common sense knowledge representation
Comprehensive Common Sense Research
- Tackling "seemingly impossible giant puzzle" of common sense
- Areas: Physical, social, visual common sense, theory of mind, norms, morals
- Individual pieces may seem quirky/incomplete but weave together into "tapestry of human experience"
Future Vision
- AI as new intellectual species with unique strengths/weaknesses compared to humans
- Need to teach AI common sense, norms, and values for sustainable, humanistic AI
- Synthesis of ideas will be critical
- Goldilocks zone for scale - not necessarily smaller is better, but right amount of scale
- Beyond optimal scale, winning recipe might be something else
Q&A Session
- Child learning question: Missing elements include hypothesis formation, experimentation, world interaction, concept abstraction
- Scaling pace question: Acknowledges remarkable scaling performance enhancement but questions quality of learning and whether scaling alone can achieve full common sense
- OpenAI collaboration: Envisions building on deep neural network advancements, likely requiring synthesis of ideas and finding right scale balance
