AI summary by TwinMind
Are Hallucinations Popping the AI Bubble?
Sabine Hossenfelder
8m 37s
Personal Investment Context
- Speaker bought Nvidia stocks weeks ago to capitalize on AI boom
- Stocks dropping ever since purchase
- AI investor panic causing stock market decline
Current AI Bubble Analysis
- Current bursting bubble is specifically large language models (LLMs), not AI generally
- Predicts AI enthusiasm will resume when people understand AI extends beyond LLMs
- Expects personal stock recovery
Large Language Model Problems
- Hallucinations: Main known issue where models confidently produce nonsense
- Historical examples:
- 2022 lesson learned but ignored
- Lawyer used ChatGPT for defense, cited non-existent cases
- Improvements made:
- ChatGPT now recommends books that actually exist
- Still refers to made-up papers/reports
- MidJourney now generates 5 fingers per hand (even when asked for 6)
- Root cause: Tying output closely to training set appears to solve problem but doesn't address fundamental issue
Fundamental Problem with LLMs
- Core issue: For LLMs, correct output is quantifiably close to wrong output
- Colin Fraser example: Wolf, goat, cabbage riddle
- Original: Farmer must transport all three across river, boat carries one + farmer
- Constraints: Wolf eats goat, goat eats cabbage if left unattended
- Solution requires taking one item back on return trip
- When constraints removed from question, LLM gives same answer (now nonsensical)
- Pattern recognition problem: Models extrapolate from training data based on word similarity, not content accuracy
- Metric mismatch: Humans and models use different metrics for "good output"
- Unfixable by scale: Can't solve by training on more input data
- Missing information: Models lack understanding of what makes output good for humans
Solution: Neuro-Symbolic AI
- Approach: Teach AI logic using symbolic language (similar to math software)
- Implementation: Combine with neural networks = "neuro-symbolic"
- Benefits: Fixes many LLM problems
Existing Success Examples
- DeepMind mathematical proofs:
- Mentioned in January: remarkable progress on AI for mathematical proofs
- Recent update: AI reached silver medalist level in Math Olympics
- Provides human-understandable proofs
- Key point: AI can parse problems and formulate logically correct answers
- Potential application: Apply logical rigor to verbal arguments
- Hypothetical impact: AI that wins every internet argument would make "Reddit a ghost town overnight"
Expert Validation
- Not new concept: Pointed out for decades by computer scientists
- Key advocates: Gary Marcus, J Kerr, others
- Speaker's position: Agrees you can't just train LLMs on more text/images expecting understanding
Implementation Challenges
- Wittgenstein's linguistic confusion: No two people use words to mean exactly same thing
- Training data contamination: Text from billions of people washes out logical relations
- Human logic limitations: People aren't naturally good with logic
- Consequence: Already trained LLMs will need retraining
Fundamental Philosophical Issue
- Core problem: "World is not made of words"
- Deepest level: World is mathematics
- Proper AI approach:
- Start with mathematics
- Build models of physical reality
- Layer words on top
Market Implications
- Investment risk: Companies heavily invested in LLMs might never recover expenses
- Future winners: Those building AI on logical reasoning and real-world models
- Example winner: DeepMind
DeepMind's Approach
- Virtual mouse project: Created virtual mouse in virtual environment
- Neural network: Modeled after real mouse brain
- Future prediction: Virtual cats chasing virtual mice on computer screens during Zoom calls
- Google connection: DeepMind acquired by Google in 2014
- Speaker's position: Hasn't lost faith in Google stocks
Future Predictions
- AI development direction: Less focus on words, more on physics
- Market application: People will use AI to manage stocks
- Overall outlook: "It'll be great"
Sponsored Content
- Brilliant.org recommendation: Courses on neural networks and LLMs
- Features: Interactive visualizations, follow-up questions
- Topics: Science, computer science, mathematics
- Offer: 30-day free trial, 20% off annual premium subscription via brilliant.org/sabina
