AI summary by TwinMind
10 differences between artificial intelligence and human intelligence
Sabine Hossenfelder
6m 46s
Introduction to Artificial Intelligence
- AI today = neural networks
- Neural networks = computer algorithms imitating certain human brain functions
- Contains virtual neurons arranged in layers, connected to each other
- Neurons pass information, perform calculations like human brain neurons
- In neural nets: neurons = numbers in code (typically 0-1 values)
- Connections have associated numbers called "weights"
- Weights determine how much information from one layer matters for next layer
- Neuron values + connection weights = free parameters of network
Neural Network Training Process
- Training = finding parameter values that minimize loss function
- Optimization problem solved through back propagation
- Back propagation: if poor result → go back, change neuron weights + connections
- Allows network to learn from failure
- Recommendation: 3Blue1Brown 20-minute video for neural net introduction
Key Differences: Human Brain vs Artificial Intelligence
1. Form and Function
- Neural net = software on computer
- AI neurons = not physical, encoded in bits/strings on hard disks/silicon chips
- Physical structure looks nothing like actual human brain neuron
- Human brain: form and function integrated
2. Scale/Size
- Human brain: ~100 billion neurons
- Current neural nets: few hundred neurons
3. Connectivity
- Neural nets: each layer fully connected to previous/next layer
- Human brain: no true layers, relies on predefined structure
- Not all brain regions equally connected
- Brain regions specialized for specific purposes
4. Power Consumption
- Human brain: dramatically more energy efficient
- Brain uses ~20 watts (comparable to standard laptop)
- With 20 watts, brain handles 1 million times more neurons than AI
5. Architecture
- Neural networks: layers neatly ordered, addressed sequentially
- Human brain: extensive parallel processing, no particular order
6. Activation Potential
- Real brain neurons: binary (fire or don't fire)
- Neural networks: continuous values mimicking firing
- AI neurons can smoothly slide from off to on (real neurons cannot)
7. Speed
- Human brain: much slower than AI systems
- Standard computer: ~10 billion operations per second
- Real neurons: fire at maximum 1,000 times per second
8. Learning Technique
- Neural networks: produce output → if low performance per loss function → change weights
- Human learning: mechanism unknown in detail, but not same as neural nets
9. Structure
- Neural nets: start from scratch every time
- Human brain: pre-wired structure + connectivity
- Brain draws on evolutionary-proven models
10. Precision
- Human brain: much more noisy, less precise than neural nets
- Brain cannot run same learning mechanism as neural net
- Probably uses entirely different mechanism
Consequences of Differences
- AI needs extensive training with carefully prepared data
- Very unlike human intelligence operation
- Neural nets don't build world models
- Instead learn to classify patterns
- Pattern recognition fails with small data changes
AI Vulnerabilities
- Famous example: small noise added to images (invisible to human eyes)
- Can fool AI into misclassifying (turtle → rifle)
- Not good at generalizing learning from one situation to next
- Success depends heavily on defining correct loss function
- Poor loss function = optimizing wrong thing
- Example: simulated self-driving car spinning in circles to maintain constant high speed
AI Strengths
- Excel at image classification
- Good at extrapolating data without well-understood trends
- Point may not be to mimic human intelligence
- Most useful machines (cars, planes) useful because they don't mimic humans
- May want machines specialized in tasks humans aren't good at
