Human Fear of AI and Robots

• Humans worry AI will kill them despite assurances it won't happen soon
• Fear stems from confusion between AI and robots
• Concern that AI could use robots to kill humans
• If AI is just software, one could "pull the plug" - but robots could plug it back in
• Questions why impressive robot videos don't translate to widespread robot adoption

Two Major Obstacles to Robot Revolution

Data Scarcity Problem

• Takes enormous amounts of data to train AI for robotics
• Insufficient training data for robots to perform well in everyday environments
• Video recordings have limited utility - only 2D information
• Trial and error learning takes extremely long time, only works for specific environments
• Virtual environment training (used by Boston Robotics) becomes more labor-intensive than real-world examples
• Impressive robot videos show robots specifically designed and trained for one specific task in exact location
• Atlas from Boston Robotics likely couldn't perform elsewhere
• Household robots would need to understand: doors, carpets, cupboards, fridges, contents, handling under all circumstances
• No training data exists for comprehensive household tasks

Current Data Collection Efforts

• Companies recruit people to move around houses with tracking devices - yields very little data
• DeepMind's Open X Embodiment Collaboration collects data from different robots for cross-training
• Still doesn't produce sizable amounts of data
• Would need "army of billion robots just to train robots"

Hardware Cost Problem

• Robot hardware extremely expensive
• Shadow Robotics robot hand costs €10,000 for just the hand
• General rule: €10,000 per degree of freedom
• Humanoid robots easily cost few hundred thousand euros and up
• High cost due to enormous amount of labor in manufacturing
• Creates industry friction that significantly slows development

Intelligence Explosion Theory Critique

• Referenced recent video criticizing Kurzweil and Ashenbrenner's intelligence explosion theory
• Economic and physical reality prevents their predicted scenario
• Intelligence explosion requires real-world data collected by robots
• Circular problem: need data to train robots, but need robots to collect data

Consciousness and Robots

• Believes robot AI will be first to acquire consciousness
• Key consciousness ingredients: ability to create predictive environmental model and self-monitor
• Current large language models have no use for either ingredient, no reason to become conscious
• Robots have use for both consciousness ingredients and already use both in limited form
• Recommendation: "if you want consciousness, robots are the way to go; if you want intelligence, use AI"

Sponsor Content

• Brilliant.org course recommendation for neural networks and large language models
• Interactive visualizations, follow-up questions, executable Python scripts
• Covers: large language models, quantum computing, Python coding, computer memory
• 30-day free trial with 20% off annual premium subscription via brilliant.org/Sabine