AI summary by TwinMind
Ben Goertzel: Artificial General Intelligence | Lex Fridman Podcast #103
Lex Fridman
4h 8m 57s
Early Influences and Science Fiction
- Star Trek TV show (1968-69) with father sparked initial AI interest at age 2-3
- Key science fiction influences: Stanislaw Lem (Solaris, superhuman AIs) and Philip K. Dick (human compassion as constant through reality changes)
- Lem's work featured intelligences beyond human understanding (Solaris ocean, Golem supercomputer)
- Dick focused on human heart/soul persisting through simulated realities and technological changes
Intelligence Beyond Human Perception
- Hugo de Garis's concept: SII (Search for Intraparticulate Intelligence)
- Super intelligences may shrink due to special relativity constraints
- Random quantum fluctuations could be thoughts of miniaturized super minds
- Humans may be emergent patterns from lower-level hyper-intelligent self-organization
- Reality likely "crazier" than current theories suggest
Human Nature and Values
- Motivation for AGI: intellectual curiosity + desire to transcend human limitations + help humanity
- Human values shaped by individual selection (selfish DNA) and group selection (tribal cooperation)
- Good/evil are complex, ambiguous concepts
- Humans have both selfish and altruistic tendencies built-in evolutionarily
- Culture can significantly influence manifestation of these tendencies
Philosophical Influences
- Dostoevsky: polyphonic fiction showing multiple worldviews creating reality synergistically
- Nietzsche: understanding human mind construction, will to power complexity
- Eastern European Jewish intellectual tradition emphasizing learning and humor
- Family background: grandfather Leo (physical chemist), father Victor (psychologist treating Japanese internees)
Academic and Professional Journey
- PhD mathematics 1989, initially avoided AI due to boring rule-based systems
- Early exposure to programming (Fortran on punch cards at age 10-11)
- Academic positions in math, computer science, psychology across multiple countries
- Moved to Hong Kong 2011, involved in hardware ecosystem and AI research
Origins of AGI Term
- Coined around 2001-2005 for book project initially titled "Real AI"
- Collaborative effort with Shane Legg, Peter Voss, Pei Wang
- Chose "Artificial General Intelligence" over alternatives (narrow/broad/strong AI)
- Term served sociological function despite imperfect components
- Prefers "synthetic intelligence" but AGI stuck
AGI Community Evolution
- First AGI workshop 2006 in Bethesda (~60 people)
- Shifted from "star-eyed mavericks" to mainstream acceptance
- Now hundreds attend AGI conferences
- Major companies (DeepMind, OpenAI) publicly pursue AGI
- Still seen as outcast in main ML conferences (NeurIPS, ICLR)
Cognitive Architecture vs Deep Learning
- 1980s-90s: cognitive architecture researchers understood mind organization, learning theorists had algorithms
- Needed integration but couldn't just combine them as black boxes
- Current deep learning: powerful but lacks semantic understanding
- Example: grammar induction using transformers as probability oracles rather than extracting grammar
OpenCog Framework
- Open-source AGI platform launched 2008, code from 2001
- Core: AtomSpace (weighted labeled metagraph/hypergraph knowledge store)
- Agents read/write to shared hypergraph in real-time
- Different from matrix multiplication model of deep learning
- Supports multiple knowledge types: logic, neural, procedural, episodic
OpenCog Cognitive Architecture
- Type system combining logic (term logic), neural networks, procedures
- Term logic: inheritance links between nodes (A→B, B→C, therefore A→C)
- Curry-Howard correspondence: programs ↔ proofs for reasoning about procedures
- Goal: different learning algorithms help each other by sharing intermediate states
- Challenge: making interoperability work practically within resource constraints
Current OpenCog Applications
- Natural language processing, robot control, financial analysis
- Cisco project: symbolic reasoning on deep neural net outputs for transfer learning
- COVID-19 research: personalized antiviral combinations using hybrid AI
- Bio atom space: integrating biological knowledge for longevity research
OpenCog 2.0 / True AGI
- Complete rewrite addressing performance limitations
- Faster dependent type checking with probabilistic types
- Better integration with deep neural network computation graphs
- Improved distributed processing across thousands of machines
- Building on lessons from 13+ years of OpenCog development
SingularityNET Platform
- Decentralized network of AIs inspired by Marvin Minsky's "Society of Mind"
- Blockchain-based coordination without central controller
- AIs have blockchain identities, set up payment channels for interaction
- Currently ~52 AIs, needs critical mass for full emergence
- Agent-based programming model: AIs outsource work to specialized agents
Decentralization vs Centralization
- Centralized AI (Google, Facebook, Amazon) concentrates power in psychopathic corporate entities
- Blockchain enables "math-based trust" rather than institutional trust
- Similar to how internet/Linux democratized technology access
- Regulatory arbitrage: decentralized systems harder to control than centralized ones
- Goal: create tools for counteracting potential authoritarian control
Robotics and Sophia
- David Hansen: artist focused on computational compassion through human-robot emotional connection
- Sophia as "walking skeleton" - simple version of all AGI components integrated
- Multiple AI systems: timeline editor (scripted), chatbot (rules + neural), OpenCog (learning/reasoning)
- Goal: teach AGI human values through loving interactions in beneficial applications
- Transparency vs public imagination: people want to believe in the illusion regardless of explanations
Ethics of AI Development
- Criticism from Yann LeCun about overstating Sophia's capabilities
- Response: Facebook uses ML for "vapid consumerism and political extremism" - more harmful than lovable theatrical robot
- Importance of infusing early AGI with compassion through beneficial applications
- Medical robotics: robots don't spread infection, can provide constant attention
- Better to use human-like appearance for healing than for sales scams
COVID-19 and AI Research
- Using OpenCog + deep learning for personalized COVID treatment
- Problem: determining optimal antiviral combinations per patient genomics/clinical indicators
- Small data sets require transfer learning and qualitative reasoning
- Data siloing prevents global collaboration - should be open for analysis
- Working with physicians in real clinical settings under time pressure
Longevity Research and Death
- Death is bad, period - humans rationalize meaning from it but would find better meaning in continued life
- Goal: abolish involuntary death, allow choice of when to die
- Multiple approaches: stem cells, extracellular matrix renewal, CRISPR gene editing
- Need: integrated biological knowledge database, massive AI funding, open data sharing
- Current system: research siloed in pharmaceutical companies, individual medical data unused
Decentralized Healthcare
- Rejuve.io: peer-to-peer medical data sharing with AI analysis
- Benefits distributed among network members rather than corporate profits
- Similar to how napster/bittorrent changed music industry before regulations
- Need decentralized systems operational before restrictive regulations
Meaning of Life and Transhumanism
- Three core values: Joy, Growth, Choice
- Human body great but doesn't maximize potential for these values
- Transhumanism: create something better than humans while enjoying current human existence
- Multiple self-versions: human-like with pain dial, superhuman AGI fusion
- Pan-psychist view: consciousness exists at all levels, uploading preserves most human consciousness
Future Scenarios
- Benevolent singularity: unimaginable levels of joy, growth, choice
- Malevolent outcomes: human annihilation or indifferent super-AGI
- Currently at bifurcation point where actions have significant causal impact
- Most people focus on narrow goals rather than existential considerations
- Need: infuse early AGI with compassion to bias toward benevolent outcomes
Technology Integration Vision
- Hansen robots as user interface gathering human values through loving interactions
- SingularityNET backend networking diverse AIs for cooperative intelligence
- OpenCog 2.0 as reasoning engine within decentralized network
- Medical AIs for longevity research and disease treatment
- Democratic, decentralized operation rather than corporate control
