AI summary by TwinMind
What is Wolfram Language? (Stephen Wolfram) | AI Podcast Clips
Lex Fridman
40m 25s
Wolfram Language Overview & Philosophy
- Wolfram Language is symbolic computational language, distinct from traditional programming languages
- Traditional languages built around computer intrinsic operations; Wolfram Language represents things existing in world/imagination
- Symbolic nature: typing "X" doesn't produce error - could represent city of Boston, spacecraft trajectory, etc.
- Built from transformation rules for symbolic expressions as fundamental computation building blocks
- 6,000 primitive functions covering broad range of real-world concepts
- Examples demonstrated: BooleanConvert, DollarRequestAddress, DiscreteWaveletData, ImageIdentify, GeoNearest
Mathematica & Wolfram Alpha Products
- Mathematica: first instance of Wolfram Language, released 1988
- Used for technical computations via typing computational language pieces
- Wolfram Alpha: natural language question-answering system
- Converts natural language queries into Wolfram Language computations
- Uses underlying Wolfram knowledge base
- Example queries: population comparisons, nearest volcanoes with mapping
Machine Learning Integration
- Image identification function demonstrated with webcam
- First attempt: 57% plunger, 8% hominid probability
- Second attempt: 89% person probability
- Modern ML techniques integrated for natural language processing
- Hybrid approach combining symbolic/knowledge-based methods with statistical ML
- State-of-the-art transformers (BERT) integrated
- Loop between NLP (machine learning) and NLU (computational methods) improving both
Knowledge Base Construction
- Started as "scary" project to ingest entire reference library
- Methodology: implement area by area rather than grand unified theory
- Leveraged world experts in various domains for input
- Principle: any area without expert help wouldn't be done right
- Goal: automatic answering of any question based on general civilizational knowledge
- Marvin Minsky testing anecdote - tried random medical, chemistry, astronomy queries
AI & Expert Systems Historical Context
- 1980s-90s expert systems were "too modest" - focused on particular areas
- Natural language understanding requires broad world knowledge
- "Easier to do whole thing than piece of it"
- Wolfram approach: encode civilization's accumulated knowledge computationally
- Contrast with pure learning approaches that "start from scratch"
- Merger of symbolic knowledge representation with statistical pattern recognition
Distribution & Adoption Challenges
- High penetration among high-end R&D experts
- Lower adoption among "plodding" practitioners
- Company emphasizes product quality over commercial promotion
- Free Wolfram Engine for developers available
- Site licenses at major universities (effectively free for students)
- Wolfram Alpha main website is free
Computational Contracts & Symbolic Discourse
- Goal: represent legal contracts in computational form
- Example applications: self-driving car ethical decisions, content ranking
- AI ethics modules needed for automated content selection
- Multiple AI systems with different ethical "brands" (conservative, liberal, libertarian, etc.)
- Market-based approach where individuals choose their preferred ethical system
- Symbolic discourse language to express everything in computational symbolic form
Technical Infrastructure & Innovation
- Wolfram Language existed for >50% of total computer language history (33 years of ~60 year field)
- More innovation in software distribution than programming language structure recently
- Simple private company structure allows flexibility without investor constraints
- Challenge of maintaining coherent vision over decades while handling "infinite tails" of projects
Future Directions & Turing Test
- Computational contracts within reach in "limited number of years"
- Symbolic discourse language to handle common sense, human desires/wants
- Turing test: Wolfram Alpha fails traditional test due to obscure knowledge queries
- Argues Wolfram approach legitimately addresses Turing's original intent
- Turing had similar encyclopedia digitization ideas but lacked automation layers
