AI summary by TwinMind
A Cult AI Computer’s Boom and Bust
Asianometry
45m 21s
Origins of Lisp and Symbolic AI (1950s-1960s)
- John McCarthy coined "artificial intelligence" in 1956 Dartmouth research project budget proposal
- Herbert Simon, Alan Newell, Clifford Shaw presented first AI program at Dartmouth
- Proved 38 theorems from Principia Mathematica
- Developed symbolic AI philosophy: brain creates abstract internal representations using symbols
- Created Information Processing Language (IPL) for logic theorist program
- McCarthy developed Lisp (List Processing) in fall 1958 with grad student Steve Russell
- Inspired by IPL's list processing concept from magnetic drum memory machines
- Designed for symbolic manipulation rather than just numerical computation
- Featured algebraic programming style, function composition capabilities
Lisp Language Characteristics
- All objects (code and data) expressed as lists that can grow/shrink dynamically
- Supports metaprogramming: programs can take code as input and output new programs
- Uniform syntax with heavy parentheses structure facilitating recursion
- Extensible language allowing user-created variants
- Memory-hungry due to extensive list usage
- Dynamic typing with tagged architecture for type tracking
Government Funding and AI Development (1960s-1970s)
- DARPA created post-Sputnik to ensure American technological superiority
- J.C.R. Licklider hired as IPTO director in 1962
- Believed computers should be creative partners augmenting human intelligence
- Funded AI research at universities nationwide
- His projects led to graphical user interfaces and ARPAnet
- Early Lisp memory constraints on PDP-10 and VAX minicomputers
- PDP-10 had 256K words (36 bits each) ≈ 1 MB RAM
- Time-sharing systems with 20+ users caused hours-long delays
- Lisp nodes used full words, limiting total list capacity
Birth of Lisp Machine Project (1973)
- Richard Greenblatt started MIT Lisp Machine project in 1973
- Inspired by Peter Deutsch's Xerox PARC paper on mini computer for Lisp microcode
- Steven Lukasik at DARPA approved funding: $100,000-$200,000 for special single-user machines
- Cons machine built with reprogrammed general-purpose processor at microcode level
- Featured tagged architecture (36-bit words with data + type tags)
- Operating system entirely written in Lisp
- Graphical user interface with mouse-manipulated windows
Cadr Development and MIT Success (1977)
- Second generation machine called Cadr (Lisp function reference)
- Improvements: microcode data type checks, real-time garbage collection
- MIT made ~35 Cadr machines at $50,000 each ($285,000 today)
- Offered to anyone with sufficient budget
- Became significant hardware operation within MIT AI lab
Company Split: Symbolics vs LMI (1979)
- Russell Noftsker approached Greenblatt about commercializing Cadr
- Philosophical split over company direction:
- Greenblatt: bootstrapped "hacker" company, self-funded growth
- Noftsker: professional company with VCs, managers, salespeople
- Two companies formed:
- Symbolics Inc (Noftsker) - got majority of original team, VC funding
- Lisp Machines Inc/LMI (Greenblatt) - bootstrapped approach
- Richard Stallman (RMS) supported LMI, later founded GNU project
- Both companies licensed software from MIT with improvement-sharing agreement
Early Commercial Competition (1980-1983)
- LMI launched first with $100,000 LMI Cadr in 1980
- Made $1M first year selling to Control Data and Texas Instruments
- Greenblatt's hand-production methods limited scale
- Symbolics LM-2 launched September 1981 due to VC funding delays
- Both companies developed second-generation machines:
- Symbolics 3600 ($78,500 entry level = $250,000 today)
- LMI Lambda
- Symbolics sold 100 first-generation LM-2s at $70,000 each
- By end 1983: LMI sold only 23 machines vs Symbolics' rapid growth
Expert Systems Boom and Market Expansion (1980s)
- Expert systems: rules-based AI capturing human expert knowledge
- Two components: knowledge base (facts/rules) + inference engine
- Notable examples:
- Prospector (SRI International): geological expert system found molybdenum deposit
- Featured in Business Week 1984 "Artificial Intelligence is Here"
- Market drivers:
- DARPA Strategic Computing Program funding
- Japanese Fifth Generation Computer Project ($500M announced, inflated to $4B)
- Fear of Japanese technological competition
- 1984 industry report: $64M spent on AI tools, predicted 50% annual growth 1985-1990
Peak Years and Financial Success (1984-1986)
- Multiple AI startups raised significant funding:
- Syntelligence: $4.3M oversubscribed round
- Inference Corp: $32M valuation, sold expert system frameworks for $60K each
- IntelliGenetics IPO: $9M December 1983, later pivoted to IntelliCorp
- Symbolics IPO November 1984: raised $12M at $6/share
- 1985 Symbolics revenue: $69M (64% of AI hardware market)
- Stock tripled to $16 by December 1985
- 1986 peak: $114M revenue, $11M profit
- CEO Russell Noftsker: "We started the market...unlimited range of opportunities"
Texas Instruments Competition and Challenges
- TI licensed LMI technology, forced licensing deal, created TI Explorer
- October 1984: TI announced 400 Explorer sales to MIT
- Only 200 paid, 200 donated = effective $27K/machine price
- Undercut Symbolics/LMI pricing significantly
- TI's scale and business tactics pressured smaller competitors
- LMI raised outside funding: 25% stake to Texas Instruments
- Greenblatt lost control to VCs and new management
Decline and Unix Workstation Threat (1987-1988)
- Expert systems failed to meet hype expectations:
- Knowledge acquisition bottleneck from human experts
- Tacit knowledge difficult to encode
- Brittle systems, hard to edit and maintain
- Companies bought expensive machines without knowing how to use them
- Unix workstations (Sun Microsystems, Apollo) became competitive:
- Benefited from faster VLSI chips (Intel, Motorola)
- Sun: 6% of $210M revenue from AI applications, growing 50% annually
- Sun workstations: $14K vs Symbolics low-end $40K machines
- Broader software ecosystem Unix offered
- Lisp never achieved industry adoption like C language
- Shortage of Lisp programmers in industry
Symbolics Internal Struggles and Bankruptcy
- Internal divide between CEO Noftsker and COO Brian Seer:
- Noftsker: develop AI chip attachments for workstations
- Seer: pursue computer graphics market (Hollywood animation)
- Fiscal 1987: $11M profit became $25M loss
- High operating costs from boom period: expensive R&D, real estate
- 1986: signed 10-year leases on California and Massachusetts facilities
- Stock fell from $16 peak to under $1 by February 1988
- March 1988: TI-Apple partnership announced Micro Explorer
- Macintosh 2 with Lisp AI chip inside
- $15K price point devastated Symbolics
- Mass layoffs: 225 of 640 remaining workers (once had 1,000)
- 1993: Symbolics filed bankruptcy, liquidated assets
- Genera OS and first-ever registered domain name sold to new owners
LMI's Earlier Demise
- LMI never recovered from TI competition
- Attempted pivot to K Machine architecture
- Went bankrupt April 1987 before K Machine release
Legacy and Retrospective
- Lisp remains second-oldest programming language after Fortran
- Python became new AI language, Java absorbed Lisp's core ideas
- Former employees' testimonials:
- "Best development environment ever invented"
- Used Lisp machines "until PCs were 50,000 times faster and 100,000 times cheaper"
- Symbolics produced high-resolution display hardware, beautifully written Lisp OS
- Open-source Genera OS now runs in emulation on modern hardware
- Active hobbyist projects and emulators preserve the technology
