The History of AI Explained: Crash Course Futures of AI #1

Scientific concepts, discoveries, and nature/technology phenomena

Exponential growth in computing power (historical trend)

Computing performance has grown rapidly over decades, evolving from room-sized computers to ubiquitous personal devices with vast access to information.

Moore’s Law (semiconductor scaling)

• Prediction (Gordon Moore, 1965): transistor counts on chips would double about every 2 years.
• Outcome: the trend held for roughly 50 years, enabling major leaps in computational capability.

Artificial Intelligence as a spectrum

AI can be seen as a range of advanced computing systems that emulate intelligent behavior, including:

• Narrow AI: task-specific systems.
• General-purpose / general AI: systems intended to handle many tasks.

Narrow AI via chess and early programs

Chess programs illustrate how “intelligent” behavior can emerge from specialized methods.

• IBM Bernstein chess program (1957)

  • Example of narrow AI using algorithms that mimic human strategies.
  • Uses board evaluation, plausible move selection, and simulation/search.
  • Performance characteristics: slow move generation (minutes per move).

• Kaisa (1970s, Russian computer chess)

  • World champion-level computer chess engine.
  • Still lost to human champion David Levy.

• Gary Kasparov vs. 32 bots (1980s)

  • Highlighted strong human advantage over earlier AI chess engines.

• IBM Deep Blue (1997)

  • Symbolic AI approach:
    • Logic and decision-making encoded directly (hard-coded rules/structures).
    • Uses search and evaluation, not learning.
  • Reported capability: evaluating ~200 million positions/second.
  • Milestone: first machine to beat a human world chess champion (Kasparov) in a match.

Deep learning and neural networks (shift away from pure symbolic AI)

• Neural networks: brain-like architectures with layers of connected nodes and weighted pathways.
• Deep learning: learning from experience/data by adjusting internal parameters.
• Key ingredients (as described):
  • Data (information available for learning)
  • Algorithms (method for learning)
  • Compute (processing power, memory, storage)

Neural-network chess breakthrough: Stockfish (2020)

• Stockfish introduced an “efficiently updatable neural network (NNUE)” approach.
• Core advantage: incremental updates instead of re-evaluating the entire board each move.
• Result: improved efficiency, enabling deeper thinking per move.
• Impact: since 2020, it reportedly dominates computer chess championships.

Transformers and rapid sequence processing (general-purpose AI)

• Transformer architectures enable models to process entire sequences in parallel rather than strictly word-by-word.
• This accelerates improvements across many domains (e.g., language and images).

AI benchmarks (measuring capability)

• Benchmarks: standardized tests used to measure model performance.
• For general systems, benchmarks span multiple domains (e.g., language tasks, driving tasks, academic-style questions).
• Benchmark saturation: when models reach high performance and further improvements become extremely rapid.

Scaling laws (predicting future AI capability)

• Observation: AI performance tends to correlate with:
  • More data
  • More compute (processing resources)
  • plus algorithmic improvements
• Scaling laws: empirical formulas relating AI performance to data/compute scaling.
• Claimed uses:
  • Predicting what future larger models might achieve
  • Informing expectations about future capability and societal risk

Caveats/limits mentioned

• Imperfect fit: not all scaling matches expectations (example: DeepSeek competing with models using less compute).
• Diminishing returns: additional compute in pre-training may yield less benefit than earlier gains.
• Uncertainty: scaling laws may eventually encounter bottlenecks.

Speculative implications

• Potential future breakthroughs suggested:
  • Discovering new cancer treatments
  • Achieving superintelligence
• Risk framing:
  • Advanced AI could become harmful or uncontrollable (e.g., an “AI apocalypse” scenario).

---

Methodologies / frameworks outlined

Symbolic AI chess methodology (Deep Blue / early bots)

• Hard-coded logic based on human-style reasoning
• Search + evaluation of future moves
• No learning/improvement during operation

Deep learning training setup (general)

• Gather data
• Use a learning algorithm
• Provide sufficient compute (CPU/GPU resources, memory, storage)
• Train neural networks to update internal weights from experience

Neural-network inference improvement (Stockfish NNUE concept)

• Use neural evaluation
• Incrementally update evaluation after each move
• Reallocate compute to deeper search/analysis

AI capability measurement via benchmarks

• Use standardized tasks across domains
• Compare performance across model families
• Track saturation levels over time

Predicting capability via scaling laws

• Identify performance patterns vs. data and compute
• Fit formulas (“scaling laws”) to extrapolate to larger future systems

---

Researchers or sources featured (mentioned by name)

• Gordon Moore
• David Levy
• Gary Kasparov
• IBM (as the source/creator of Bernstein and Deep Blue)
• Kaisa (named chess program/engine)
• DeepSeek (DeepSk) (as a model/company example)
• Crash Course Futures of AI (series source)
• Future of Life Institute (production partnership)