Deep Blue vs Kasparov
Context: FIT1061_MOC · the W2 milestone and Kialo motion — the first machine win judged by result, argued over by method · the case that stresses Turing Test (Imitation Game) and The AI Effect (Defining AI) What it weighs: whether a victory counts as intelligence because of what it achieved or because of how it was produced.
Quick Revision
- 🎯 Test question: When an AI wins, does it matter how? — the – word reflective response, and the Kialo motion “Deep Blue’s win tells us nothing about intelligence.”
- ⚠️ Key Constraint: a verdict without the machinery named earns nothing ➔ say brute-force tree search over positions/s with a hand-tuned evaluator, then judge that, not “a computer”.
📝 Core Commitments
- The facts you argue from ➔ 1997-05-11, Game 6: Kasparov resigns after moves, match lost –. IBM, over a decade, \approx\1030+480\approx2\times10^8$ positions/s.
- The mechanism ➔ no intuition, no experience, no learning ➔ it searched (Search Problem Formulation), evaluated the leaves, picked the best reachable outcome. Same skeleton as Uninformed Search (BFS and DFS), plus a heuristic evaluator and pruning.
- Behavioural rule ➔ judge only the observable output; the machinery is off-limits by design — Turing’s whole move was to replace “can it think” with a performance test.
- Mechanist rule ➔ judge the process; a result produced by exhaustive enumeration is a statement about hardware and search, not about understanding.
- What each refuses to count ➔ the behavioural rule refuses “but it’s only search” as evidence; the mechanist rule refuses the scoreboard as evidence.
- The deflation trap ➔ the moment chess fell, chess stopped counting as intelligence (The AI Effect (Defining AI)) ➔ any purely mechanist verdict must answer why the goalpost moved after the result.
- Hsu’s framing ➔ “The computer doesn’t play chess like a human, any more than an airplane flies like a bird.” ➔ concedes the mechanism differs and denies that this diminishes the achievement.
⚖️ Framework Contrast
| Position | Test applied to Deep Blue | Verdict it reaches | Where it breaks down |
|---|---|---|---|
| Behavioural (Turing’s substitution) | did it produce grandmaster-beating play under fair conditions? | intelligent — the output is what was ever testable | licenses a lookup table of every position as “intelligent” |
| Mechanist (“only search”) | what machinery produced the moves? | not intelligence — enumeration plus an evaluator | the same objection dissolves human cognition into neurochemistry |
| AI-effect deflation | is the task still hard now? | not intelligence — chess is solved-ish, so it never counted | unfalsifiable: no achievement can ever survive the test |
🧩 Case Application Drill
Case 1: The Kialo motion — "Deep Blue's win tells us nothing about intelligence."
Model verdict
- Duty-holder / claim-maker: IBM, which staged the match as a scientific milestone and marketed the result.
- Framework applied: the mechanist rule fires cleanly on this machinery — depth-limited search plus a hand-tuned evaluator, zero learning, zero transfer; the same program cannot play draughts, let alone route a car.
- Verdict: the win tells us little about general intelligence and a great deal about search: that a well-pruned tree plus positions/s outperforms the best human pattern recognition in a bounded, perfect-information domain. The strongest counter is behavioural — that “it’s only search” is exactly the objection Turing pre-empted, since we grant humans intelligence without inspecting their machinery. Answer it by bounding the claim to general intelligence rather than denying the achievement.
- Debate note: W2 Kialo — citing is optional this week (expected in W7 and W11); the graded move is prompt → response → counter-argument → reply, so the counter-position must be engaged, not listed.
Case 2: Game 2, move 36 (Be4) — the move Kasparov could not explain, that an IBM engineer later attributed to a software bug producing a random move when the machine could not decide.
Model verdict
- Stakeholders: Kasparov (who read creativity into it and alleged human intervention), IBM (which refused a rematch and dismantled the machine), and everyone downstream who cannot now check.
- Framework applied: the behavioural rule is falsified by its own success here — the move looked intelligent precisely because the observer could not see the mechanism, and the mechanism turned out to be a coin flip.
- Verdict: attributing intelligence from behaviour alone is unsafe when the observer cannot audit the process; the case is the strongest available argument that how matters. Note the asymmetry that keeps it honest: a bug explains one anomalous move, not a – match.
- What it costs: IBM dismantled Deep Blue and split the hardware between the Smithsonian and the Computer History Museum ➔ the experiment is unrepeatable, which is the reproducibility failure the unit’s how does it work refrain exists to prevent.
Case 3: In 2026 a phone runs Stockfish at nodes/s, Elo , and beats Deep Blue (Elo ) comfortably — on fewer positions per second.
Model verdict
- Framework applied: the mechanist rule, turned productively — compare the machinery, not the scoreboard.
- Verdict: hardware shrank but that is not why the phone is stronger; it searches fewer positions and wins, so the gain is algorithmic — better pruning, better evaluation, better move ordering. The 1997 result was a hardware milestone; the 2026 result is the argument that strategy beats scale, which is exactly why the unit spends W3–W5 on heuristics rather than on faster loops.
⚠️ Common Mistakes
- 💡 Retelling the match ➔ the dates, the moves and the accusation are evidence, not an argument; every fact cited must be attached to a claim about intelligence.
- 💡 Asserting a verdict with no rival named ➔ the response is graded on engaging the counter-position; “it was just brute force” without the Turing objection answered is half an argument.
- 💡 Sliding the goalposts silently ➔ if the verdict is “chess never counted”, say so and defend it against The AI Effect (Defining AI) — otherwise the position is unfalsifiable.