Daejeon, July 27–31
Introduction
How does information infrastructure shape long-term cultural evolution? The availability, bandwidth, and structure of social information depend on material technologies and social institutions that can change rapidly. (Aker, Mbiti 2010; Valverde 2016)The current shift toward AI-mediated, online knowledge makes it urgent to examine how changes in "infostructure" reshape cultural evolutionary dynamics over long time spans.
This study uses the game of Baduk (also known as Go or Weiqi) as a historical observatory for such dynamics. Baduk is a complex, open-ended strategic system with a curated record of professional play spanning multiple centuries, and it recently experienced a major infostructural shock with the arrival of superhuman AI (SAI) programs after 2016 (Shin, et al. 2023). Using four centuries of professional game records from the GoGoD database (1600–2024; >118,000 human games after exclusions), I examine how opening strategies change across historically defined eras, and how those changes relate to population size and the structure of interaction networks among players. The primary outcome is the distribution of the first two moves (Black’s first move and White’s response), which compactly indexes "families" of openings; deeper opening sequences are used to validate this representation and visualize branching diversification.
The analysis builds on recent work on *information architecture* (Smaldino, et al. 2025) but uses the term "infostructure" to emphasize that these systems are not coherently designed: they are partly evolved, partly institutional, and partly technological. Infostructures shape (1) who can learn from whom, (2) how much information can be transmitted, and (3) how information is filtered or regulated. Baduk’s infostructure plausibly changed across feudal fragmentation, professionalization and mass print, mid-20th-century geopolitical disruption, international tournament circuits, the internet, and superhuman AI tutoring. This provides a long-run window into how opening diversity and turnover vary across eras, how population size and network structure shape strategic diversity, and whether SAI produced a lasting "revolution" or only a short-lived perturbation.
Methods
Getting game data
Game records were drawn from the July 2024 GoGoD database and processed in R. Records were cleaned to remove duplicates and mis-specified entries. Handicap games, games lacking valid dates, games prior to 1600, and games involving SAI players were excluded to focus on human play in the Early Modern period onward. Because Baduk openings are symmetric under rotation and reflection, all games were standardized so Black’s first move falls in the same board quadrant, enabling direct aggregation of opening distributions.
Temporal coverage is highly uneven: early centuries contain few recorded games per year, while post-1945 records expand sharply in both volume and international representation. To reduce sample-size artifacts, diversity and divergence statistics were computed via bootstrapping within time bins, repeatedly subsampling a fixed number of games per bin and averaging estimates.
Characterizing different eras of social learning in Baduk
To represent historically distinct infostructures, I partition the record into six eras tied to major historical and institutional transitions:
1. Early Modern (1600–1867): feudal patronage, limited travel and contact; hereditary schools and elite play.
2. Imperial (1868–1945): industrialization and professionalization; major institutions; print dissemination; rule experimentation including komi.
3. Cold War (1946–1972): postwar reconstruction; growing broadcast and publishing; geopolitical divisions; limited Chinese participation during the Cultural Revolution; rise of South Korea.
4. International (1973–1991): cross-national tournaments and increasing connectivity.
5. Internet (1992–2015): online servers and digital archives; expanded access to records and analysis.
6. Superhuman AI (2016–2024): widespread SAI tutoring following AlphaGo.
These eras provide a scaffold for comparing diversity, turnover, tempo, and network structure under different infostructural conditions.
In Baduk, the first few moves structure the game
Similarity between opening sequences was measured using edit distance on strings of move coordinates. A multidimensional scaling (MDS) analysis of sampled games shows strong clustering by the first and second moves: major regions in latent space correspond to specific Black–White opening pairs, with subclusters reflecting third-move choices. This demonstrates that the first two moves reliably index opening families while allowing meaningful follow-up diversification.
Quantifying cultural diversity
Opening diversity is quantified using Hill numbers, focusing on order-one diversity (the exponentiated Shannon entropy, exp(H')), interpretable as the effective number of equally frequent opening variants. To quantify change over time, Jensen–Shannon divergence (JSD) is computed between opening distributions in successive time bins, capturing both turnover and reweighting among existing variants. Both measures are estimated with bootstrap subsampling to control for changing sample sizes.
Analysis
Tracking opening diversity through time
Opening strategies exhibit substantial long-run turnover (Figure 1). Early records include openings that persist for centuries and later become rare or extinct, while some patterns vanish and reappear with new continuations. A sharp diversification occurs in the Imperial Era, consistent with historical accounts of "Shin Fuseki" experimentation in the early 20th century, during which many novel opening pairs rose in frequency.
The mid-20th century shows contraction and extinction of many experimental branches, coinciding with war, reconstruction, and political disruption. In later decades, diversity rises into the International Era while year-to-year divergence declines, indicating more incremental change within established families rather than wholesale replacement.
During the Internet Era, diversity falls again to historically low levels despite a growing and more connected player population, with openings concentrating into a small number of canonical choices. After the arrival of SAI, diversity increases modestly and JSD spikes around 2018, but this disruption is brief; divergence soon returns toward pre-AI levels. Similar short-lived effects appear deeper in the opening, suggesting transient adjustment rather than permanent restructuring.
Figure 1 - (top) Opening variant frequencies over time (middle) Opening move Shannon diversity over time (bottom) Opening move Jensen-Shannon Divergence over time
Diversification of families of openings
Decision trees of openings to seven moves show that some major families maintain stable continuation diversity across eras, while others expand during the Internet Era and contract in the SAI Era. Many early-era continuations are extinct. While SAI coincides with some revival of rare branches and reweighting among continuations, it does not produce diversification comparable to the Imperial Era; the major families established in the 20th century remain dominant.
Collective exploration of the strategic landscape of Baduk
MDS representations of longer opening sequences show constrained exploration in the Early Modern period, expansion into new regions during the Imperial and later eras, and consolidation during the Internet Era as previously explored regions empty out. Post-2016 play shows small expansions within existing regions but no formation of new clusters, suggesting that the macro-structure of openings stabilized in the 20th century and that recent dynamics involve faster cycling within this structure.
The pace of opening evolution
Since 1945, the magnitude of annual frequency fluctuations increases with the advent of the internet and peaks briefly after AlphaGo, then slows in recent years (Figure 2). This distinguishes diversity from tempo: even when few openings dominate, frequencies can oscillate rapidly among them.
Figure 2 - Magnitudes (in standard deviations) of the fluctuations in opening move frequencies since 1945, spanning the Cold War, International, Internet and SAI Eras
Opening diversity and population structure
Match networks reveal fragmentation in early periods, increasing connectivity with professionalization and internationalization, and high global connectivity in recent decades (Figure 3). Across periods, opening diversity follows an inverse-U relationship with player population size, peaking at intermediate sizes and reaching its lowest levels when populations are largest. Community detection shows a parallel shift from many smaller communities during high-diversity periods to fewer, larger communities in recent eras, consistent with theories predicting rapid convergence in highly connected networks.
Figure 3 - (left) Shannon diversity of the first two moves and number of players for defined periods. (right) Player network latent group size and count by community detection on the main component of each time period's match network.
Discussion
These results identify long-run phases of collective innovation and homogenization in Baduk openings that align with major infostructural transitions: a professionalization-driven burst of novelty in the early 20th century, mid-century contraction under geopolitical upheaval, late-century stabilization, and recent high-tempo cycling under digital connectivity.
The impact of superhuman AI on opening theory is relatively small
Despite the appearance of an AI-driven revolution in Baduk, the long-run record suggests that SAI has had a modest and transient effect on opening distributions. The post-2016 spike in divergence is brief and small compared to earlier upheavals. SAI mainly reshapes evaluations and accelerates adjustment within established families rather than generating new ones.
Contemporary openings are human–SAI convergent
Because modern Baduk AIs can learn without human training data, their convergence with human opening families suggests that centuries of cumulative human search captured robust strategic structure. At the same time, this convergence raises historical questions about why those families stabilized when they did, and why earlier eras invested heavily in now-extinct regions of strategy space.
Group size and structure mediate cultural diversity
The inverse-U relationship between diversity and population size, together with the association between high diversity and many smaller communities, supports the view that network structure mediates the maintenance of variation. Intermediate fragmentation appears to sustain exploration without forcing rapid global convergence.
Sabermetrics, aesthetics, and evolving perceptions of the game
Finally, Baduk mirrors broader data-driven transformations in cultural and athletic domains: optimization and improved performance can coincide with reduced stylistic diversity and changing aesthetic values. In the SAI era, evaluation is increasingly shaped by non-human agents optimized purely to win, potentially reshaping what players consider elegant or meaningful. Baduk’s long record offers a concrete case for digital humanities research on how infrastructures of information and evaluation co-produce cultural change.