Beyond Optimality: Lessons from Shin Jin-seo's Victory over KataGo

Shin Jin-seo's recent victory  (2026-07-21) over KataGo demonstrates a subtle but profound distinction between optimal decision making and winning against a particular opponent. The result suggests that maximizing an objective function is not always equivalent to maximizing the probability of defeating an adversary whose own decision process is constrained by optimization.

One moment in the third and final game was particularly striking. Shin intentionally accepted a slight local loss in the corner, but in return secured outside thickness by steering the game into an essentially single-path sequence that even KataGo was forced to follow.  Although this exchange was locally suboptimal, it significantly simplified the subsequent game and ultimately became one of the decisive factors in his victory. 

KataGo is designed to maximize its expected winning probability. Formally, it selects an action $a\in A$ by solving $a^*=\arg\max_{a\in A}\mathbb{E}[U(a)]$, where $U(a)$ denotes the utility (or winning probability) associated with action $a$, and $\mathbb{E}[.]$ denotes the expectation with respect to the uncertainty represented in the AI's evaluation model, including future game evolution. Every move is chosen because it maximizes the expected value of the objective function.

Consequently,  KataGo has little incentive to choose moves that intentionally sacrifice objective value in order to induce errors by its opponent. Such moves are generally excluded from the optimization process because they reduce the immediate expected utility. In other words, KataGo is designed to optimize its objective function, not to employ strategic deception.

Shin Jin-seo appears to have understood these characteristics remarkably well. Rather than engaging in the highly tactical and calculation-intensive battles where KataGo is strongest, he deliberately guided the game toward positions with low combinatorial complexity. More importantly, he recognized that, because KataGo is designed to maximize expected utility, it rarely accepts temporary losses solely to increase the probability of inducing errors by its opponent. While AI is highly effective at exploiting mistakes once they occur, it is generally not optimized to create conditions that make such mistakes more likely. By intentionally favoring relatively forced, one-way sequences—even at the cost of a slight territorial disadvantage—Shin prevented KataGo from fully exploiting its overwhelming tactical search capabilities. Since he already held an advantage of approximately 18 points, his strategic objective was to steer the game toward a stable and essentially deterministic continuation that would quickly and decisively resolve the game, rather than allowing it to develop into complex tactical fighting where KataGo's computational strength would be most effective.

Human experts, however, solve a different optimization problem. When faced with a losing position, they often abandon the objective of minimizing the current deficit and instead seek to maximize the probability that the opponent will make a mistake. This frequently requires introducing tactical complications or strategic imbalances that increase the variance of the game. Although such decisions may further reduce the objective evaluation if the opponent responds correctly, they simultaneously increase the opportunity for human error and may therefore improve the practical probability of victory. From this perspective, the optimization objective can be formulated as $a^\dagger=\arg\max_{a\in A}P(\text{opponent makes an error}\mid a),$ or more generally, $a^\dagger=\arg\max_{a\in A}\mathbb{E}[U(a,\pi_{\mathrm{opp}})],$ where $\pi_{\mathrm{opp}}$ denotes the actual behavioral policy of the opponent rather than an ideal rational agent. The optimal solution therefore depends not only on the game itself but also on the opponent's decision mechanism.

The essential insight is that KataGo optimizes over the game, whereas Shin Jin-seo optimized over KataGo. In effect, he exploited the constraints imposed by the AI's own optimization principle.

This distinction can be expressed mathematically. Let $\pi^*=\arg\max_{\pi}\mathbb{E}[U(\pi)]$ be the globally optimal policy. Against a specific opponent with policy $\pi_{\mathrm{opp}}$, however, the policy that maximizes the probability of victory is$\pi_{\mathrm{exploit}}=\arg\max_{\pi}\mathbb{E}[U(\pi,\pi_{\mathrm{opp}})]$. In general, $\pi_{\mathrm{exploit}}\neq\pi^*$.Th is inequality represents the fundamental distinction between optimization and exploitation. A globally optimal strategy need not be the strategy that best exploits a particular adversary.

The broader implication extends far beyond the game of Go. Most contemporary AI systems are designed to optimize explicitly defined objective functions. Unless the objective itself incorporates opponent modeling, strategic deception, or information asymmetry, these behaviors emerge only indirectly, if at all. Consequently, there exists a class of adversarial environments in which a human expert can obtain an advantage not by exceeding the AI's computational capability, but by understanding the limitations imposed by its optimization criterion.

Future AI systems may overcome this limitation by explicitly incorporating opponent modeling into the optimization process. Rather than solving only $\max_{\pi}\mathbb{E}[U(\pi)],$ they may solve $\pi^{**}=\arg\max_{\pi}\mathbb{E}[U(\pi,\hat{\pi}_{\mathrm{opp}}(\pi))],$ where $\hat{\pi}_{\mathrm{opp}}$ denotes the predicted behavioral response of the opponent. Such systems would no longer compute merely the optimal solution to the game; they would compute the optimal strategy against a particular intelligence. At that point, AI would transition from pure optimization to adaptive strategic reasoning, fundamentally altering the nature of competition between humans and machines.