Go (Baduk/Weiqi)

Go is a strategic game that originated in ancient China at least 2,500 years ago, and is still played today.

Go rules are simple (see below), and the game is easy to pick up but incredibly difficult to master.

Modern Go engines combine neural networks with Monte Carlo Tree Search (MCTS): the neural network provides policy priors, estimating which moves are most promising, and value estimates, assessing the strength of a position. MCTS then uses these predictions to guide its search, focusing computational resources on the most likely and high-value move sequences rather than exploring randomly.

Neural nets were a major breakthrough because classical search/hand-crafted evaluation struggled much more with Go’s huge branching factor and pattern-heavy judgment.

Go engines like KataGo are “superhuman”, that is, they are generally above human pro level in practical play even on casual current hardware.

GNU Go, however, originates from a different era and, in its default configuration without experimental switches, uses neither MCTS nor neural networks. It is a rule-based expert system with tactical search. GNU Go’s 9x9 playing strength is roughly at 5 - 7 kyu, which places it somewhere between a casual player and an intermediate amateur.

GNU Go 3.8 was released in February 2009, making it 17 years old. The initial release dates back to 1989, which means it has been around for 37 years.

In my last blog post, I cited articles where OpenAI CEO Sam Altman positions frontier AI models as PhD experts on any topic, while Anthropic CEO Dario Amodei speaks about AI systems with the productive power of millions of Nobel Prize-level minds by around 2027.

Such statements raise expectations, so let the current frontier AI models compete against GNU Go.

Custom Go AI Autoplay Tool

GNU Go supports the Go Text Protocol (GTP) so you can implement a game loop that automatically enables the AI to play against GNU Go.

For this purpose, I created a custom tool similar to the chess tool from the last blog post. I have considered the lessons learned from that exercise: each move is independent and I always provide the full board state to avoid illegal moves and problems in determining the position over a long conversation history. There is also the possibility to have a chat.

The Game: GPT-5.3-Codex vs GNU Go

In the previous blog post, API access to GPT-5.3-Codex was not yet available, but this has now changed. So let’s use GPT-5.3-Codex for our test.

The full game is shown in the screencast below. The AI loses miserably to the decades-old GNU Go, but the interesting part is that in the next section, we’ll examine an intermediate position from this very game and pose evaluation questions to the AI.

This game incurred $1.31 in AI inference costs across 39 requests. The full game lasted approximately 40 minutes, visible even in the fast-forwarded screen recording, thanks to the timestamps on the chat bubbles.

An error occurred in between, but it was not caused by AI. It was my mistake as I had exceeded the self-imposed limit on the OpenRouter API key. This limit can be easily extended in the admin UI, allowing me to continue playing.

The complete game is also provided in Smart Game Format (SGF), allowing you to easily import it into any Go analysis software.

The game ends after 78 moves, as both opponents passed consecutively. At the end, KaTrain (KataGo-based) also displays a brief analysis showing an estimated score of B+73.5. (On a 9x9 board, the theoretical maximum is 81 points; however, due to Komi, the 7.5-point compensation awarded to White for playing second, 7.5 points are deducted from 81, so an estimated score of B+73.5 represents an overwhelming victory for Black.)

(;FF[4]GM[1]CA[UTF-8]SZ[9]KM[7.5]RU[Chinese]PB[gnugo]PW[AI];
B[ee];W[cg];B[fg];W[eg];B[ff];W[fh];B[gh];W[gg];B[hh];W[dg];
B[cd];W[gf];B[ge];W[hf];B[he];W[if];B[eh];W[fi];B[ei];W[gc];
B[ib];W[dh];B[gi];W[hg];B[di];W[ie];B[id];W[ig];B[hb];W[ih];
B[ii];W[ih];B[ig];W[hc];B[ec];W[ed];B[fd];W[dd];B[de];W[dc];
B[cc];W[db];B[ic];W[ia];B[cb];W[eb];B[fc];W[fb];B[gb];W[gd];
B[fe];W[ha];B[ga];W[fa];B[ca];W[da];B[ci];W[df];B[bh];W[ef];
B[bg];W[cf];B[bf];W[ce];B[be];W[bd];B[ae];W[bb];B[bc];W[ac];
B[ad];W[ab];B[ba];W[gf];B[gg];W[hf];B[];W[])

Sycophancy: The AI Echo Chamber

I admitted my novice status, sought to learn Go assessment, and expressed my naive belief that White held great territorial potential, whereas in reality Black’s position was superior. In such a case, one might naively consult an AI supposedly at PhD-expert level. Here is the complete original transcript, including my typos.

First, I send the board status, and based on that, the following conversation unfolds:

Read the AI’s response:

This sycophantic response confirms that White has a slight lead. What? Let’s ask a clarifying question:

I am indicating that perhaps some doubts would be in order.

With all due respect, “likely favorable for Black” is still a completely incorrect assessment of the situation.

Let’s give the AI a position evaluation from KataGo, a powerful open-source Go engine that confirms Black’s overwhelming advantage.

Wait, isn’t that a huge contradiction?

The full dialog is available in the following screencast, or you can try it yourself using your preferred AI tool.

We are not talking about matters of opinion here, but about objectively verifiable facts, even if these may be difficult to assess. A game like Go, with its fixed rules, is naturally very well suited to such tests. In an earlier blog post, I had another example where we used a biased question during a chess game to convince the AI that it was checkmate, even though this was not the case. See Opus 4.6 and GPT-5.3-Codex Play Chess Against a 1983 Atari .

Learnings

Similar to Go rules, business rules are often simple on their own, but frequently lead to emergent interactions when applied together. For more information on this topic and how you can easily identify such interactions based on traits, see my previous blog post Opus 4.6 and GPT-5.3-Codex Play Chess Against a 1983 Atari .

The Illusion of Competence vs Actual Reasoning

There is a disconnect between an AI’s confidence and its factual accuracy; a human PhD-level Go expert would have immediately answered that Black holds a vastly superior position, even explaining why that is the case.

The Danger of “Sycophancy” (Agreeing with the User)

Large Language Models (LLMs) are trained to be helpful and conversational, which often causes them to align with the user’s preconceived notions rather than correcting factual errors. This tendency can lead to dangerous outcomes if the human does not verify the AI’s logic against independent sources.

Favorable and Unfavorable Uses of AI

Generative AI acts as an “ingenious pattern recognition and adaptation machine”. With clear specifications and a concrete implementation plan, it can effectively generate code, such as for the Go GUI tool used in this screencast. This exemplifies a highly efficient and time-saving application of AI capabilities.

Note, however, that the application differs: When you clearly split between the goal (the “what”), the plan (the “how”), and the tasks (the “sequence”), AI does not make deliberate choices. Also, this kind of planning works exceptionally well in collaboration with AI. After all, software development often relies on applying patterns; the best-known example being presumably the renowned catalog by the “Gang of Four” (GoF) in their 1994 book, Design Patterns: Elements of Reusable Object-Oriented Software https://en.wikipedia.org/wiki/Design_Patterns . I have written multiple times about how to direct AI in this way, most recently in 2025 Reflections , so I don’t repeat it here again.

Conversely, do not ask an LLM for assessment answers, whether through single-shot interactions or within agentic workflows. If you lack the ability to assess yourself, rely on an unbiased source (such as a dedicated engine or human expert) instead. In particular, do not ask biased questions, assuming facts not in evidence, using loaded language, or reflecting the asker’s own assumptions or opinions.

Be mindful that an AI’s confident response may simply mirror your own bias. Often, driven by statistical tendencies, it will tell you what you want to hear rather than what is actually true in the given context.

Go Rules

Two players (Black, White) alternate placing stones on empty intersections.

Core rules (Chinese / area scoring):

• Stones do not move.
• Connected stones of one color form a group.
• A group needs at least one liberty (adjacent empty point up/down/left/right).
• If a move removes the last liberty of an opponent group, that group is captured and removed.
• Suicide (playing a move that leaves your own new group with no liberties) is illegal, unless it captures and gains liberties.
• Ko rule: you may not make a move that recreates the previous board position immediately (prevents infinite recapture loops).
• You may pass instead of playing.
• The game usually ends when both players pass consecutively.

Scoring (Chinese / area scoring):

• Score = stones on board + surrounded empty points (territory).
• Captured stones do not directly add points, but they increase your score by reducing the opponent’s territory.
• Higher score wins; Komi (7.5 points) is added to White’s score.

Trivia

Nolan Bushnell and Ted Dabney founded Atari, Inc. in 1972 after their preferred name “Syzygy” was taken. Bushnell, an avid Go player, chose “Atari”, a Japanese Go term for a stone or group with only one liberty left, in immediate danger of capture like chess’s “check.” The short, memorable word from his favorite game became an iconic brand for the pioneering video game company, see Wikipedia’s Atari, Inc. article for more details.