
Golden Board is a fascinating experiment in self-describing information, mathematical communication, and AI evaluation. It also exposes a surprisingly difficult problem: decoding a message is not the same as understanding what it means.
Imagine receiving a mysterious binary message from another civilisation.
No documentation.
No English.
No shared alphabet.
No helpful little README explaining that you should install Python before attempting to communicate with the inhabitants of another star system.
Just a sequence of zeros and ones.
Somewhere inside, apparently, is a complete introduction to chess.
The rules.
The pieces.
The legal moves.
Examples.
Exercises.
And 64 complete games.
Your task is to figure out how to read the message, discover what the game is, and learn how to play it.
This sounds like the beginning of an excellent science-fiction story.
It is also the premise of Golden Board, a project by developer Artem Legotin that recently appeared on Hacker News.
The project consists of a 512 KiB binary message designed to explain its own structure and then teach chess to a recipient who does not share our language or culture.
And while the interstellar angle is charming, the most interesting implications may be much closer to home.
Because Golden Board is really asking something fundamental about information and intelligence.
How much can you teach through structure alone?
And how do you know whether the recipient has actually learned anything?
Start With 4,194,304 Bits
Golden Board contains exactly 4,194,304 bits.
That is:
2²² = 4,194,304
Arrange them into a square, and you get:
2048 × 2048
The square is not merely a decorative illustration.
It provides a physical layout in which regularities can be discovered.
The first 64 rows contain 32 frames, each 64 × 64 cells.
Together, those frames form the bootstrap.
Their job is to teach the recipient how to interpret everything that follows.
This is the clever part.
Rather than beginning with chess pieces, Golden Board begins by establishing how information is represented.
First, counting.
Then binary numbers.
Then the relationship between groups of bits and numerical values.
Then structured data.
Then ways to verify that information has been decoded correctly.
Eventually, the message establishes enough machinery to reconstruct images and simple interactions.
Only then does it introduce chess.
It is rather like sending someone a book that first teaches them how to build the reading system required to read the book.
The difference is that this book does not begin with an alphabet.
It begins with mathematical relationships.
A Message That Teaches Its Own Decoder
Most digital information depends on conventions that already exist.
A JPEG file makes sense because software knows how JPEG works.
A JSON document is useful because the recipient knows its syntax and what the relevant fields mean.
A chess PGN file assumes that someone understands chess notation.
Even a plain text document presupposes a character encoding.
These conventions are usually invisible to us because our computers already contain the necessary machinery.
But remove that shared infrastructure, and a simple file becomes an arbitrary collection of bits.
Golden Board attempts to reconstruct part of that infrastructure from scratch.
Its bootstrap gradually introduces the operations required to read the message.
For example, it teaches elementary binary operations and then demonstrates a checksum calculation.
The chosen checksum is CRC-32C.
The recipient is not expected to recognise the name CRC-32C.
Instead, the message provides examples intended to let the recipient reconstruct the procedure.
Once the algorithm is understood, it can be used to check whether later sections have been recovered consistently.
This matters because a plausible-looking interpretation is not necessarily correct.
The recipient might read the bits in the wrong direction.
Swap zeros and ones.
Choose the wrong boundaries.
Or construct a convincing-looking image from an incorrect interpretation.
A checksum provides a way to reject inconsistent readings.
Not proof of meaning, but evidence that the numerical reconstruction matches the stored data.
That is an important distinction.
A correct checksum tells you that the decoding procedure has recovered the expected bits.
It does not tell you that the recipient understands chess.
Or why anyone would want to play it.
From Bits to Bishops
Once the representation has been established, the payload introduces the game itself.
The recipient learns about an 8 × 8 board.
Two sides.
Six kinds of pieces.
Starting positions.
Alternating turns.
Legal moves.
Captures.
Check.
Checkmate.
Draws.
And eventually more complicated rules and strategic relationships.
The material uses reconstructed images, examples, contrasts, exercises, and feedback.
There are also 64 complete games.
According to the project's documentation, the selected games span 1909 to 2021 and contain 6,890 individual moves, counting each player's turn separately.
The original human players are not identified in the binary message.
Neither are the tournaments, dates, or familiar opening names.
The intended recipient does not need to know who Garry Kasparov or Bobby Fischer was.
They need to recognise positions, moves, and the relationships between them.
Which is an elegant idea.
Chess is unusually well suited to this kind of experiment.
Its rules are finite.
Its board is small.
Its actions can be represented mathematically.
And its legal transitions can be checked.
Yet the space of possible play is enormous.
You can teach the formal rules relatively compactly, then introduce an extraordinarily rich world of strategies and ideas.
A message that teaches chess is therefore more than a collection of arbitrary examples.
It is an attempt to transmit a system of constraints from which the recipient can generate new possibilities.
That's considerably more interesting than transmitting a list of answers.
Why This Is More Than Data Compression
It would be easy to mistake Golden Board for an exercise in efficiently packing chess data into a small file.
But compression is not its central achievement.
In fact, much of the message is devoted to making information understandable rather than merely storing it compactly.
The bootstrap occupies about 3.1% of the total bit budget.
Lessons and exercises account for roughly 78.8%.
The complete games occupy about 17.7%.
The rest covers indexing, checks, and unused capacity.
A conventional chess archive would not need most of this instructional material.
A chess program already knows what a board is.
It already knows the rules.
It only needs the moves.
Golden Board assumes that the recipient may know none of those things.
So it spends information on explanation.
The design is not trying to minimise the number of bits required to store chess.
It is trying to reduce dependence on knowledge outside the message.
And that is a very different engineering objective.
The interesting quantity is not simply the length of the data.
It is the amount of prior agreement required before the data becomes meaningful.
The Problem With a Truly Universal Language
Here is where the project becomes philosophically interesting.
Golden Board does not actually assume nothing about the recipient.
The author is explicit about this.
The intended reader must recognise the material as a deliberate message.
It must be capable of counting, detecting patterns, reasoning mathematically, and constructing simple computational procedures.
It must also discover how the message's physical representation relates to its logical structure.
Those are substantial assumptions.
And there is nothing dishonest about making them.
A communication system without any shared assumptions would be an extraordinarily difficult thing to define, let alone build.
The real question is whether the assumptions are modest enough to be plausible across very different intelligences.
Mathematical regularities are promising candidates.
Counting, symmetry, repetition, and consistent relationships between quantities do not depend on English grammar.
But translating these relationships into a practical reading procedure still requires interpretation.
Consider a simple arrow.
To us, an arrow indicates direction.
But is that convention universal?
A Hacker News commenter raised precisely this objection.
Legotin's response was that the message does not require the recipient to know the arrow convention beforehand. Its meaning is intended to emerge from examples and repetition.
That is a reasonable design strategy.
But it also illustrates the underlying challenge.
Every apparent primitive must be examined.
Is the recipient discovering a concept from evidence?
Or recognising something familiar from prior experience?
This is not a minor problem.
It is the central problem of communication without shared conventions.
Golden Board makes progress by building an increasingly explicit chain of examples.
It does not abolish the need for inference.
Can an LLM Really Learn Chess From This?
The author used large language models to test the message.
The author reports 54 trials across different stages and conditions, including raw reconstruction, interpretation, guided work, and continued or assisted sessions. The detailed records include continuations and administrative or unrun checkpoints, so they should not be treated as 54 independent experiments.
Some experiments began with raw bits.
Others used information that had already been decoded.
Some allowed more guidance than others.
This distinction matters.
The most important question is not whether a capable model can correctly explain chess after encountering the file.
Modern language models already know chess.
They have encountered descriptions of the rules, board diagrams, games, chess notation, and discussions of strategy during training.
If a model sees an unfamiliar binary representation and successfully identifies chess positions, that is interesting.
But it does not prove the message taught the model chess from zero.
The model may have recognised familiar structures.
It may have used prior knowledge to fill in missing meaning.
Or it may have inferred the format by matching the data to patterns it already knew.
These are different achievements.
The author's documentation explicitly separates them.
One type of test asks whether the representation can be recovered from raw information.
Another asks whether the recovered material can be interpreted and applied.
A third concerns the operation of the supplied viewing and interaction machinery.
The detailed records also describe failures, partial results, corrected approaches, and cases where a successful decoder was tested on transformations of the input.
That degree of differentiation is valuable.
It is much more informative than simply announcing that an AI successfully decoded the transmission.
Decoding Is Not Understanding
Suppose a model reconstructs the exact message bytes.
Excellent.
That establishes a meaningful technical result.
Now suppose it correctly displays a chessboard.
Also excellent.
But what has it learned?
Can it identify legal moves?
Can it explain why one move is legal and another is not?
Can it respond correctly to a position it has never seen?
Can it distinguish checkmate from stalemate?
Can it apply rules that were taught through examples rather than through familiar chess vocabulary?
And most importantly, would it have been able to do any of this without its existing knowledge?
These are separate questions.
A system can reproduce a representation without understanding its subject.
A model can demonstrate subject knowledge without learning that knowledge from the supplied representation.
And a model can sometimes produce correct answers without a reliable internal procedure for applying the relevant rules.
This is one of the reasons Golden Board is interesting for AI research.
It provides an unusually explicit environment for separating stages that are often conflated.
Signal interpretation.
Data reconstruction.
Structural understanding.
Rule application.
Generalisation.
The author's public trial account supports bounded technical conclusions, including successful recovery and checking under documented conditions.
It does not establish universal comprehension, entirely unassisted learning, or generalisation to arbitrary unfamiliar formats.
That is a limitation.
But it is also a useful starting point for better experiments.
The Experiment I Would Like to See
The obvious next step would be to test the educational component more independently from existing chess knowledge.
For example, construct a new board game.
Not chess with renamed pieces.
A genuinely unfamiliar rule system with different movement constraints and victory conditions.
Encode it using the same bootstrap.
Then ask several models to reconstruct the representation and learn the rules without being told what the game is.
The evaluation should include new positions and tasks that cannot be solved by recalling familiar chess rules.
And the experiment should carefully separate information supplied at each stage.
Ideally, it would compare several conditions.
Raw message only.
Decoded material without explanatory guidance.
Guided material.
And a control condition without the relevant lessons.
This would help distinguish whether a model was genuinely using the message to learn a rule system or relying on prior associations.
It would not solve the problem of communicating with an alien intelligence.
But it would provide a stronger test of whether the protocol teaches unfamiliar structured knowledge.
A model that correctly applies rules for a newly designed game is more interesting evidence than a model that recognises a chessboard.
Especially if the model can also explain which examples establish which rules.
That would bring the experiment closer to its most ambitious claim.
Not merely transmitting information.
But transmitting enough structure for a recipient to construct usable knowledge.
Why This Matters for AI Agents
The extraterrestrial premise is delightful.
But there is an immediate engineering analogy.
AI systems increasingly exchange structured information with tools, databases, applications, and other agents.
Today, much of that communication relies on natural-language instructions, JSON schemas, API specifications, and established software protocols.
These approaches work.
But they also depend on shared conventions and documentation.
What happens when an agent encounters an unfamiliar tool or dataset?
Can the interface explain itself through machine-readable examples?
Can the recipient discover valid operations?
Can it verify that its interpretation is consistent?
Can a protocol teach the recipient how to use it without requiring a lengthy natural-language manual?
Golden Board offers a useful conceptual model for those questions.
Not because production agents should start exchanging enormous chess-like bitmaps.
That would probably be an exceptionally inefficient use of infrastructure.
But because it demonstrates an alternative design principle.
Instead of supplying only a schema, supply enough examples and checks that the recipient can reconstruct important aspects of the schema's meaning.
A machine-readable system could progressively explain its own structure.
Show valid operations.
Show invalid operations.
Provide expected outputs.
Require consistency checks.
And expose the difference between merely parsing a request and applying its semantics correctly.
There are practical limitations, of course.
A system that infers an interface incorrectly can produce harmful actions.
A checksum cannot validate the meaning of a financial transaction.
And examples do not automatically establish that a model has understood every edge case.
Critical systems still need explicit contracts, validation, authorisation, and test coverage.
But self-explaining structured messages could be useful in systems where agents encounter unfamiliar capabilities and must learn how to interact with them.
The interesting future is not necessarily AI agents communicating in mysterious binary telepathy.
It is interfaces that are easier for machines to inspect, verify, and learn without relying entirely on long prose instructions.
What Golden Board Does Not Prove
There are several reasons not to overstate the result.
First, successfully decoding a clean digital artifact is not equivalent to discovering an unknown signal from space.
The project starts with the recipient having access to the bits or their physical arrangement.
Detecting a transmission, recovering its boundaries, distinguishing signal from noise, and selecting a suitable interpretation are separate challenges.
Second, mathematics does not automatically provide a complete shared language.
Two intelligences might recognise the same numerical relationships without assigning the same significance to every visual convention or interaction.
Third, existing language models are not blank recipients.
Even a new chat begins with a model containing extensive prior knowledge.
Fourth, successful technical decoding does not establish genuine learning or retained understanding of the subject.
And finally, a project designed to be decoded is not necessarily a project that has been independently demonstrated to teach its subject to completely unfamiliar minds.
The author acknowledges these distinctions.
Another practical caveat concerns licensing.
The repository is publicly available and the project can be studied and adapted under its stated terms, but the original material is distributed under noncommercial licensing conditions.
It should not be described as unrestricted open-source software in the conventional licensing sense.
Commercial reuse requires attention to the project's permissions.
None of these limitations makes Golden Board uninteresting.
They define what the experiment actually demonstrates.
And separating an achievement from a larger ambition is essential if we want to learn anything useful from it.
The Most Interesting Part Is Not Chess
The project's inspiration comes from a tradition of attempts to communicate with unknown intelligences.
The Voyager Golden Record.
The Arecibo message.
Hans Freudenthal's Lincos.
Other efforts to construct messages that a sufficiently capable recipient might interpret without already understanding a human language.
But Golden Board makes an unusual choice.
Instead of beginning with scientific facts or historical information, it transmits a game.
And there is something appealing about that.
Science communicates what we have discovered.
A game can communicate something about how we think.
Chess contains rules, constraints, strategies, sacrifices, mistakes, and surprising combinations.
Its possibilities emerge from a relatively compact formal system.
To understand a chess game is not merely to identify the movement of pieces.
It is to recognise choices made within constraints.
And to discover that simple rules can generate situations of extraordinary complexity.
Of course, recognising the beauty of a sacrifice is not guaranteed simply by understanding the legal moves.
That is another layer of interpretation.
But this is precisely why the experiment is interesting.
The author is not trying to send a message that says only, Here are the rules of chess.
He is attempting to transmit enough of the game that another intelligence could explore it.
Perhaps eventually discover interesting positions of its own.
And perhaps recognise something in the decisions made by human players.
That is a much more ambitious form of communication than sending a factual statement.
It is an invitation to participate in a structured intellectual activity.
The Verdict
Golden Board is not proof that mathematics provides a universal language.
It is not proof that today's language models can learn chess without prior knowledge.
And it is certainly not evidence that an extraterrestrial civilisation would understand our favourite board game.
What it is, however, is an unusually thoughtful experiment in self-describing information.
Its most valuable feature is not the size of the binary file.
It is the attempt to make decoding, interpretation, verification, and learning separate problems with separate evidence.
The project also raises an important question about AI evaluation.
When a model produces the correct answer, how do we know which information it actually used?
Did it learn from the supplied material?
Did it reconstruct a representation?
Did it recognise something already familiar?
Or did it combine all three?
Golden Board does not completely resolve that problem.
But it makes the problem visible.
And that may be more useful than another demonstration in which an impressive AI model appears to understand something and nobody is quite sure why.
There is also a certain elegance in the author's choice of subject.
If we ever communicate with another intelligence, we might send mathematical constants, scientific discoveries, and descriptions of the physical world.
All sensible choices.
But perhaps one of the most human things we could send is a system of rules that allows someone else to play.
Not a declaration of who we are.
Not a demand that they understand our culture.
Just an invitation.
Here are sixty-four squares.
Here are the pieces.
Here is how they move.
Now see what you can discover.
And if the recipient eventually plays a move we never anticipated, that might be the most interesting reply of all.
Not because the message has transmitted a complete understanding of humanity.
But because it has created the conditions for another mind to do something new with an idea we shared.
That is a rather good ambition for 512 KiB.
And an even better one for AI.
