We Trained a Baby AI on Shakespeare.
Here's What It Taught Us About AI Provenance.

Most AI models meet the world only after they've become impressive. By then, the awkward beginnings are gone. Early mistakes have been overwritten. Training decisions are buried. What remains is a polished system, a benchmark score, and a story told after the outcome is already known.

Baby Shakespeare starts somewhere different.

At the beginning.

Meet Baby Shakespeare — a 117,888-parameter character-level AI model trained from scratch on The Complete Works of William Shakespeare. Pure NumPy. One CPU. Eight minutes of training.

Today, he mostly speaks nonsense.

And that is exactly the point.

He has begun to recognize character names, dialogue formatting, punctuation rhythms, recurring fragments, and pieces of the statistical structure of Elizabethan English. He has not learned to write like Shakespeare.

His attention weights were intentionally frozen at random initialization, so the fairest description is also the least glamorous: a character-frequency model inside a transformer-shaped architecture.

Modest by design.

Because this experiment was never really about building an impressive language model in eight minutes. It was about asking a more important question:

Can someone who was not in the room independently reconstruct what happened — without simply trusting us?

In artificial intelligence, we spend a great deal of time talking about scale, capability, fluency, and benchmarks. We spend far less time asking about lineage.

What data was used? What was the training intent? Which artifacts existed before training began? What changed during the training run? What was preserved? What is missing? And what can another person independently verify?

For Baby Shakespeare, we preserved the full declared lineage:

Each preserved artifact carries a SHA-256 fingerprint. The corpus catalog and provenance records are permanently preserved on Arweave. Bitcoin reference metadata and Bitcoin commitment status are disclosed separately — not blended together or presented as proof of something they cannot establish.

The goal is not to manufacture certainty.

The goal is to create an inspectable record.

Cryptography cannot make a weak model strong. Preservation cannot turn a claim into truth.

Provenance establishes something narrower — and essential: what existed, what was declared, what changed, what evidence accompanied it, and what another party can independently verify.

He is a baby.

He babbles — which is exactly what we should expect from a tiny character-level model trained for eight minutes with frozen attention. He knows something about the shape of Shakespeare before he knows Shakespeare.

That is not literary achievement. It is a documented developmental state.

Baby Shakespeare is not remarkable because he is intelligent. He is remarkable because his development is inspectable.

The declared corpus can be inspected. The hashes can be checked. The manifest, weights, certificate, and declared limitations can be reviewed. No account is required. Independent reproduction remains open.

Baby Shakespeare will never again be exactly the model he was at eight minutes old. Future versions may improve. They may produce better language. They may make different mistakes. They may reach new milestones.

But this version does not have to disappear.

Most AI history begins when a system becomes successful. We believe history should begin when something begins. Every consequential AI model deserves a documented beginning.

Because the future of responsible AI will not only depend on what models can do. It will also depend on whether we can understand how they came to be.

Be kind. He'll never be eight minutes old again.