Sáastun AI – A New Breakthrough Language Model Capable of Deciphering Maya Glyphs

A Mexican-Canadian Engineer wows attendees at Mayanist conference in Mérida, Mexico, with his Sáastun AI language model

At the INAH Yucatán Ichkaantijoo Symposium, researcher Benjamín González showcased a groundbreaking artificial intelligence model called Sáastun, capable of converting ancient Maya to English.

While demoing Sáastun AI, González made it clear his work is still an early prototype, but left attendees at the conference impressed. The Sáastun AI model takes its name from a seeing stone used by Maya priests for divination. 

“I think everyone was a little sceptical about this AI model. Maya writing is famously complex, but while there is still obvious work to do, just how far it has already come is nothing short of amazing,” said archaeologist Zachary Lindsey.

González is not an epigrapher or archaeologist but rather an engineer born in Mexico, and living in Montreal, who specialises in developing artificial intelligence applications to improve productivity among a wide range of industries. He described Sáastun AI as a kind of smart digital assistant for archaeologists in the Maya world. This assistant is programmed to do three things in order: find the glyphs in an image, figure out what each glyph is, and then translate the whole sequence.

While the AI model is not open to the public yet, González says it will go live for everyone ‘not too long from now’.

How Sáastun AI Works

The first step is for the AI to separate the glyphs from everything else in the picture, like artwork or background patterns. González trained his system to do this using a first batch of just over 200 example images. Generally, Sáastun AI does this well most of the time, but it can get confused. For example, it might mistake decorative carvings on a ruler’s clothing for written text. Fixing this requires showing the AI many more examples of what is and isn’t a glyph.

Next, the system must arrange the glyphs in the correct order for reading. This is tricky because Maya scribes did not always write in simple straight lines. Texts can curve around the edges of a sculpture or be arranged in complex patterns. González’s current method groups glyphs of similar size together to read them. He admits this solution isn’t perfect and needs more work to handle all the different layouts.

The biggest challenge was teaching Sáastun AI to recognize individual glyphs. There isn’t a giant library of labeled photos for this purpose. González’s solution was to use existing databases where experts have already identified glyphs in drawings. His AI studied over 70,000 of these drawings to learn the patterns. It became quite good at identifying glyphs from these clean illustrations. 

However, a major limitation emerged: Sáastun AI struggled when given real photographs of worn, ancient stone monuments rather than clean drawings. Bridging this gap is a key area for improvement.

Finally, the system translates. It compares the sequence of glyphs it found to a huge database of known translations. It looks for the closest match. Then, it uses a modern language model to turn that match into a clear English sentence. In a demonstration, the system successfully translated a section of text from a Maya panel.

During the presentation, Gonzalez explained the importance of the segmentation of elements which make up any particular Maya glyph to make sense of its meaning. Photo: Carlos Rosado van der Gracht / Yucatán Magazine

Sáastun AI’s Current Limits and Future Potential

González was open about the system’s flaws. It is not perfect and can sometimes make things up, a common problem with AI known as “hallucination.” It also doesn’t fully understand Maya grammar or the creative ways scribes sometimes combined glyphs. For now, it’s a tool to suggest a possible reading, not to give a final answer.

Despite these limits, the possible uses are significant. González imagines it could help field archaeologists get a quick sense of what a text says. It could help piece together broken monuments by suggesting how fragments might fit together. On a larger scale, it could quickly analyze thousands of inscriptions to find patterns. It might even tell what a worn-away glyph could have been based on the ones around it.

But even more importantly for González, Sáastun could be used for education. González showed how the technology could work in tandem with virtual reality, letting anyone, anywhere in the world, explore a 3D monument, stelae or text and see translations appear. This could make Maya history much more accessible to students and the public.

A Call To Come Together

González is currently working independently on Sáastun AI, but looks to partner with universities and archaeologists. At the end of the day, like any AI model, it can only get better with more and better data—exceptionally high-quality photographs of actual monuments, codices, and pottery. “If we have the data, we can train a model,” he said.

The researcher also noted that, although he is open to working with researchers from anywhere in the world, he is especially keen to collaborate with students and researchers in Mexico, his country of birth. “All too often, we look to the rest of the world to advance our technology, but we are more than capable of doing it ourselves; though a little outside help never hurts, González told Yucatán Magazine.