ARION
arionresearch.org 2026 · founding initiative

What if the next frontier model could begin to understand whales?

ARION proposes injecting tokenized cetacean phonetic data into frontier language models — using the same cross-lingual alignment that already lets models understand 100+ human languages.

h · human text
la pluie tombait doucement la pluie tombait doucement the cat sat on the mat 猫はマットの上に座っていた la pluie tombait doucement the cat sat on the mat 猫はマットの上に座っていた el gato estaba sobre la der Wind war kalt heute le chat est sur le tapis the cat sat on the mat la pluie tombait doucement 猫はマットの上に座っていた le chat est sur le tapis el gato estaba sobre la le chat est sur le tapis the cat sat on the mat we walked along the shore 猫はマットの上に座っていた 猫はマットの上に座っていた the cat sat on the mat la pluie tombait doucement le chat est sur le tapis we walked along the shore 猫はマットの上に座っていた le chat est sur le tapis el gato estaba sobre la
w · whale codas
O.heavy V.o O.ext T.fast T.med T.fast R3.irr V.u R5.orn R4.reg R3.irr V.o · T.slow O.heavy · O.heavy O.none V.i O.heavy O.ext V.a T.med V.o T.fast O.ext O.ext R3.irr · T.slow R3.irr R3.irr T.med R5.orn V.a V.u O.none · T.slow T.fast O.heavy O.heavy R5.orn R5.orn RB.flat RB.rise O.none · V.u · R2.reg · T.med O.ext R3.irr O.ext V.o T.fast RB.rise T.slow V.a T.fast V.a V.a · T.slow · V.a · RB.rise · O.ext R2.reg R3.irr V.o
§ 02 The wall

Specialized models hit a wall.

Current AI models for animal communication — DolphinGemma, WhAM, and others — have made remarkable progress discovering patterns within species-specific vocalizations. They can predict next sounds, cluster behaviors, and even generate realistic calls.

But they remain siloed. They never share an embedding space with the rich conceptual world encoded in human text. Without that shared foundation, they cannot describe what a whale coda means in English, nor map it to human-understandable ideas like cooperation, kinship, or navigation.

FIG.A   Siloed vs. sharedΔ semantics

   SILOED
   audio  ──▶  clusters  ──▶  ∅ (dead end)

   SHARED
   phonetic text   
                   ├──▶ shared model ──▶ explanations
   human text                                in both domains
The siloed approach hits a semantic ceiling. The shared-modality approach inherits human concepts for free.
§ 03 The bridge

Phonetic text is the bridge.

Project CETI has already cracked the first step: a phonetic alphabet for sperm whale codas. Each coda is encoded as a structured string capturing rhythm, tempo, rubato, ornamentation, and vowel-like qualities.

This is text. The same modality as English, Mandarin, Python code, or ancient Sumerian. A standard tokenizer can process it. A frontier model can train on it. No cross-modal adapters. No architectural changes.

Example coda notation
R4.regT.fastO.heavyRB.riseV.a
Rhythm · Tempo · Ornamentation · Rubato · Vowel quality
§ 04 Named for

Named for the poet saved by dolphins.

In 625 BCE, the poet Arion sang on the deck of a ship. Dolphins gathered, drawn by his music. When he leapt into the sea, a dolphin carried him to safety.

Acoustic signal. Cross-species understanding. A shared modality bridging two worlds. The story we're trying to write with AI is 2,600 years old.

Read the full story

"He sang, and they came.
He leapt, and they carried.
Acoustic signal across
the species boundary."
— fragment, after Herodotus
§ 05 Built on published research from

A pipeline assembled from existing work.

See the research landscape

§ 06 Now

The next frontier run is coming.
Include the data.