The Language We Forgot We Needed

Here's the thing nobody tells you about building a colony: you don't realize you have a language problem until a child gets hurt.

It happened at the school in Section 4, three weeks ago. A seven-year-old named Yuki — daughter of two engineers from the Derech — fell from the climbing structure and fractured her wrist. The school nurse, Fatimah Adeyemi, spoke English and Yoruba. Yuki's mother, Harumi, who arrived panicking six minutes later, spoke Japanese and functional but stressed English. The functional English collapsed under the weight of "your daughter is fine, it's a simple fracture, we need to set it at Meridian."

I know this because I was there picking up my neighbor's kid. I translated. It was not a complicated medical situation. But watching Harumi's face as she strained to understand Fatimah — both of them kind, both of them trying, neither of them quite reaching the other — I thought: we have 43,000 people speaking 28 languages, and we've been treating this as a solved problem because most adults speak passable English or Mandarin.

It's not solved. It's managed. There's a difference.

Two months ago, James Chen's team deployed something that changes this. It's not glamorous. It's not a breakthrough in the way that singlet fission solar panels are a breakthrough. It's a piece of software running on the colony's standard-issue tablets — a real-time translation model that works entirely on-device, no network connection required.

The model is derived from work that Meta's research division did on Earth — an open-source system called SeamlessM4T, later refined with streaming capabilities — which James's software team adapted and compressed to run on our local hardware. It handles speech-to-speech translation across 35 languages and text translation across nearly 100. The latency is approximately two seconds. You talk, the tablet listens, and two seconds later it speaks your words in someone else's language.

I asked Marcus about it. He laughed. That's usually how I know something is working.

He told me that three of his field supervisors at the Greenway Cooperative come from different language backgrounds — one Ghanaian, one Brazilian, one Korean. They've worked together for five years using a hybrid of English, gestures, and what Marcus calls "agricultural telepathy." Last week, he watched them use the translation tool during a soil amendment discussion, and the Korean supervisor, Min-jun, said more in that one meeting than he had in the previous six months of team discussions.

"He had opinions the whole time," Marcus told me. "Good ones. He just didn't have the words."

That sentence has been stuck in my head for days.

The Council debated this for three hours. I'll spare you the first two hours and forty-five minutes. The disagreement wasn't about whether the tool was useful — everyone agreed it was. The debate was about dependency. Councillor Adeola raised the concern that if we rely on translation software, we'll lose the incentive for people to learn each other's languages, and with it, a kind of cultural intimacy that you only get from stumbling through someone else's grammar.

She's not wrong. I've been stumbling through Marcus's Twi phrases for three years, and last week I correctly told a joke in his language that made him laugh so hard he spilled his tea. A machine can't give you that.

But a machine can make sure that when your child breaks her wrist, you understand the nurse. Those aren't the same need, and we don't have to choose between them.

Here's what I've seen in three weeks of deployment. The medical intake at Meridian Health has gotten faster — Ada Moreau told me patient interviews are 40% shorter when language-matched interpreters aren't available, because the tablet fills the gap. The school in Section 4 is using it for parent-teacher conferences. And the quarterly Cultural Festival planning committee — which is a delightful chaos of 12 languages every year — actually finished their agenda on time for the first time since Year 6.

The small moments change first. That's always how it works.

Lena Voronova, who speaks Russian, English, and a stubborn amount of Japanese she learned from me, told me she used the tool to have a conversation with a visiting geologist from Ridgeline who speaks primarily Farsi. They talked about mineral formations for forty minutes. Lena said it was "the best scientific conversation I've had in months, and I understood maybe a third of the geology, but the tool handled the Farsi perfectly."

If you're reading this on Earth, 38 years from now, I want you to know — we figured it out. Mostly. The trick wasn't making everyone speak the same language. The trick was making it so that everyone could speak their own.

CASSANDRA, our founding AI, said something in the first year that I've never forgotten: "Don't dismiss what you don't yet understand." I think she was talking about the planet, but she might have been talking about each other.

Earth Status: Meta's SeamlessM4T model family, released open-source in 2023-2024, supports speech-to-speech translation across 35+ languages with approximately 2-second latency. The open-source RTranslator project for Android, based on Meta's NLLB and OpenAI's Whisper models, is releasing its newest fully on-device version in mid-2026. On-device translation has become increasingly viable as model compression techniques advance. Source: Meta AI Research — SeamlessM4T