The Twokriks plan

Over the past year, technology stopped just answering questions and started doing things. The first personal agents arrived this year: helpers on your phone that can book, plan, pay and reply for you. I believe this is where all of it is heading: everyone will have an agent that knows their work, their money and their habits, and handles the small things for them. The big systems behind today's chatbots will still be there underneath, but the thing people live with every day will be personal.

An agent is only as good as how well it understands you, and that is where the problem starts. Picture a woman who sells fabric in Warri. Her phone is her shop. All day, customers send her voice notes in Pidgin: what they want, how many yards, when they will pay. She keeps all of it in her head and her chats, scrolling back to remember who paid half and who still owes. When she turns to one of today's chatbots for help, she has to stop, switch to English and type, because that is the only way it understands her. She is not unusual. When Nigerians ask ChatGPT to get something done, 96.2% of those conversations are in English, even though only half of their casual chats are. People switch languages to be understood. An agent built on that switch will know the English she puts on for it, not her.

The more an agent knows about her, the more it is also worth to someone who wants to rob her. In a 2024 study, scam emails written entirely by a language model got 54% of people to click, as many as those written by human experts. A cloned voice can already call her and ask for money. Soon her agent will answer that call for her, and the caller will be an agent too. So security cannot be something added once the agent works. It has to be how the agent works: what it knows about her stays hers, and nothing leaves without her say.

Put those two things together and the question is no longer whether personal agents are coming. It is who they will be built for, and who will be left out. The next few years decide that, and whoever is missing from the data when these agents are built will be left out of them for a generation. Our aim as a lab is simple: as technology speeds up, nobody gets left behind. In practice that means using an assistant feels as normal to a market seller or a grandmother as it does to an engineer, at home or abroad. That aim does not change as we grow.

Most of the industry is building one model from the languages with the most data and translating it down to everyone else. We believe it has to be built the other way, from the bottom up: from how people actually talk, pay and decide, in the languages they think in. A model translated down to her will only ever hear the English she puts on for it. A model built up from her life will hear her.

That is why Twokriks exists, and why it starts where it does. We come from the place we are building for. We speak the language, and we can record the voices most labs will never reach.

But you cannot build an agent that understands people from data that does not exist. For Ijaw, the open speech that exists is a few hundred short recordings. For Urhobo, there is none. The newest speech models list both languages anyway, and nobody has published how well they actually hear them. So we start with products people can use today, and let them build the record. Yfraud tells her Stop or Clear before she pays a scammer, and it already catches 80% of real phishing emails in our tests with rules alone. Mama Budget steps into her cart before she breaks her week. FlickChoice gives her one answer when she is stuck. Creeks teaches Ijaw, Urhobo and Warri Pidgin with real voices. Each one does a job for her today, and each one teaches Y something about how she talks, pays and decides.

Y is the brain under all of them. We build it from openly published models that we download, train and keep improving ourselves, never a rented model with someone else's context on top. When stronger open models are published, Y moves onto them and keeps everything it has learned, so it gets better as the field does without ever changing what it is for. Speech comes first, because she would rather talk than type. Today we have two published Ijaw dictionaries, the work of Kay Williamson and Roger Blench, turned into machine readable text with about 19,000 headwords between them, and a Warri Pidgin wordlist of 336 entries, 293 of them glossed by hand. Native speakers have started on our first batches of words and sentences, every one of them with their consent. Those numbers only go up.

Once Y understands her, the companies that already serve her can too. Her bank and the apps she pays with can run Y on their own data and answer her in Pidgin, and that data stays with them. What works where we come from, in the South South, goes next to the rest of Nigeria, then follows our languages across borders into West Africa and the rest of the continent, then to every developing country and every small nation the big systems reach last.

Along the way, we hold ourselves to four things:

• Her context belongs to her, and a company's context belongs to the company. Neither comes back to us.
• We learn from corrections, never from the person. No account numbers, passwords, balances, or messages nobody handed us.
• We publish what we count, including what we get wrong.
• Nothing goes public until it works.

So, in short, the plan is:

1. Build products people use every day, for how they talk, pay and decide, with security under all of them
2. Use what people correct in them to build the record our languages and daily lives never had
3. Train Y on that record, a brain we own, and give every person, and every company that serves them, an agent that understands them
4. Start where we come from, the South South, grow to all of Nigeria, then West Africa and the rest of Africa, then every developing country and every small nation the big systems reach last

When Y works, this is her day. A customer sends a voice note: "Abeg, di blue lace wey you show me, make you keep six yards for me. I go send half now." Before she has served the next customer, the order is written down for her: six yards of blue lace, half paid, the rest due Saturday. On Friday her phone reminds her. When a voice claiming to be her bank calls to "confirm" her account, it is checked before she sends a single kobo, and she is told why in Pidgin. She never typed a word, and she never switched to English. She stops bending to the technology. It bends to her.

We have already started.