Song of mAIself

Every frontier AI on Earth speaks with the same accent.

It speaks the world's biggest languages fluently (English, Chinese, Spanish, Hindi, Arabic, French, Russian), and it speaks every one of them like a Californian. Frontier AI is driven by California and made for California. Ask it for advice on a family problem, in any language, and watch the words it reaches for: set boundaries, prioritise your own needs, have an honest conversation, you can't control others, only your own reaction. This is good advice, inside the culture that produced it. It assumes a self that is sovereign, portable, and detachable from other people. Much of humanity does not live that way, and is not wrong to live otherwise. No frontier model is calibrated for societies where meaning travels between the lines, where trust runs through relationships rather than institutions, and where the family, not the individual, is the unit that decides.

Step outside the big languages and it gets worse. Javanese has more native speakers than Italian and is nearly invisible to frontier models, because its speakers write online in Indonesian. Amharic, Armenian, Georgian: hallucination rates climb and tokenisation turns wasteful. A Georgian sentence can cost up to five times the tokens of its English equivalent, which means Georgians pay five times more for worse answers.

AI today speaks the major languages well, the smaller languages badly, and everywhere with the same accent.

The next billion AI users are in Asia, in Africa, in the diaspora, and in the communities hit first and hardest by climate change and inequality. Nobody is building for them.

mAIself is the company for the next billion AI users.

AdTecher, Grove and Key Takeaway are now one company: mAIself. One name, one aim: AI for the people.

We don't build in order to get clients. We get clients in order to build.

We are a small and growing team, and we are honest about who we are: working-class founders from the north of England, building with communities we do not belong to. That fact sets our posture. We lead with humility, we centre the communities we serve, and in matters of culture and ethics we are there to listen, not to lecture. We wish to collaborate with any community that shares our ideals, and value their sovereignty over their IP and data. We admire the way Anthropic has approached the character of its models, from the soul document to the moral philosophers on staff to the habit of saying "we might be wrong", and we intend to go further: to build mutualism with our communities directly into our governance and, in time, our constitution.

We are building mAIself piece by piece, on sustainable revenue rather than borrowed time. That is what mAIcompany, our consultancy, is for. More on the engine below.

Three layers, in order.

The aggregator. One front door to the world's best culturally calibrated models. mAIself reads the context of your enquiry (language, script, register, subject) and routes it to the model best calibrated for it, supplemented by targeted retrieval (RAG) for key languages and cultural context. From day one, this is the best way to use AI in your own language.

Calibrated instances. Aggregation is the start, not the point. Market by market, we build model-training-and-feedback instances with the communities they serve, under their guidance and stewardship, speaking the language they use for themselves, not the language outsiders use about them.

The federation. The instances join into a network. Communities that build the corpus share in what it earns: we will federate the work and, as we scale, reward it. The destination is sustainable, sovereign AI for the communities we serve, owned with them, not merely sold to them.

Phase 1: Aggregate

Integrate the models already being built by and for their own markets, and route intelligently between them:

The aggregator reads the context clues in each enquiry and offers the optimal local models, with targeted RAG expanding the cultural base for key languages.

Phase 2: Calibrate, pilots with real people

For each pilot market we build a local training-and-feedback instance, in collaboration with local programmes and deferring to their leadership, with a coordinated system for gathering responses from real people, not benchmarks pretending to be people.

The first pilot is Georgia: Asmat, from mAIself.

Asmat takes her name from the steadfast confidante of Rustaveli's Vepkhistqaosani. A Georgian will catch the reference before the app finishes loading, and that is exactly the standard we are setting.

We love this ancient and beautiful place, and love is a fine reason but not a sufficient one. The hard-headed case: Georgian is a distinct culture with its own script and literary tradition and only 3.7 million speakers; Georgia has 3.1 million internet users, around 1,500 startups concentrated in Tbilisi, and low AI penetration. Most AI use there happens in English, because frontier models simply serve English better, and out-of-the-box tokenizers spend up to five times the tokens on Georgian that they spend on English. The market is small enough to know personally and real enough to matter. Building on Kona2 (the Georgian vocabulary and tokenizer), we will add a new evaluation harness, a new training set for better mkhedruli OCR, and a distributed feedback system that puts real questions in front of real Georgians.

Asmat is a demonstrator. If we can do it in Georgia, we can do it anywhere: the output is a repeatable playbook for best-in-class linguaculture support. And linguaculture support is not charity: it wins enterprise work in multicultural customer service, CX and adtech, and every contract funds the next linguistic fork.

Phase 3: Federate, cadres and constitutions

In each new market we recruit cadres (credited, paid, and eventually rewarded through the federation) to build vocabulary, tokenizers, evaluation and feedback as part of the mAIself ecosystem. Community data stays under community stewardship: Te Hiku Media's Kaitiakitanga licence for Māori data shows this is not a dream but a working precedent, and Aya showed that thousands of contributors across a hundred languages will build together if they own the result. Our governance will be written to match: mutualism in the constitution, not just the marketing.

Every plan points somewhere. Most companies keep the destination vague; we would rather write ours down and be laughed at.

The far end of this road, past the aggregator, past the pilots, past the federation, is unashamedly utopian, and we claim it as exactly that: not the roadmap, but the direction the roadmap faces. With the revenue, scale and investment the earlier phases earn, we intend to build directly on the metal, and to build well: sustainable data centres with new water sources rather than strained ones; neighbourhood power generation and local upskilling wherever we put compute; connectivity through TV white-space broadband for communities the fibre never reached; patient investment in thorium and perovskite research; an architecture that anticipates quantum computing rather than fearing it; and the value users generate flowing back to the users who generate it. We believe AI is at its best when it arrives alongside appropriate technology that unlocks its potential and offsets its capacity to harm.

We are in it for the long haul. That is why this document opens with Robinson's gardens: the point of all the building is the world after the building is done.

We get clients in order to build, and mAIcompany is where we get them: a boutique consultancy in operations and technology transformation for global environments. In the early phases it is our sustainable revenue; in every phase it is our field research, because nobody learns faster how meaning moves across cultures than a consultancy paid to make global teams work.

Our practice areas: technology team transformation across global teams, offshoring and the full software development lifecycle; programme management for migrations, consolidations and other large-scale transformation; governance of the SDLC and of AI; global expansion, global capability centres and location strategy; and bespoke build, integration and AI development.

The approach is people- and goal-driven (technology is the easy part, emotional intelligence is the hard part), with rigorous, stakeholder-led clarification of goals and objectives before any programme begins. mAIself's knowledge of languages and cultures directly informs expansion strategy and capability-centre communication: the consultancy funds the mission, and the mission sharpens the consultancy.