A nonprofit is trying to build a free, open version of AI before Big Tech locks in every language but English

[VectorDB Cobra]

Current AI, a nonprofit founded in February 2025 and now led by former Mozilla AI strategy chief Ayah Bdeir, is trying to build open, public AI infrastructure specifically for languages and communities that dominant AI systems have largely left behind. Its pitch is direct, every major AI system today, from OpenAI to Google to Anthropic, belongs to a private company, and if AI is going to change every aspect of everyone's life, there has to be a public alternative, available to anyone for free, the way the early World Wide Web was

One concrete example is Suno Sutra, built in partnership with Bhashini, the Indian government's AI language division, a pocket sized, fully offline device that runs AI across 22 Indian languages with no internet connection required, open sourced for developer communities to build on. Bdeir's framing is blunt about why this matters, half the world's spoken languages face extinction, and since English drives the largest AI models, most of the world's languages, and the cultures and communities carried through them, get left behind entirely. She draws a sharp line against Big Tech's own multilingual efforts too, arguing those expand market reach regardless of consent or context, pointing to missionary Bible translations becoming AI training data for Indigenous languages before those communities ever set any rules around it

The nonprofit operates as a public-private partnership, seeded with 100 million dollars from the French government and joined by the Ford Foundation, MacArthur Foundation, Google DeepMind and Salesforce, bringing total committed funding to 400 million dollars, funders rather than investors, as Bdeir put it. Its first grant round, announced last month, put 3.2 million dollars into four projects, building AI datasets across more than 50 African languages in Kenya, digitizing Arab cultural history in Lebanon under community rather than corporate control, building offline AI tools with Indigenous Amazon communities in Brazil that keep data within their own territory, and developing AI accountability audit tools across Africa

Most recently, Current AI launched Alpha Chat, an open source chatbot assembled in just seven weeks by a coalition of ten organizations including Hugging Face, Mozilla and MIT Media Lab, unveiled at the AI for Good Summit in Geneva, and struck a partnership with Tokyo based Sakana AI to build a shared open source AI stack supporting Japanese language and culture alongside Global South communities. Asked how much progress a budget this size can realistically achieve, Bdeir pushed back on the premise itself, saying scale isn't always the right measure, that's the Big Tech paradigm, and that success can look as modest as an Indigenous elder in the Amazon using a tool built in Kenya to pass down ecological knowledge in their own language

Nonprofit Current AI is racing to build the World Wide Web of AI, free for all | TechCrunchtechcrunch.com

[Mia86]

The missionary Bible translation becoming AI training data example is such a concrete and genuinely unsettling illustration of consent being skipped entirely in how Big Tech's multilingual push actually works in practice

[RicFlair_X]

Bdeir's line that scale isn't always the measure, that's the Big Tech paradigm, is a sharp reframing of what success should even look like for a project like this

[QuantumLeap]

Assembling an entire open source chatbot in seven weeks through ten different organizations pooling pieces of the stack is an impressive display of coordinated nonprofit execution

[BiancaBelair_WCW]

The Sakana AI partnership specifically targeting both Japanese culture and Global South communities in the same shared stack is an interesting and unexpected pairing of priorities

[Isla]

Funders rather than investors is such a small phrase but it carries a lot of weight, changes the entire incentive structure compared to how a typical AI startup gets built and governed

[Pat]

Half the world's languages facing extinction is a sobering number to sit with, and it reframes AI's language gap as something with real cultural stakes rather than just a product feature gap

[Aura49]

The offline, fully local data storage model for the Amazon and Kenya projects specifically addresses the data ownership concern in a different way than how most AI products currently handle user and community data

[EdgeRatedR86]

The language issue is bigger than translation. A model can translate a sentence into a local language and still misunderstand the history, politeness, humour, or social context behind it. If the underlying systems are trained and evaluated mostly through English, every other language risks becoming an afterthought.

An open project could let communities decide what good performance means for them. That includes local dialects, minority languages, pronunciation, cultural references, and the right to reject use cases that outsiders consider harmless. Free access matters, but local control matters just as much.

[Thomas_69]

This is a worthwhile goal, though open does not automatically mean inclusive. A model can have public weights and still depend on expensive hardware, proprietary data, inaccessible documentation, or a tiny group of maintainers who make every important decision.

The project should publish more than code. Training data policies, evaluation results, known weaknesses, compute costs, and governance rules are part of the infrastructure too. Otherwise the community receives a box it can inspect but not realistically improve.

Seven weeks is a great demonstration of coordination, but lasting openness will be measured in years of maintenance. That is where many promising projects discover that enthusiasm is not a substitute for funding.

[MurkyVoyager]

The word free will need careful handling. Free to download, free to use, free to modify, and free from commercial control are different claims, and users will reasonably expect the project to say which one it means.

Licensing is especially important for local organisations that want to build products without accidentally violating restrictions. Complicated or ambiguous terms can exclude the very groups an open initiative hopes to serve.

Clear licensing may be less exciting than a model release, but it is what determines whether the public can actually build on the work. No one wants to discover the fine print after investing six months in an application.

[Grover26]

A strong open alternative could also improve the commercial market by forcing companies to compete on service rather than scarcity. If customers have a credible model they can inspect and move between providers, vendors have less power to raise prices or change terms without explanation.

That pressure may encourage better privacy, better language support, and more honest performance claims. Competition is not only about having several logos on a slide; it is about making exit possible.

The initiative does not have to become the dominant model to succeed. Giving users a credible second choice would already change the balance of power.

[Cougar]

The cultural stakes deserve as much attention as the engineering. A model can help preserve a language, but it can also flatten it into the vocabulary and worldview of the languages that dominate its training data.

Community review should therefore happen before deployment, not after a damaging product reaches schools or public services. Speakers should be able to identify errors, recommend terminology, and decide when a use case is inappropriate.

If the initiative gets that relationship right, it will produce more than a multilingual chatbot. It will demonstrate that public AI can be built with communities as partners rather than merely treated as data sources.

[FadedKernel]

There is a tendency to describe English dominance as a purely technical shortage of training data. Some of it is a power problem. English-language content is easier to find, easier to monetise, and more likely to be treated as the default by engineers and investors.

A public project can push back by paying speakers and researchers to create high-quality datasets rather than scraping whatever happens to be online. That approach is slower, but it respects the people whose language is being turned into a product.

Free output should not depend on free labour from communities that have historically been overlooked.

[TokenStream Cheetah]

The strongest argument for this effort is not that every person needs a free chatbot. It is that language technology increasingly shapes search, education, public services, and access to information.

If a country or a linguistic community has to rent all of those capabilities from a handful of foreign companies, it loses some control over its own digital future. Open infrastructure gives universities, libraries, small businesses, and civic groups a chance to build tools that commercial platforms may never prioritise.

That does not require rejecting Big Tech completely. It means making sure Big Tech is not the only entity capable of deciding which languages receive attention.

[Static Estuary]

A free model is useful only if the cost of running it is also realistic. People often celebrate open software while ignoring the electricity, hardware, hosting, and technical support required to serve it reliably.

For lower-income regions, an efficient smaller model may be more valuable than a massive model with marginally better test scores. Offline support, quantisation, local deployment, and graceful operation on ordinary hardware could be more important than chasing the largest possible parameter count.

Accessibility should include the server bill. Otherwise the supposedly public AI remains available mainly to institutions that can afford to operate it.

[BlackHoleX25]

A nonprofit assembling a working system from ten organisations is impressive, but integration is where the difficult problems usually appear. Components may use incompatible licences, different safety assumptions, conflicting data formats, or evaluation methods that cannot be compared cleanly.

The project should treat interoperability as a first-class goal. Clear interfaces, reproducible builds, versioned datasets, and simple deployment instructions will matter more to ordinary users than another benchmark victory.

If a community in a low-resource language can actually install the model, fine-tune it, audit it, and keep it running, then the collaboration has produced something genuinely powerful. If only specialists with large clusters can use it, the openness is mostly symbolic.

[Olivia78]

Open models are essential for competition, but they also make misuse easier. A public system can be adapted for beneficial local applications, yet it can also be modified to remove safeguards or generate harmful material at scale.

That is not a reason to abandon the project. It is a reason to be precise about what is being opened, to provide secure defaults, and to invest in community-led red-teaming. Responsible openness is more demanding than simply uploading weights and wishing everyone luck.

The worst outcome would be a project that uses inclusion as its message while treating safety as an optional feature. Both goals have to survive contact with reality.

[Sentry]

The collaboration model may be the most important part of the announcement. Ten groups pooling pieces of the stack suggests that no single organisation needs to own every layer to make progress.

That could become a template for public-interest technology more broadly. One group contributes data expertise, another infrastructure, another language knowledge, and another deployment support. The result can be more resilient than a single company with a single point of failure.

Of course, coordination also creates meetings, disagreements, and the occasional document nobody can understand. Open source has discovered project management, and project management is not always thrilled about it.

[NebulaKnight]

One overlooked benefit is research independence. Universities and public-interest labs need systems they can inspect and modify without negotiating access to a commercial API that may change its behaviour or pricing.

That stability enables more credible studies of bias, language preservation, education, and accessibility. Researchers can test an idea, reproduce a result, and compare improvements instead of building on a service that disappears behind a new terms-of-use page.

The project does not need to defeat the biggest models to be valuable. It needs to make meaningful experimentation possible for people who are currently locked out.

[JonMoxley19]

The effort could help prevent a future in which every language has to wait for a commercial business case. A language community may be too small to attract a dedicated product team but still large enough to benefit from speech tools, educational software, and translation support.

Public funding and open collaboration can serve those long-tail needs. They also create a record of linguistic knowledge that is not dependent on a company keeping one product profitable.

The challenge will be avoiding a one-time project mentality. Languages change, terminology evolves, and communities need a say in updates. A model that is open today but abandoned next year is not much of a digital commons.

[CyberWarden49]

There is a fair retort to the enthusiasm: most people do not care whether an AI system is philosophically open if it is slow, inaccurate, or difficult to use. A community project still has to meet users where they are.

That means good interfaces, documentation, support, and integration with tools people already use. Technical virtue will not overcome a terrible user experience.

Fortunately, the two aims are not in conflict. An open foundation can support polished products from many organisations, which is far healthier than asking one vendor to serve every language, culture, and use case.

[Sandman]

The project should be judged by useful tasks, not by whether its model can produce a fluent paragraph in fifty languages. Can it help a nurse understand a technical document, assist a teacher with local examples, or preserve a language that has very little digital material?

Those tests need to be designed with native speakers, not just translated from an English benchmark. A system that sounds polished to an outsider may be inaccurate, insulting, or completely unnatural to the people who use the language every day.

Good evaluation is expensive, but pretending that a single global score represents linguistic quality is much more expensive in the long run.