THE FUTURE IS FOR EVERYONE. IS IT REALLY?

by Vania La Rocca

Last year, in Kampala, Ernest Mwebaze went looking for an artificial intelligence that could speak Uganda's languages. A former research scientist at Google, he tested American models against Chinese ones to see which could handle the country's dozens of tongues.

A Chinese model won.

It handled Uganda's dozens of languages better than anything from Meta or Google; it cost far less, and he could retrain it on his own data. He built a tool called Sunflower on top of it. Today, a coffee farmer named Muganzi asks Sunflower about his crops in his own dialect and receives an answer. The model underneath, the New York Times reported this month, came from Alibaba. On August 10, Mark Zuckerberg published over 6,000 words describing a future in which every child on Earth has a personal tutor with a PhD in every subject. In Uganda, something like that tutor arrived. It was not his. It was not American, and it did not come from Silicon Valley.

The promised future

Zuckerberg's essay, "The Future is for Everyone", does not make its central promise about efficiency. It is invention. "Invention, not automation," he writes, will be the greatest contribution of the technology Meta is spending tens of billions to build. (Is it, really?) Meta's founder describes AI that helps scientists cure disease this century, that discovers new drugs, that teaches every child. He argues that this future should be open, that spreading the technology widely is safer than concentrating it, and Meta released a new open-weight model the same day to make the point. Take the argument at face value, because much of it is right. The purpose of a general-purpose intelligence should be to invent what we cannot yet invent. To buy us capability we do not have. To give us, in the most literal sense, more. And if it is going to be for everyone, everyone has to be able to get their hands on it. Now look at what has actually been shipped.

What the machine does

At scale, what has shipped is mostly automation, not invention. In 2025, the MIT initiative NANDA studied enterprise AI and found that 95 per cent of company deployments had no measurable effect on the bottom line, despite spending 30 to 40 billion dollars. That is a verdict on enterprise return, not on the frontier of science, but it shows where the money went: into making existing office work marginally cheaper.

The economist Daron Acemoglu estimates that AI will add at most about two-thirds of one per cent to productivity over a decade, because only a fraction of what people do is exposed to it, and even less of that is worth automating. Real invention is happening in the labs. Most of the money does not go there. The people building the machine are not pretending otherwise.

Six days after the Meta manifesto, Dario Amodei, who runs the rival lab Anthropic, was asked about the public's distrust of AI. He did not reach for better messaging. The most accurate criticism of the industry, including his own company, he said, is that it has not delivered on its promises to benefit the world. Promising to cure cancer, he added, is closer to a cliché than to anything inspiring. What would be inspiring is curing cancer.

Richer for whom?

Invention is slow, and capital flows to what monetises soonest. A new drug can take decades to be produced, tested, and approved, and usually fails. A climate intervention pays out on a schedule no earnings call can meet. Automation is legible and immediate. It has a line item and a number a buyer will sign this year.

So the company that writes "invention, not automation" ships an ad system and a coding assistant because that is what the business model can be priced at. There is no conspiracy behind this automation.

This is the same logic that shapes who the technology gets built for, and it comes down to a quiet disagreement about what "best" even means. Silicon Valley optimised for one question: how powerful can the model become? Chinese firms, locked out of the top chips by U.S. export controls, could not win that race, so they optimised for a different one: can you download it, afford it, run it on your own machines, and bend it to your own data? They gave their models away. American firms moved toward closed, metered systems. When Meta pulled back from its open Llama line, Chinese open models filled the gap. Zuckerberg's manifesto is now an attempt to reclaim that ground.

"The Future is for Everyone" is, in part, a response to the loss of the rest of the world to Beijing's open models. In Kampala, the second question is the only one that matters. Sunflower, the tool Mwebaze built, is a public resource released by a Ugandan non-profit under Alibaba's open model, and, across 31 Ugandan languages, it outperforms ChatGPT and Gemini on 24 of them. Across Kenya, Uganda, Ghana and Nigeria, the Times found, thousands of developers have made the same choice.

The winning model is not the most capable one, but the one that fits the constraints of the person using it. These constraints include cost, control, and the languages people actually speak. When the constraint is raw precision, the American model still wins, and a Kenyan firm digitising its invoices reached for OpenAI. Capability and usefulness are not the same thing.

None of this means America is losing the market. Ordinary Kenyans still mostly reach for ChatGPT, which one 2025 survey put in the hands of 42 per cent of the country's internet users, and American clouds earn money even when Chinese models run on them. What went elsewhere was the building. China widened the opening politically, too. In July, 10 African countries joined 19 other states as founding members of a China-backed World AI Cooperation Organisation. But openness is not independence. Building on someone else's model, Chinese or American, means it can be yanked, altered or repriced without warning. When Washington had Anthropic cut off access to one of its most powerful models this year, John Tanui, a senior official in Kenya's digital-economy ministry, called it a wake-up call and said his country would lean neither West nor East.

Mwebaze learned the other half of the lesson. After Sunflower began describing China as a democracy and skirted questions about its record in Uganda, and after Huawei came courting, he grew cautious about the politics and started rebuilding on Google's open Gemma model instead. The point was never America or China. The point is that he could choose.

A future handed down by the richest companies, built for the highest-paying customers, is not a future for everyone. A future that is truly for everyone must let people shape it to their own constraints. In Uganda, that already happened. It was not American.

Betting on time

Invention versus automation is not the main differentiating factor. Automation can be the right kind of AI. Sunflower is automation, and it gives a Ugandan farmer something new, in a language he actually speaks. The wrong kind of automation just compresses the life you already have, so you can return to the inbox faster. One expands what a person can do. The other only lowers the cost of what they already did. The expanding kind buys things that did not exist before: the years a cured illness gives back, the decade a stable climate protects, the schooling a child in Kampala actually receives.

If we are honest about what curing a disease or cooling a planet means, "more" includes more time on the planet itself, not a faster way to spend the time we have. The writer Joanna Maciejewska put the test better than any manifesto when she said that she wanted AI to do her laundry and dishes so she could make art, not to do her art so she could get back to the laundry.

We do not need to stop building. We need to stop confusing speed with direction, and capability with usefulness. The most powerful tool we have built is being aimed at the smallest targets. The decision about what we are willing to fund when the payoff is slow, and whether the people holding the compute can be moved by anything other than what increases ARR this year, is not a technical problem.

The future is for everyone, the Meta essay says. In Kampala it arrived anyway, on a model a developer could download, afford, and shape to a farmer's own language. Just not the one built by the company that wrote the sentence.