Small AI means artificial intelligence models compact enough to run on an ordinary mobile phone with a weak internet connection. The World Bank is urging developing countries to build such models for their own local needs.
Defining small AI
“Small AI” is not a formal technical category. It is a convenient label for artificial intelligence systems designed to be modest in size, cheap to run and usable without powerful hardware or a fast network connection. The most familiar AI products, such as large chatbots and image generators, usually run on large clusters of specialised processors in data centres. A user sends a request over the internet, the data centre does the work and the answer comes back. That arrangement depends on reliable connectivity, on someone paying for the computing power, and often on a model trained mainly on languages and content from wealthy countries.
A small AI model is built the other way round. It is compact enough to run directly on a device such as a mid-range or basic smartphone, or on a modest local computer, and it can keep working when the connection is slow, patchy or absent. It usually does a narrower range of tasks than a general-purpose chatbot, but it can do them in places where the large systems are impractical.
According to the Guardian, the World Bank is now encouraging developing countries to start creating models of this kind. The bank’s argument is that the benefits of AI will reach emerging economies fully only if the technology is adapted to local conditions, rather than imported unchanged from products designed for richer markets.
Origins of the idea
Over the past several years, the main direction of the AI industry has been towards ever larger models. Big technology companies, many based in Silicon Valley, have competed to train systems with more parameters, more data and more computing power, and to turn them into increasingly sophisticated consumer products. The Guardian describes this contrast directly: while large tech firms concentrate on advanced consumer offerings, the World Bank is pointing to a different route for poorer countries.
Interest in smaller models grew alongside the large ones for practical reasons. Running a very large model is expensive, uses a lot of energy and requires a dependable network. Researchers and companies therefore developed methods to shrink models while keeping much of their usefulness, partly to cut costs and partly so that AI features could run on phones and laptops. Development organisations saw that the same methods could suit places where mobile phones are common but broadband, steady electricity and data-centre capacity are scarce.
The Guardian reports that the World Bank’s chief knowledge officer argued that AI could shorten the process of economic development from a matter of centuries to a matter of decades, on the condition that it is fitted to local needs. The detailed reasoning behind that estimate, and any timetable or funding attached to the bank’s call, is not set out in the material available here.
How it works today
Several established techniques are used to make AI models small enough for everyday devices:
- Distillation: a large “teacher” model is used to train a smaller “student” model that imitates its behaviour on selected tasks.
- Quantisation: the numbers that make up a model are stored at lower precision, which reduces memory and processing needs, often with only a modest loss of accuracy.
- Pruning: parts of a model that contribute little to its output are removed.
- Narrow training: rather than trying to do everything, a model is trained for a defined job, such as answering farming questions, transcribing a particular language or reading a specific kind of document.
Once a model is small enough, it can be installed on a phone and run locally. This has several consequences that matter in low-income settings. Users are not charged for data each time they make a request. The service keeps working where coverage is unreliable. Sensitive information, such as health or financial details, can stay on the device. Local developers can also adapt the model to languages, dialects and circumstances that global products tend to overlook.
Because data for many local languages is limited, building useful small models often depends on collecting and curating new local datasets. That is one reason the World Bank’s call is framed as something developing countries should start themselves, rather than wait to receive.
Common misconceptions
That small AI is simply worse AI. Small models are less capable across the board than the largest systems, but a well-designed small model can perform a narrow task well. For many development uses, reliability, cost and offline operation matter more than breadth.
That it needs no infrastructure. Running a model on a phone removes the need for a constant connection, but someone still has to train, update and maintain it. That requires skills, data and some computing resources, even if far less than for frontier systems.
That it is a single product or programme. As reported, the World Bank is urging countries to begin creating such models. That is a policy direction, not a specific piece of software. Which countries, sectors or projects will be involved is not known from the available reporting.
That faster development is guaranteed. The claim that AI could compress development timelines is conditional. The bank’s own framing, as the Guardian reports it, depends on local adaptation, and the outcome will also depend on many factors beyond technology, including education, governance and investment.
That it competes directly with big tech. Small AI often builds on research and openly released models from large companies and research groups. The difference lies in where and how the technology is deployed, not in a complete break from the wider industry.
Where to look next
Readers who want to follow the subject can start with the World Bank’s own publications on digital development and artificial intelligence, which set out the institution’s priorities and any programmes that follow from this call. Academic and industry research on model compression, including distillation and quantisation, explains the technical limits of what small models can do. Work by researchers and community groups building datasets for under-represented languages shows one of the main practical bottlenecks. Reporting from technology and development journalists, including the Guardian’s original article, gives context on how the policy debate develops. Telecommunications regulators and mobile industry bodies publish material on phone ownership and network coverage, which determines where on-device AI is most useful.
Frequently asked questions
What is small AI?
Small AI is an informal term for artificial intelligence models built to be compact and inexpensive, so that they can run on everyday devices such as mobile phones rather than in large data centres. These models usually focus on specific tasks and can keep working with a weak or intermittent internet connection, which makes them suited to places where connectivity and computing resources are limited.
Why is the World Bank promoting small AI models?
The Guardian reports that the World Bank is urging developing countries to start creating small AI models that can run on phones with low connectivity. Its view is that AI could speed up economic development in emerging economies, but only if the technology is adapted to local needs instead of relying on products designed mainly for wealthier consumer markets.
Can AI really run on a basic mobile phone?
Yes, within limits. Techniques such as quantisation, distillation and pruning can shrink models enough to run on modest phones. Such models cannot match the breadth of the largest cloud-based systems, but they can handle defined tasks, such as answering questions in a particular field or processing a local language, without sending every request to a remote server.
How is small AI different from chatbots run by big tech companies?
Large chatbots typically run on powerful servers and need a steady internet connection, with costs carried by the provider or the user. Small AI runs locally on a device, works offline or with poor coverage, and is usually narrower in scope. It can also be trained for local languages and conditions that global products often serve poorly.
Will small AI speed up development in poorer countries?
The World Bank’s chief knowledge officer suggested, according to the Guardian, that AI could shorten development from centuries to decades if adapted to local needs. That is a conditional projection, not a guarantee. Results will depend on local data, skills, investment, governance and whether the tools address problems people actually face.
Sources and further reading
- The Guardian: technology reporting on the World Bank’s call for developing countries to build small AI models.
- World Bank: publications and policy material on digital development and artificial intelligence in emerging economies.
- Academic research literature: studies on model compression, including distillation, quantisation and pruning for on-device AI.
- Mobile industry and telecommunications bodies: reports on phone ownership and network coverage in low- and middle-income countries.
Surfaced from the rss:guardian_tech signal “AI for developing economies”. AI-assisted draft, editorially reviewed.
