Why I Started LowRes.AI

From accessible knowledge to constraint-native intelligence

For more than ten years, much of my work with Audiopedia has revolved around a simple question:

How do we make useful knowledge accessible when the usual assumptions behind digital technology do not hold?

What if connectivity is weak? What if literacy cannot be assumed? What if smartphones are shared or unavailable? What if the relevant language is poorly represented online? What if there is nobody nearby to maintain a complicated system?

These questions shaped Audiopedia long before generative AI entered the picture.

We learned to use audio because text was often the wrong interface. We worked with memory cards, basic phones, community radio, solar players, QR codes and WhatsApp because there was never one universal distribution channel.

And we learned something that now feels highly relevant to AI:

constraints rarely arrive one at a time.

Capability is not the same as viability

AI has become extraordinarily capable.

Models can work across languages, summarise complex material, generate speech, interpret images and help people navigate enormous bodies of knowledge.

But much of today’s AI still assumes a particular world: persistent broadband, personal devices, cloud services, user accounts, subscriptions, highly digitised knowledge and technical support.

That world exists.

But it is not the whole world.

A model can understand a language perfectly and still fail the people who speak it if connectivity is unreliable, electricity is scarce, prompts are impractical, subscriptions are unaffordable or nobody can maintain the system.

The problem is not only low-resource languages.

It is the low-resource context as a whole.

Start with reality

That is the idea behind LowRes.AI.

Instead of starting with a model and asking how to adapt it later, we start with the environment itself.

What resources can actually be relied upon?

How do people communicate?

Where does trusted knowledge come from?

Who maintains the system?

What happens when connectivity disappears?

What happens when funding ends?

Only then do we ask what kind of AI belongs there.

That can lead to very different choices: local rather than cloud inference, voice rather than text, shared infrastructure rather than individual accounts, curated knowledge rather than unrestricted web access, or smaller models rather than the most capable ones available.

Sometimes it may even mean deciding that AI is not the right answer.

I have started calling this constraint-native AI.

LowRes.AI is an attempt to build a systems-level framework around it.

Non-extractive by design

There is another issue I think we need to take seriously.

Much of modern AI was built by ingesting huge amounts of data from the internet. We should not reproduce that logic by treating Indigenous or community knowledge as the next underused dataset.

The answer to underrepresentation cannot simply be:

Give us your data and we will build a better model.

For whose model?

Who controls it afterwards?

Who benefits?

Some knowledge should remain local. Some should remain under the authority of communities or knowledge holders. Some should never become training material at all.

That is why non-extractive has become one of the principles behind LowRes.AI.

Not every useful AI system needs to absorb the knowledge it works with into a model.

SILAC is the first test

The first LowRes.AI reference implementation is SILAC, a local knowledge environment we are developing for Indigenous communicators.

Its initial focus is climate resilience, biodiversity, forests and agriculture.

The idea is not to build an “Indigenous LLM.”

It is to create an environment where Indigenous communicators can work with trusted external knowledge and community-governed Indigenous knowledge while keeping provenance and authority visible.

AI can help search, compare, structure and adapt those sources.

But it does not decide which knowledge system is superior.

It does not silently merge everything into one dataset.

And it does not turn community knowledge into model training material simply because that is technically possible.

The goal is a practical diálogo de saberes - AI supporting people as they compare, interpret and communicate different forms of knowledge on their own terms.

Developing the framework in public

SILAC is the first implementation, but the broader design problem appears in agriculture, health, climate adaptation, humanitarian response, education and public services.

LowRes.AI is an initiative of Audiopedia Foundation and grows directly out of what we have learned there. I am also bringing into it what I am learning through my work as a member of the UNDP Global Expert Group on Closing the Language Gap in AI.

This is still early.

I do not want to declare a new field and then defend the terminology.

I would rather build, test, compare and learn.

There is one principle I am fairly confident about already:

Start with reality.

Then decide what AI should become.