Mpaka

Making progress on the biggest problems in development at the speed of the frontier

London, September 2026.

I wrote these essays because it seems increasingly likely powerful AI is coming very soon, and I wanted to explore some early thoughts on how we can use it to solve big problems in global development.

My first question was: of the many places we could use AI in development, which ones will make the biggest difference? Rather than start from existing interventions, I asked two questions:

1. What are the biggest barriers to development and wellbeing? (not very novel)
2. How much of the barrier gets lifted if we add unlimited expert work? (more interesting!)

This highlighted many existing ideas, like AI tutoring and clinical AI — although I think we could raise our ambition by a lot here too. But it also pointed towards state capacity: giving revenue authorities unlimited tax auditors, or helping clear court backlogs.

Then the question is what should we actually build, when the models just keep getting better? I think within a year, a frontier model won't just be a high-quality tutor: it'll build, ship and keep improving the tutoring product itself. That means we should focus on what stays hard: building government and public trust, certainly, but also figuring out how to rethink a service around unlimited expert work.

I suspect the private sector's shift to 'owning the work' instead of selling tools is one answer. I'm interested in the potential of creating a few 'lighthouse' institutions (clinics, schools, etc) where we can experiment fast on what the AGI-native service looks like before expanding to national scale.

I also think we need to consider letting AI do more of the actual work. In rich countries, care and teaching are pretty good, so we won't lose much impact (for now) if we restrict AI to providing advice. But in the developing world, this same caution will leave millions of people with worse care and learning than we could give them for (almost) free.

Finally, I did a bit of research to understand what makes moving fast with AI hard in our sector? Every development organisation, whatever its interventions, can multiply its impact by running more experiments and automating busywork. But few people in development have been close to how fast things are moving, and I think that's part of why our ambition is lower than it could be.

Essays

I. Free Thinking: where unlimited expert work can make the biggest difference for development

II. Build What Stays Hard: where we should focus now, if we believe AGI is coming soon

III. Forge: how AI makes learning faster, and what makes this hard today

A note on methods

I wrote these pieces primarily to shape my own thinking as I come up to speed on the opportunities in AI for development. As a result, many of the conclusions here aren't new. But I tried to come at these questions from first principles, and I hope it might prove interesting anyway. Although I worked hard to keep the research unbiased (and many of my early hypotheses ended up being wrong!), I make no claims of deep rigour.

I do not examine the arguably more-important question of what a world with AGI means for poor countries' development paths. I defer there to the real economists (and my friends at the Africa Jobs Fund).

I worked heavily with Claude to do the heavy lifting of gathering and crunching data, and to shape the argument. The factbase and exhibits (e.g. sketches of interventions, reasoning for scores) were largely written by Claude. I wrote the essays.

Updated: September 28, 2026