Building Smarter Systems Under Constraint: Fireside Chat with Google DeepMind's Lucio Dery

Future Caribbean hosted a fireside chat with Lucio Dery, Senior Scientist at Google DeepMind, led by Lily Dash and co-hosted by Roniesia Gittens, exploring the future of agentic AI, continual learning, modular systems, and the opportunities these technologies could create for emerging markets.

Dery, who grew up in Ghana before studying at Stanford and Carnegie Mellon and later joining Google DeepMind, brought a particularly relevant perspective to the conversation: how to build intelligent systems when compute, capital, data, and specialist talent are not unlimited.

Continual Learning: AI That Improves Over Time

One of the major themes was continual learning — the idea that AI systems should be able to learn and adapt after deployment rather than remain largely frozen after their initial training.

This matters especially in markets that are underrepresented in global datasets. Local languages, institutions, regulations, workflows, and cultural contexts may not be well understood by general-purpose models. Continual learning creates the possibility of systems that become more useful as they interact with real users and environments over time.

Agentic AI Can Lower the Cost of Expertise

The discussion also focused on agentic systems: AI models connected to tools, memory, workflows, and other software so they can do more than simply answer questions.

For emerging markets, one of the biggest opportunities is that these systems can reduce the cost of accessing technical expertise. A project that once required a large engineering team may increasingly be prototyped by a much smaller group using AI-assisted development.

Dery highlighted the potential for agentic systems to help founders and organisations get solutions off the ground even where specialist skills or funding are limited.

The important takeaway was not that AI replaces expertise, but that it can expand access to expertise.

Modularity: Powerful, but Not Without Trade-offs

Another key theme was modularity.

Instead of creating one system that attempts to do everything, developers can build specialised agents for different tasks and domains. This can make systems more efficient and easier to scale.

However, modularity introduces a new challenge: coordination.

Once intelligence is distributed across many specialised components, the system must determine which component to use, when to use it, and how information should move between them.

Dery noted that intelligently switching between specialised solutions can itself be a legitimate and powerful form of problem-solving.

What This Means for the Caribbean

The Caribbean is a fragmented but interconnected market, with multiple currencies, regulators, legal systems, languages, and institutions. That complexity raises an important question: could agentic systems help create better coordination across the region?

The conversation suggested a pragmatic approach.

Start by strengthening individual sectors and solving specific problems well before attempting to connect everything together.

Rather than beginning with an enormous regional AI system, builders can focus on tightly scoped opportunities in areas such as healthcare, finance, transport, climate, agriculture, or public services.

This is particularly important because the biggest obstacle may not always be technology. Dery pointed to human and administrative bottlenecks — data access, regulation, institutional processes, adoption, and coordination — as major barriers to deployment.

The Bigger Opportunity: Build for Constraint

Perhaps the strongest theme of the discussion was that emerging markets should not simply copy the AI development model of Silicon Valley.

Frontier AI increasingly depends on enormous amounts of compute, capital, and infrastructure. But countries and companies operating under tighter constraints may need a different approach: smaller models, smarter systems, better use of data, targeted applications, and architectures designed around real local conditions.

That could become an advantage.

The opportunity for the Caribbean, Africa, and other emerging markets is not necessarily to build the largest AI models. It is to build the most useful intelligent systems for the problems that matter locally.

And in a world where AI is making intelligence increasingly accessible, the ability to build efficiently under constraint may become one of the most important advantages of all.