Beyond Chatbots: Why World Models and Embodied AI Matter
A useful glimpse of what comes next
A session with Google DeepMind at the London AI Hub prompted me to revisit a question that keeps coming up in my work: are we asking language models to do too many different jobs?
The session covered agent development tools, including Antigravity and the Interactions API. There was also discussion of techniques for reducing token use. That combination interested me because it linked practical development with the economics of operating AI systems.
I enjoy seeing tools while they are still being demonstrated in a development environment. It makes the distance between an idea, a working prototype and an operational product more tangible.
Token efficiency is a systems question
Reducing token consumption can make a workflow cheaper and faster. But it also invites a broader question: which parts of the job need generative reasoning in the first place?
Retrieving a known record, applying a fixed calculation or tracking a process state may be better handled by a database, a conventional program or another specialised component. A language model can help interpret the request and coordinate those functions.
Token optimisation is not proof that language models are the wrong technology. It is a reminder to choose an architecture around the job, rather than assume every step should be another model call.
From language to action
Much of the public AI conversation centres on chat and coding. Both matter enormously, but I am particularly interested in what happens as intelligence connects more deeply to physical environments.
A robot needs to account for objects, movement, uncertainty and the consequences of its actions. Explaining a task in language is only one part of completing it.
World models offer a way to learn useful representations of an environment and anticipate how it may change. Combined with perception, planning and control, they could help systems choose and evaluate actions.
That is why I see embodied AI as an important next frontier. It creates a much more demanding test of whether a system's predictions correspond to reality.
Why I am watching Google DeepMind
My view is that Google's research across several areas of AI could become particularly valuable as these capabilities converge. The opportunity lies in combining systems, rather than treating a single model as the answer to every problem.
That remains a strategic judgement. Impressive research does not guarantee a successful commercial product, and physical deployment brings challenges that a software demonstration can avoid.
For businesses, the practical lesson is to keep looking beyond the chatbot interface. Ask what the system needs to perceive, remember, predict and do. Then choose the components that serve those requirements.
The next phase of AI could be shaped as much by how well those components work together as by how fluently any one model can speak.