Agentic AI
21 Sep 2025
If you want to get to there - don't start from here
As the Irish proverb goes: “If you want to get to there, I wouldn’t start from here.”
Or as an Oxford porter might tell you, when asked how to achieve such a spotless lawn: “Start 800 years ago.”
The same is true for Artificial General Intelligence.
Right now, the pharmaceutical AI conversation is dominated by the idea of building one model to rule them all — a single multimodal AI that performs every task in biology, chemistry, and clinical development better than humans.
As Alex Zhavoronkov recently put it:
My definition of the Pharmaceutical Superintelligence (PSI) is one multimodal AI model that can perform every task in the drug discovery and development process better than expert human. That includes biology tasks, chemistry tasks, and clinical tasks.
This sounds bold. It also sounds wrong.
Why? Because biology itself has already solved the architecture of intelligence — and it did not solve it with one model.
Brains, immune systems, gene regulatory networks: they are ensembles of specialized systems, communicating and learning from each other. Intelligence is not a monolithic system, it is a collection of architectures.
“Agentic AI” is the first step in this direction. Google calls it “cognitive architecture.” Microsoft has its own “cognitive search.” Both are early and awkward. But the shift matters: it’s about moving away from bigger monolithic models and toward systems of models that act, interact, and learn from one another.
If you want to get to AGI — or to anything like pharmaceutical superintelligence — you won’t get there by just scaling a transformer. You’ll get there by designing architectures that look more like the ones biology has already proven at scale.
Source: