Horizon 01
Artificial Intelligence
Systems that increasingly decide, and what follows from placing them beneath things that matter.
Intelligence is becoming infrastructure.
The shift
For most of computing history, software did what it was told. The instruction set was explicit, the failure modes were enumerable, and responsibility for an outcome could usually be traced to a line someone wrote. That property is eroding.
Contemporary models are statistical, general, and increasingly embedded in workflows rather than sitting beside them. They summarise, classify, draft, route, approve, and act. The effect is not that a task gets faster. It is that judgment moves — from a person, to a system, to a place where it is harder to observe.
The interesting questions are rarely about the model itself. They are about what an organisation stops doing once the model is reliable enough, and what happens on the day it is not.
A model does not need to be autonomous to become load-bearing.
What we ask
- What decisions has the system quietly absorbed?
- Which human checks were removed because the system made them feel redundant?
- How would anyone notice degradation before it became consequential?
- What does the system do when its inputs are adversarial rather than merely noisy?
- Who is accountable for an output that no one specifically authored?
Lines of inquiry
- 01
Models as infrastructure
When a capability becomes reliable, it stops being a feature and becomes a dependency. We examine where models are moving from the application layer to the operating layer, what that concentrates, and what it makes fragile.
- 02
Evaluation and assurance
Benchmark performance is not evidence of fitness for a particular deployment. We are interested in evaluation that reflects the conditions a system will actually meet — including the ones an adversary selects.
- 03
Model and pipeline security
Training data, weights, retrieval sources, tool access, and prompts are all attack surface. Each expands the set of things that must be trusted before an output can be.
- 04
Agentic operation
Systems that plan and take action change the question from whether the output is correct to what the system did. Authority, scope, reversibility, and audit become design problems rather than policy afterthoughts.
The dependency is easier to acquire than to inventory.
If a model has become load-bearing somewhere in your organisation, the useful question is what would notice if it quietly stopped being reliable.