A hundred years is a working horizon.

This page is a position, not a claim. The lab’s rule is that we do not report what we did not measure, and nobody has measured the next century. So read this the way it is written: as the direction we are building toward, and the standard we hold ourselves to while building.

Intelligence became abundant. Direction did not.

The models are shared infrastructure now. What is scarce is everything around them: the judgment that decides what is worth building, the structure that lets fleets of agents do real work without drifting, and the evidence that lets anyone trust the result. That scarcity is why this lab exists. We study how agent systems behave on real work, and we build the tooling that makes their behaviour observable, testable, and safe to rely on.

Benevolent systems are built, not hoped for.

We think the future where humans and intelligent systems flourish together is not a prediction to argue about — it is a body of work to contribute to. It gets built the unglamorous way: skills written down and versioned, boundaries and stop conditions in plain files, rubrics that say what good means before the work starts, receipts that record what actually happened, including the failures. Every system built this way is a small argument that capability and accountability can grow together. Enough of those arguments, shared openly, become a craft. A craft, held long enough, becomes a culture.

The question we refuse to skip.

There is an open research question at the centre of all this, and it is the one we have the strongest possible conflict of interest in: does sustained agent assistance degrade the operator’s own ability to defend the conclusions the system produced? The early external evidence on cognitive offloading is not flattering, and it would be convenient for everyone selling AI systems to ignore it. We would rather design against it: state your intent before the system answers, require it to surface counter-evidence, record your decision and your reasons separately, and check later whether you can still defend them unassisted. A system that makes you faster while quietly making you weaker is a bad trade. We want the version of this where the human gets sharper too — and we intend to measure it rather than assert it.

What compounds for a hundred years.

Not any single model, and not our tooling either. What compounds is method: the growing, credited, openly shared set of skills and guidelines for how humans and intelligent systems work together — what we call the canon. Every builder who ships a system with receipts adds to it. Every operator who keeps their judgment sharp while their leverage grows proves the point. That is the work we think the next hundred years asks of the people doing this now: build systems worth trusting, keep the human in command and in growth, write down what you learn, and give it away.

That is what “build toward the stars” means here. Not escape — direction.

— Frank Riemer, Starlight Intelligence, Amsterdam

Written with AI assistance; reviewed, edited, and signed by a human. The research question above is tracked in the lab’s open agenda; when we have data, it will be published with sample sizes — n=1 stays n=1.