How we think about AI.
AI is now a general-purpose technology. That framing matters. It means the interesting question is not "which model is best this quarter" but "which systems, habits, and institutions do we need so that we can deliver value and transparency with this technology". Our work is organised around that second question.
We are optimistic about what AI can do for industry, research, and public life. We are realistic about the conditions under which that optimism is earned: good data, good evaluation, good engineering, and the institutional patience to let systems mature before they carry real weight.
Our principles in practice.
Legibility before scale.
Before we scale a system up, we make sure the current version is understood. Scale without legibility is a liability.
Evaluation as a first-class product.
Every Glassbox engagement delivers an evaluation suite alongside the system itself.
Data you can defend.
We will not build on data pipelines we would be uncomfortable defending in front of a regulator, a journalist, or an affected customer.
Small teams, senior teams.
A Glassbox engagement is staffed by a small number of senior practitioners, not a pyramid.
Paid to tell the truth.
Our clients hire us because we will tell them, plainly, whether an AI project is a good idea, whether a vendor's claims hold up, and whether a system is ready to go live.
Sovereignty, plainly stated.
That is a problem for three reasons.
- Reason 01
Resilience.
A country that cannot build, operate, and audit its own AI systems has outsourced a growing share of its decision-making to providers it cannot compel.
- Reason 02
Fit.
Australian institutions have data, regulatory, and cultural contexts that off-the-shelf systems do not understand.
- Reason 03
Opportunity.
The economic surplus of AI will be captured, in part, by the countries that can do this work themselves.
Glassbox Labs is a small piece of that connective tissue. We build here, we employ here, we train here, and we keep the option open for our clients to keep their data, their weights, and their accountability here too.
Safety and responsibility.
We take the dual-use nature of AI seriously. We decline work we think is harmful. We run our own internal review before taking on engagements in sensitive sectors.
None of this is marketing. It is how we stay a company we want to work at.
Writing.
We write because good thinking about AI is still scarce, and because writing is how we keep our own thinking honest.
Technical notes.
From inside our own engineering, aimed at practitioners.
Policy & governance.
Commentary aimed at the people making decisions about how AI is procured, deployed, and regulated in Australia.
Essays.
On the longer-arc questions: institutional design, the economics of intelligence, the shape of a working AI ecosystem in a middle power.