What Base76 Actually Does (and Why)
The AI debate in Sweden often swings between two extremes:
- Hype about productivity
- Fear without technical grounding
Base76 operates in the space between those two.
We are building an independent research lab with one clear focus: reliable AI.
That means we do not only look at model outputs. We look at what happens inside models when they reason, guess, or start to lose stability.
The Problem
Today’s language models can sound confident even when their internal state is fragile.
That is a practical risk for anyone using AI in real decisions:
- organizations
- public agencies
- research environments
- critical social functions
If we evaluate only output quality, we risk missing early warning signs.
What We Do Technically
At Base76, we work on mechanistic interpretability and reliability evaluation.
In short:
- We train sparse autoencoders (SAEs) to identify latent features in model representations.
- We run interventions and trajectory blocking to test what is causally relevant.
- We build reproducible evaluation panels to compare baseline behavior against interventions.
- We develop compact indicators (coherence/residual-state signals) that can function as early risk markers.
The goal is not pretty visualizations. The goal is actionable signals for risk assessment.
Why This Matters in Sweden and the EU
Sweden and the EU are now talking about AI sovereignty.
Sovereignty is not only about where models are hosted.
It is also about epistemic control:
- Do we know when a model is robust?
- Do we know when it starts to drift?
- Do we have methods that are transparent, reproducible, and auditable?
This is exactly where Base76 aims to contribute.
How We Work
We build in open artifacts, with clear experimental protocols and documented limitations.
Negative results are as important as positive ones.
For us, research quality means:
- reproducibility
- methodological clarity
- scope discipline
- practical relevance for safety and governance
What We Are Building
Base76 is being built as a clear Swedish node for independent AI safety research.
Not a consulting label.
Not a PR surface.
A real lab.
Our next step is deeper collaboration with public and private actors that want to test reliability methods and internal model observability in real environments.
If You Want to Collaborate
We are currently looking for:
- pilot projects (8-12 weeks)
- research collaborations with clearly scoped work packages
- partners who want to strengthen robust, auditable AI in Sweden
If this is relevant for your organization, get in touch.
/Bjorn Wikstrom
Base76 Research Lab