In 2017, a Google Brain team published Attention Is All You Need, introducing a new type of neural network, the Transformer, an invention that increased processing speed and efficiency by orders of magnitude by leveraging the power of parallel processing.
Language models existed prior to the existence of Transformers, but they were a far cry from the systems we know today. Think of the clumsy smartphone spell-checkers of a decade ago, when "Sorry for the inconvenience" becoming "Sorry for the incontinence" felt less like a bug than the norm. Transformers introduced the architecture that made large-scale improvement possible.
A few years later, an OpenAI team followed with another breakthrough: by observing the performance of Transformer-based systems at the surface level as they scaled, they were able to identify patterns without first needing to fully understand the internal mechanics. That work became Scaling Laws for Neural Language Models (2020). These scaling laws made AI development more predictable by showing that increases in compute, data, and parameters produced consistent performance gains, which in turn guided model design and accelerated investment. In that sense, Transformers became the critical infrastructure that allowed language models to scale into LLMs, and scaling laws supplied the predictability that gave model scaling technical and financial credibility.
AI safety and governance followed a different path. They were treated mainly as problems of policy and oversight, at some distance from the underlying technology itself: frameworks, audits, and jurisdictional debates over regulation. Meanwhile, AI capability was being industrialized at extraordinary speed, under competitive pressures that left safety and governance fragmented, reactive, and difficult to operationalize.
That asymmetry stayed with me. My conviction was that until these tensions were resolved, AI could never fully realize its potential, either ethically or economically. I began to wonder whether AI safety and governance needed their own version of the step change that Transformers created for capability: an infrastructure layer capable of scaling with deployment, paired with structured observation of system behavior that begins to clarify what is actually happening inside. If so, it could become a foundation for trust in AI grounded in evidence and help unlock wider adoption and sustained investment.
AI safety is advancing on several fronts at once: alignment tries to shape what models aim for, interpretability tries to explain what they are doing, evaluations probe what they can do and how dangerous they may be, and governance tries to set the rules for their use. But when a lab, deployer, regulator, auditor, or insurer has to decide whether a real system should be deployed, under what constraints, and with what ongoing evidence, those insights are still fragmented. We can produce findings, benchmarks, audits, and policy recommendations, but we still lack the scalable infrastructure to turn them into measurable operating limits that can be agreed upon, computed, monitored, and enforced in realtime. That is where I believe the next breakthrough is needed.
Ludulluu is built around a simple idea: AI safety will remain incomplete until we can translate safety concerns in real time from across the stack into measurable operating conditions for real systems at scale.
Our approach borrows from the way scaling laws were discovered: working from system behavior inward. In interpretability terms, that means trying to reveal structure in the black box without opening it first. If the observable signals are precise enough, and tied tightly enough to evidence and outcomes, they may begin to surface recurring signatures of unstable or deceptive internal dynamics, proxy bundles that predict failures before they occur, and the conditions under which specific failure modes emerge.
That would be more than a governance improvement. It could become a new scientific interface to AI systems: a way of discovering which external behavioral patterns reliably co-vary with hidden internal dynamics, and of guiding where deeper model-internal theories should be sought. If those correlations prove robust enough, the result would not just be better oversight, but a new path toward understanding AI systems through structured observation of their behavior at scale.
What excites me is the possibility that safety and governance could become cumulative in the way capability did. Before Transformers, intelligence could improve, but it could not yet compound through a scalable architecture. AI safety feels similar today: important work is happening everywhere, but much of it is still episodic, manual, and non-cumulative.
Without a scalable infrastructure layer, safety remains episodic and non-cumulative. What remains out of reach is a world in which safety becomes a system property: measurable, monitorable, and able to improve at the pace of deployment itself. That is how the false dilemma between safety and speed begins to break down.
The layer that makes this possible has a name: authorization infrastructure. It is the capacity to define, compute, monitor, and update the conditions under which AI systems operate. When governance becomes computable, safety becomes structural, comparable, and eventually priceable. That can fundamentally improve the economics of AI deployment. Most importantly, it stops being an episodic act of interpretation and becomes something AI systems can continuously operate within.
If capability scaling defined the last era of AI, the next question is what makes safety and governance scale with it.