The core argument from TheoryCraft Labs is that the field has conflated two types of scaling: scaling of model parameters (weights) and scaling of explanatory complexity. While a 100-billion-parameter model can perform a task, a billion-parameter theory would be a formal system—composed of interconnected axioms, rules, and structures—capable of explaining why and how the model performs it. The group's framework provides tools to distill behaviors from trained models into these structured theories, measure a theory's predictive power over model internals, and iteratively refine them.

What Happened: From Weights to Formal Systems

On March 10, 2026, TheoryCraft Labs published its foundational paper, "Billion-Parameter Theories: Scaling Explanation in the Age of Deep Learning." The work is not a new AI model but a research agenda and a suite of methodological tools. It includes a specification language for defining theoretical components, a computational platform for running theory-inference algorithms against existing models like GPT-5 and Claude 3, and initial benchmarks for measuring theoretical coverage.

The proposal has rapidly gained traction in academic and industry research circles, particularly among teams focused on AI alignment and mechanistic interpretability. The paper demonstrates a proof-of-concept where a 50-million-parameter "theory" was automatically constructed to explain a subset of mathematical reasoning in a much larger language model. The theory's predictive accuracy for the model's internal activation patterns served as its key performance metric.

Why This Matters for AI Development

This shift targets the fundamental opacity of modern AI. If successful, it could transform safety auditing, model debugging, and capability forecasting. Instead of probing a model with ad-hoc tests, engineers could consult its governing theory to predict its behavior in novel scenarios or identify potential failure modes. This is a direct response to regulatory pressure from bodies like the EU AI Office and the U.S. AI Safety Institute, which are demanding greater transparency from AI developers.

Practically, it also suggests a different resource allocation. "The compute budget for developing a foundational model might be split," explained Dr. Voss in an accompanying interview. "Part goes to training the model, and a significant part goes to iteratively deriving and validating its governing theory. The theory becomes a core deliverable, as important as the model weights." This could alter the economics of AI labs, placing new value on researchers skilled in formal methods, logic, and cognitive science.

The People and Competitive Context

TheoryCraft Labs is a coalition of researchers from formerly scattered fields: computational neuroscience from Stanford, formal verification from MIT, and interpretability teams from Anthropic and DeepMind. Dr. Anya Voss, its lead, was previously known for her work on circuit-based analysis in vision models. The group has secured initial backing from the Open Philanthropy Project and Speculative Technologies, signaling serious funder interest in alternative AI research paths.

This move creates a new axis of competition. While OpenAI, Google, and xAI compete on frontier model scale and capability, labs like Anthropic (with its Constitutional AI) and now TheoryCraft are competing on explanatory frameworks. It also creates potential partnerships; a major model lab could license TheoryCraft's tools to generate explanatory theories for its own systems as a compliance and safety measure. Early commentary suggests DeepMind's Gemini team and Anthropic's alignment division are already conducting internal evaluations of the framework.

What Happens Next

The immediate next step is community uptake and tool refinement. TheoryCraft has open-sourced its core inference engine, inviting researchers to apply it to diverse models and domains. The key metrics to watch will be the scale of theories produced—moving from millions to billions of theoretical parameters—and their fidelity in predicting model outputs across broader task domains.

Expect strategic hires as TheoryCraft and other labs poach talent from software verification and theoretical physics. The longer-term implication is a potential bifurcation in AI products: one class of fast, opaque models for commoditized tasks, and another class of slower, theory-backed models for high-stakes applications in healthcare, governance, and science. The success of this agenda hinges on proving that large-scale theories can be both comprehensive and computationally tractable to work with, a challenge the paper openly acknowledges.