Project-Labs-x / Research Preview
What comesafter.
Frontier is not a product. It is our public research into architectures, training methods, and alignment techniques that do not exist yet. We build it openly so the field can build on it.
Active research tracks
The questions we are asking.
Post-Transformer Architectures
The transformer is powerful but not optimal. We are exploring state-space models, dynamic computation graphs, and hybrid architectures that trade unnecessary compute for targeted capability.
Efficient Pretraining
Training frontier models should not require a supercomputer. We are studying curriculum design, data quality over quantity, and novel loss functions that reach GPT-4 capability at 10× less compute.
Symbolic-Neural Integration
Pure neural systems hallucinate. Pure symbolic systems cannot generalise. We are building bridges — models that can invoke formal reasoning and retreat to neural intuition when appropriate.
Mechanistic Interpretability
We want to understand what our models are actually doing, not just what they output. Probing internal representations to build trustworthy, auditable systems.