Scaling for frontier intelligence
Every change proves itself, or it doesn't ship.
Calibrated Abstention in Language Models
Knowing when not to answer: low-confidence generations become "I don't know", not a hallucination.
Four-Million-Token Context on a Single Consumer GPU
A 4-bit KV cache at 99.998% accuracy, portable between machines in milliseconds: million-token context that is cheap and reliable.
Scaling Knowledge, Not Compute
A new scaling axis for large language models that works within the scaling laws, not against them.
Continual Learning Without Catastrophic Forgetting
Resolving the stability–plasticity trade-off in deployed language models: new knowledge after release, prior capabilities preserved, no full retraining.
Dense Capacity at a Fraction of the Parameters
Matching what a larger network learns with fewer parameters and fewer FLOPs per token.
Old Problems, New Math
Revisiting long-standing failure modes in training, such as units that stop learning for good, with solutions where the mathematics works in our favor.
A New Question About Generalization
On what a single parameter can represent, and what that implies for how far a network generalizes beyond its training distribution.
Vendor-Agnostic Distributed Inference
Hybrid tensor–pipeline parallelism across heterogeneous accelerators: NVIDIA, AMD, Intel and Apple silicon on one runtime.
Sparse-First Fine-Tuning
Full-weight fine-tuning in about half the memory of LoRA, and fourteen times less than full fine-tuning. Open source as Helix.