Frontier research,
openly shared.
Every paper we publish comes with reproduction details, evaluation code, and (where applicable) model weights. Open science and safety reinforce each other: you cannot audit what you cannot examine. Basalt runs six concurrent research programs, all publishing on the same track with no internal-only results.
Monolith-1 Technical Report
142 pages. Architecture, training setup, evaluation methodology, safety testing, failure modes we didn't solve.
Read the reportPrograms
Six activeAlignment
Scalable oversight, interpretability-informed RLHF, and methods for detecting when a model's objectives have drifted from what we intended. This work runs in parallel with capability development, not behind it.
Interpretability
Mapping internal representations of large models to trace specific behaviors back to specific circuits. Our goal is mechanistic: we want to know why a model says what it says, not just that it tends to say the right thing.
Long-context reasoning
Extending coherent reasoning to one million tokens and beyond. Monolith-1 handles this natively with no retrieval fallback, no chunking, and no measurable degradation past 200k.
Multimodal foundations
Unifying vision and language into a single reasoning substrate. We're not bolting image encoders onto language backbones. We're building a joint foundation from pretraining.
Efficient inference
Reducing the per-token compute cost of frontier-quality reasoning. Our quantization, speculative decoding, and KV-cache compression work ships in every Monolith release.
Governance & policy
Empirical contributions to AI governance debates at the national and international level. We publish all policy research openly and accept no government funding for it.
“We default to openness when there isn't a compelling safety reason to restrict. So far, for us, there almost never has been.”