VectorStorm

Building intelligence that is capable, safe, and good for humanity

We are a research laboratory working at the frontier of artificial intelligence — advancing what AI can do while ensuring it remains aligned with human values, transparent in its reasoning, and beneficial to the world.

Research

What we work on

Our research spans the full depth of modern AI — from the mathematics of learning systems to the practical challenge of deploying them responsibly. We publish openly because we believe the field advances when knowledge is shared.

Alignment & safety

Teaching AI systems to pursue goals that are genuinely beneficial — and to remain aligned even as they become more capable.

Interpretability

Understanding what happens inside neural networks — building tools to read, audit, and verify AI reasoning at a mechanistic level.

Reasoning & planning

Extending the frontier of how AI systems think — improving multi-step reasoning, long-horizon planning, and reliable inference.

Multimodal systems

Building models that can perceive and reason across text, images, code, and other modalities — toward richer, more general intelligence.

Autonomous agents

Researching AI systems that can take actions in the world — and the frameworks needed to keep them reliable, auditable, and controllable.

Evaluation & benchmarks

Developing rigorous methods for measuring AI capability, safety, and behavioral drift — so we know what our systems actually do.

Recent publications

Sparse autoencoders reveal universal features in large language model representations

Chen, M. · Okonkwo, A. · Park, S. · 2026 · NeurIPS

Interpretability

Constitutional AI at scale: reward model robustness under distribution shift

Alvarez, R. · Lindqvist, E. · 2026 · ICML

Safety

Chain-of-thought faithfulness: do models actually follow their reasoning traces?

Watanabe, K. · Osei, P. · Fernández, L. · 2025 · arXiv

Reasoning

Scalable oversight via debate: empirical results on complex question answering

Singh, D. · Müller, H. · 2025 · ICLR

Safety

View all publications

Safety & Alignment

Safety is not a constraint — it is the work

We believe that building AI safely and building AI capably are the same project. Every advancement in capability requires a corresponding advancement in our ability to understand, audit, and align these systems.

Our responsible scaling policy

Before any model is deployed externally, it must clear a comprehensive evaluation suite covering dangerous capability thresholds, behavioral robustness, and alignment properties. We publish our evaluation criteria publicly and update them as our models grow more capable. Capability without accountability is not something we ship.

Read the full policy

"A model that is more capable but less understood is not progress. Understanding must keep pace with capability — that is the discipline we hold ourselves to."

— From the VectorStorm Charter, 2023

Interpretability research

We build tools to examine what AI systems have learned — not just what they produce. Our mechanistic interpretability team works to identify features, circuits, and representations inside our models so we can detect misalignment before it manifests in behavior.

Red-teaming & evaluation

Before deployment, our models undergo adversarial red-teaming by internal teams and external experts. We look for dangerous capabilities, deceptive behavior, and failure modes under distributional shift — then we fix what we find, or we do not ship.

Our safety principles

01

Corrigibility before capability

We build systems that support human oversight and correction at every capability level. A more capable system that resists correction is a more dangerous system.

02

Transparency as a default

We publish our safety research, our evaluation frameworks, and our failures. Closed development of powerful systems concentrates risk; open science distributes it.

03

No deployment without evaluation

We do not release models that have not passed our safety evaluations, regardless of commercial pressure or competitive timing. The policy has no exceptions.

04

Long-term thinking over short-term metrics

We measure ourselves against outcomes that matter in a decade, not a quarter. The incentives of frontier AI development are structurally misaligned with safety — we design our institution to resist that pull.