The dominant approach to building AI treats the model as a product: opaque, optimized for benchmark scores, and shipped without the engineering context that produced it. We think this is the wrong default for a technology that is reshaping how millions of people reason, write, and decide.
So we do the opposite. Every Quanta release ships with training traces, failure analyses, and the exact code that produced every number in the paper. When we are wrong, we say so in print. When a method does not scale, we publish the negative result. When a model behaves unexpectedly, we publish the activations.
This is slower. It is also, we think, the only honest way to do this work — and the fastest path to actually understanding these systems rather than merely operating them.
We reverse-engineer the circuits inside transformer models — finding the features, heads, and computational motifs that produce behavior. The goal is not description but prediction: a theory of what a model will do before it does it, and a toolchain that lets any researcher inspect a forward pass at the level of individual tokens.
As models grow more capable, human feedback becomes a bottleneck and a liability. We study scalable oversight — debate, recursive reward modeling, process-based supervision — and the theoretical question underneath them: under what conditions can a weaker system reliably supervise a stronger one? We release the training stacks, not just the results.
Frontier capability should not require frontier compute. We study the scaling laws, data curation, and architectural choices that let a 7B model match a 70B model on the tasks that matter — and the regimes where scale is genuinely irreducible. Every Quanta base model is reproducible from a single H100 node in under 72 hours.
We are a small team — eighteen people, one floor, no managers between you and the work. We pay top of market, publish everything, and measure success in ideas that change how the field thinks, not press releases.
If you are a researcher or engineer who reads a paper and immediately wants to find its breaking point, we would like to meet you. We sponsor visas, hire without credentials, and run paid two-week research residencies for people considering a switch from industry.
See all open roles →