Head of Pre-training.

The lab had reached the point where pre-training could no longer be run as a collection of strong individual contributors. The next training run sat in a compute tier where data, infrastructure, evaluation and the research direction itself had to be coordinated under a single owner who could carry the technical and operational accountability. The CTO wanted a Head of Pre-training who had actually run a frontier-scale pre-training programme, not someone who had supervised one from a research-director seat.

The constraints were absolute. The brief explicitly excluded candidates whose involvement in a frontier training run had been at the readout level rather than the execution level. The Head of Pre-training had to be credible with the research scientists, the training-systems engineers, and the lab's compute partners. They had to be willing to take a seat that operated below the public surface of the lab — the role was deliberately not press-facing. Compensation reflected the bench's scarcity and the lab's stage. Geography was open across the lab's two main research hubs.

The senior pre-training bench is tight and well-known to itself. A small number of researchers and engineers have actually run frontier-scale pre-training programmes; most are in seats they are not leaving on standard terms. Below that tier sits a wider group of senior research leads who have owned a major sub-component — data, evaluations, training systems, post-training — without ever owning the full stack. They are strong, but they are not yet ready for this seat.

Our read was that the search would resolve through reference networks rather than search activity. The people who could do this role were two or three references away from every other person who could do this role; the assessment work was therefore as much about which references to weight as about which candidates to interview. We invested heavily in the opening month in mapping the reference graph properly. Peter led every first conversation and the lab's CTO joined assessments from the second round.

The shortlist drew from senior researchers and engineers who had owned a frontier-scale pre-training run, and from a small number of senior sub-component owners ready to take the full stack at a smaller-scale frontier programme.

• Senior researchers who had owned a frontier-scale pre-training run end-to-end.
• Senior training-systems leaders ready to step into the research-and-engineering owner seat.
• Heads of data or evaluations from tier-one labs ready for the full stack.
• A small set of crossover candidates from elite academic groups with frontier-lab placements.

The hire was a senior researcher who had owned a frontier-scale pre-training run inside a competing lab and was ready to step into a seat with more end-to-end authority than they had carried before. What made them right was less the run history than the way they framed the next eighteen months in the second meeting: not as a scale-up but as a sequencing problem about data, infrastructure and evaluations that had to be solved before compute was committed. The CTO and the lab's founders converged on the same finalist inside a single discussion.

The close was deliberate. Compensation took two passes to settle, with the candidate's existing equity package re-priced against the lab's next planned valuation event. The offer signed on the fourteenth week, two days before the candidate's previous lab announced its next training run.

“The lesson was about reference weight. In a bench this small, the conventional outbound search produces noise; the work is in deciding which two or three references inside the field carry weight on which two or three candidates. Spectrum's role in week one was to build that reference graph honestly — including the references who would tell us where someone's involvement had really sat in the stack — and to let it drive the shortlist rather than the other way around. By the time we presented the finalists, the assessment was already a foregone conclusion.”

— Peter Wood