Memo: Sam Altman says OpenAI's next model finished pretraining, and moves Safety to Research and Security to Scaling; Fidji Simo becomes CEO of “AGI Deployment”
Alex Heath /Sources:
Context & Ripple Effects
OpenAI had already separated commercial leadership from its core technical agenda when it hired Fidji Simo as CEO of Applications, while Altman retained research, compute, and safety responsibilities. This memo extends that division: Simo’s remit is now framed as AGI deployment, while the next model has cleared its pretraining phase.
The organizational changes also give operational form to Altman’s reported shift toward fundraising, supply chains, and data-center buildout, following his transfer of safety and security oversight. Safety moves closer to model research; security moves closer to the systems responsible for scaling access and infrastructure.
First-order effects
- OpenAI’s next model advances from pretraining into its subsequent development and release-preparation stages, placing immediate emphasis on research, evaluation, and deployment decisions rather than raw training progress.
- Simo takes ownership of AGI deployment, while Safety joins Research and Security joins Scaling—clearer reporting homes for teams previously overseen by Altman.
Second-order effects
- The split makes deployment a distinct executive function from frontier-model development, building on Simo’s earlier appointment to lead applications and potentially concentrating product, customer, and rollout decisions under one organization.
- Embedding security in Scaling ties security work more directly to compute and data-center expansion; embedding safety in Research ties it more directly to the model-development cycle.
Third-order effects
- If sustained, the reorganization points to a frontier-lab structure in which research, industrial-scale infrastructure, and deployment are separately managed but tightly coordinated, rather than concentrated under a single CEO.
- That structure may make governance increasingly operational: safety evaluation is integrated into research, while security is treated as a prerequisite for reliably scaling model access and infrastructure.
The trend: This is one data point in the institutionalization of frontier AI labs as they divide research, deployment, and infrastructure into specialized executive domains.