OpenAI commits 20% of its compute to superalignment
The pledge to devote a fifth of secured compute over four years was later disputed by the team's own co-lead, who said requests for GPUs were repeatedly refused.
- Safety & alignment
- Major
OpenAI announced a new team, Superalignment, dedicated to solving the technical problem of controlling AI systems that might become more capable than the humans supervising them. Chief scientist Ilya Sutskever and researcher Jan Leike co-led the effort, which OpenAI gave itself four years to complete. The company wrote that it was “dedicating 20% of the compute we’ve secured to date over the next four years to solving the problem of superintelligence alignment.”
The framing was notable for treating superintelligence as a near-enough prospect to warrant a dedicated, well-resourced research programme rather than a speculative side project, and for putting a specific, checkable number — a fifth of the company’s compute — behind that commitment. The team’s stated approach centred on building an “automated alignment researcher”: using AI systems themselves, at roughly human level, to help evaluate and align systems beyond human level, on the theory that alignment research itself could be scaled the way capabilities were being scaled.
The commitment did not survive contact with the rest of the company. According to later reporting, drawing on multiple people familiar with the team’s operations, the Superalignment team’s compute budget never came close to the promised 20%, and repeated requests for GPU access were turned down by OpenAI leadership. Leike, resigning ten months later in May 2024, wrote publicly that his team had been “sailing against the wind” for the resources it needed and that safety culture had “taken a backseat to shiny products.” The team was dissolved days afterward, with remaining staff redistributed elsewhere in the company — closing, within a year, an effort framed at launch as a four-year project.