The constraint running through every serious debate about artificial general intelligence is a sequencing problem: capability research has historically moved faster than the safety evaluation infrastructure meant to govern what gets released. Shane Legg, a co-founder of DeepMind, has formalized that concern into an institution. Legg has launched a new institute dedicated to exploring the deployment of artificial general intelligence, adding his weight to growing calls for the industry to slow down.
His warning is blunt: AI development must not outrun its safety controls. The new institute is organized around that constraint, treating AGI deployment as a dedicated research domain rather than a secondary consideration that follows capability work.
The sequencing problem Legg is addressing has a structural dimension. Safety evaluation work requires its own tooling and benchmarks, developed through different timelines than the model training that produces frontier systems. The gap that results tends to widen as the frontier advances.
Legg's decision to build an institute rather than issue a public statement matters in its own right. It signals a commitment to sustained, methodical research and places him among senior figures from the industry's founding generation who are now focused on what happens after AGI arrives rather than whether it will.
The calls for a sector-wide slowdown that his institute enters are themselves growing in volume. Whether an organization focused on deployment methodology can materially shift how the industry handles the safety-capability gap is the question Legg's new institute now has to answer.