Post-scaling AI systems
If raw scale stops being the dominant axis of progress, the frontier moves to what surrounds the model: memory, tools, environment, and the systems that compose them.
The scaling assumption
For most of the last decade, capability has tracked computation. That correlation structured the entire field: the labs that could spend the most could discover the most. If the correlation weakens — for economic, physical, or data-availability reasons — the frontier stops being a single axis and becomes a design space.
A design space favors teams that can think, not just spend. That is the premise this direction studies.
Where progress could move
The candidates are well known individually: persistent memory that outlives a session; tool use that is reliable enough to delegate to; environments that give systems something to act on; verification loops that let a system check its own work. What is not known is how they compose — which combinations produce capability that scales with engineering effort rather than training compute.
Our angle is systems-level: treat the model as a component and study the architectures around it. Small, carefully instrumented systems that fail for legible reasons teach more than large ones that succeed for opaque reasons.
Position
This is an exploratory direction with a deliberate posture: we do not claim scaling is over, and we do not bet the lab on either answer. We build systems that are valuable under continued scaling and more valuable if it slows. That asymmetry — useful in both futures — is itself the design principle.