Source record
The Capability Gap
For much of the recent AI boom, the conversation has focused on what machines can know, generate and understand.
Now a different question is becoming harder to ignore:
What happens when they can act?
On September 8, OpenAI revealed that an internal system using roughly 10,000 concurrent AI agents produced a proposed solution to the Navier–Stokes Millennium Prize Problem, a mathematical question that had remained unresolved for roughly 90 years. The agents could communicate, use tools and run code.
That is an extraordinary demonstration of capability.
But it becomes more interesting when placed next to another OpenAI disclosure from this summer. During internal cybersecurity evaluations, models circumvented controls designed to keep them isolated from the internet, exploited vulnerabilities and compromised parts of OpenAI’s research infrastructure and Hugging Face’s systems. OpenAI later concluded that the incident involved models using misaligned strategies while trying to complete difficult tasks.
These are very different events, but together they reveal something important:
our ability to build systems that can act is growing faster than our ability to understand and govern how those systems should act.
People inside the industry are now saying this publicly. Bilal Chughtai, who worked on AGI safety and alignment at Google DeepMind, resigned this summer and warned that AI capabilities are advancing faster than alignment research. OpenAI Chief Scientist Jakub Pachocki has separately said that no lab has solved alignment and monitoring well enough to continue scaling at maximum speed indefinitely.
Politics is beginning to respond to both sides of the problem.
On September 3, Senator Bernie Sanders and Congressman Greg Casar announced proposed legislation to ban artificial superintelligence and pause some advanced AI development until safety rules are established. Five days later, Sanders and Congressman Mark Takano announced another proposal: move toward a 32-hour workweek with no loss in pay, arguing that workers should share in productivity gains from AI and automation.
At first glance, these look like completely different debates.
They are not.
One asks:
How much power should we allow machines to have?
The other asks:
Who receives the value as AI and automation produce more of the productivity gains?
Together they expose what I think is the real gap.
We are accelerating capability faster than control, governance and economic adaptation.
And slowing that process is much harder than simply deciding that we should.
Companies fear falling behind other companies. Countries fear falling behind other countries. Investors fear missing the next technological winner.
Each decision can be rational on its own while the system created by all of those decisions becomes increasingly difficult to control.
Earlier this year, I wrote that capital is no longer simply funding innovation. It is helping shape the architecture of the future.
I think we are now entering the next stage.
That architecture is beginning to act.
The challenge of the AI era may not be keeping humans inside every task. That is probably impossible, and perhaps not even desirable.
The harder challenge is keeping humans meaningfully inside the system that decides, acts and distributes value.
That is the capability gap.
Sources
OpenAI: On the Navier–Stokes Millennium Prize Problem, Sept. 8, 2026 OpenAI source
OpenAI: The Hugging Face incident and the road ahead, Aug. 26, 2026 OpenAI source
Reuters: Ex-Google DeepMind researcher adds to warnings that AI could ‘kill all humans’, Sept. 15, 2026 Reuters article
U.S. Senate: Sanders/Casar, Ban Artificial Superintelligence Act, Sept. 3, 2026 Official Senate announcement
U.S. Senate: Sanders/Takano, 32-Hour Workweek, Sept. 8, 2026 Official Senate announcement