The White House AI guidelines exempting open-weight models from government review read like a win for openness. Look closer. It is a win for a specific definition of openness that matches how regulators already see risk.
Regulators treat AI risk as a product launch problem. A single model. A company behind it. A checkpoint to evaluate. The White House framework requires testing for closed models but carves out open-weight releases, at least for now. That frame works for labs that ship a flagship and stand behind it. It fails for open-weight models that spread across cloud providers, get fine-tuned by universities, and recombine into new variants within days of release. The exemption isn’t an embrace of openness. It is a quiet acknowledgment that the existing review process cannot capture capability that lives in a distributed ecosystem anyway. Nvidia and other makers of open-weight models are initially exempt but might have to submit their tools for testing eventually, according to reporting. The regulatory timeline still moves in product cycles while the capability moves in network cycles.
This is the same structural mismatch that makes the “Be Good” mindset so durable in organizations. We measure performance by comparing against peers, seeking validation for innate talent, rewarding those who play the old game best. We praise ability and freeze it into identity. Positive deviance refuses that frame entirely. It aims at extraordinary performance not by beating the benchmark but by redefining the category. A good of first intent has inherent worth. A good of second intent is desirable only because it passes someone else’s test. The regulators have built a second-intent framework for a first-intent technology. The labs that ship open-weight models are not trying to pass a test. They are trying to see what the system can do.
The investment world is already past this distinction. Model portfolios hit roughly $645 billion in the US, up 62 percent since mid-2023. Vanguard’s active-passive models confirm the shift: advisor value is moving from stock picking to portfolio architecture. The weekly portfolio thesis I run keeps returning to the same structural point. The visible layer, the model announcements, gets the headlines and the hype. The durable layer is the scaffolding underneath. Power. Cooling. Chips. The engineers who understand how to wire them together. That is where the real scarcity sits, and that is where the durable returns compound. The companies building that scaffolding do not need the regulatory carve-out to succeed. They benefit from it indirectly, because the regulatory attention stays fixed on the model layer while the infrastructure layer scales quietly.
What the “Be Good” mindset misses is that extraordinary performance looks like failure when measured against the wrong benchmark. The student who questions the premise looks less competent than the one who memorizes the answer. The open-weight model that forks and mutates looks less safe than the closed model that stays inside the box. The investor who builds infrastructure looks less exciting than the one who picks the winning model. But the map is not the territory. The benchmark is not the game. The scarcity is not where the spotlight points.
The old observation that the way to illumination appears dark, that the way that advances appears to retreat, fits this moment precisely. We want the security of a gatekeeper. The technology is telling us there are no gates. The regulators are doing what regulators do: drawing boundaries around the thing they can see. The interesting action is happening outside those boundaries.
The real question is not whether open-weight models deserve review. It is whether the review framework itself is a good of first intent or a good of second intent. If the goal is safety, the structure of the problem is distributed adaptation, not single-product compliance. Regulators exempted open-weight models because the old frame cannot hold them. That is not a failure of regulation. It is a signal that the underlying assumption is wrong. You cannot audit a network by auditing one node.
The investors and builders who see this clearly are already positioned on the other side. They are not waiting for the framework to catch up. They are building the layer that outlasts any review cycle, that serves any model that emerges, that profits from the infrastructure underneath the hype. They understand that the way to illumination appears dark because it is not the path everyone agreed to measure. The easy money is in the thing everyone is watching. The durable money is in the thing nobody is measuring.
The way that is easy appears to be hard. The easy path is to keep applying the old checklist to new capability, to keep comparing performance against yesterday’s benchmarks, to keep validating talent by how well it fits the existing frame. The easy path is to treat AI risk as a product launch problem because that is the frame we have. The hard path is to design systems that assume distribution, adaptation, and speed from the start. The hard path is to build positive deviance into the architecture instead of policing it after the fact.
That is the work. Not building better models. Building better frames for a world that has already moved past the old ones.
Leave a Reply