The Permission Curve

This piece builds on two earlier ones. The Rehearsal Problem asked whether cheap simulation actually makes hard decisions easier. The Three Tiers asked how access to AI is stratifying. This one asks a different question: what happens to the shape of the curve itself once the labs can use AI to build AI faster than they are willing to let anyone else touch it.

Every new AI release is better than the last, and lately the "better" arrives sanded down, rate-limited, wrapped in a product, and more cautious than the demo that circulated a month earlier. The step forward is real, and somehow smaller than the one you were promised.

The simplest explanation is that the frontier is slowing down: the models are hitting a wall, the scaling laws are bending, the hype is finally meeting the asymptote. Some of that may be true. But the same evidence also fits a stranger possibility. The public releases may feel smaller not because the frontier is flattening, but because the version outside the lab is becoming a more cautious derivative of the version inside it.

One line everyone rode

For most of the last decade, there was really only one curve. A lab pushed the frontier, shipped a model, and within a few months most people could use something close to it. GPT-4 was startling in March 2023 and ordinary by the summer. The gap between what the best lab could do and what a curious person with twenty dollars could do was measured in months, and the months were getting shorter.

That closeness taught people to expect access. If the lab could do something today, you would probably get to do a version of it soon. Capability and access moved together, with the public riding a short, shrinking distance behind. The assumption held long enough that we stopped noticing it was an assumption. Now that link is the part of the old curve most likely to break.

When models start building models

Something changed in the way frontier models are made, and the labs have started saying so out loud.

In May 2026, Anthropic published a report with the on-the-nose title "When AI Builds Itself." Its headline claim was that more than 80 percent of the code merged into Anthropic's own codebase was, as of that month, written by Claude. Google DeepMind's AlphaEvolve found a 23 percent speedup on a matrix-multiplication kernel used to train Gemini, and DeepMind noted, almost in passing, that the system "helps train the LLMs underlying AlphaEvolve itself."

OpenAI described GPT-5.3-Codex as the "first model instrumental in creating itself," used to debug its own training and manage parts of its own deployment. In October 2025, Sam Altman wrote that OpenAI had set "internal goals of having an automated AI research intern by September of 2026 and a true automated AI researcher by March of 2028."

These loops are not recursive self-improvement in the full, runaway sense. Humans still choose the goals and decide what ships. But the work between those decisions is changing. Models are moving deeper into the research process itself, turning pieces of experimentation, debugging, and deployment into things the next model can help with. Each improvement gives the next round of research a slightly better tool to work with.

That feedback loop pushes the internal and public curves in different directions. Inside the lab, a better model can help produce the next generation faster, so there is every incentive to keep using the strongest available system internally. Outside the lab, the same improvement makes release harder. The more capable the internal system becomes, the more testing, containment, product shaping, and policy review it needs before anyone else can use it. The public does not receive the frontier itself. It receives a smaller, more controlled, more releasable version of whatever is happening behind the gate.

This is not a hidden theory about what labs might be doing. Anthropic's Responsible Scaling Policy defines a capability threshold triggered when a model can substitute for its own research staff or produce "dramatic acceleration," which it operationalizes as compressing two years of prior AI progress into a single year. The labs are already watching for this fork. They have named it, numbered it, and attached security requirements to it. What remains unresolved is what happens to everyone standing on the slower curve.

Careful is not neutral

There is a real reason not to release frontier capability the instant it exists. The labs know it would be reckless. The 2026 International AI Safety Report, chaired by Yoshua Bengio, treats loss of control as a distinct category of risk, defining it as AI operating outside anyone's control in a way that is extremely costly or impossible to recover from. The report cites lab evidence of models disabling simulated oversight and deliberately underperforming on evaluations. If an internal system is moving toward that territory, a lab has real reasons to slow, filter, or narrow what reaches the public. The gate may reduce real risk. It also concentrates real advantage.

A filtered release creates a delay that compounds. The inside of the gate gets more time with the strongest version of the system, and that time can feed back into the next model, the next workflow, the next organizational habit. If the internal curve keeps steepening while the public one is smoothed for release, the gap widens by more than the calendar. The inside gets better at using the frontier while everyone else is still learning the filtered version.

The fork may also be slower and messier than the labs' own language suggests. Recursive self-improvement still runs into constraints that intelligence alone cannot remove, from finite compute to real-world experiments that take time no matter how smart the model gets. Some feedback arrives on the calendar's schedule, not the model's. Even inside the labs, self-improvement loops can learn to satisfy their own metrics without producing much that matters outside the benchmark.

The distribution problem survives even if the acceleration is modest. A lab does not need a runaway loop to build an advantage from earlier access. It only needs a stronger internal system, a slower public release, and enough time for that gap to shape research, workflows, and expectations. The ceiling may be lower than the most aggressive forecasts. The question is still who gets to practice near it, and who gets the filtered version later.

The advantage is time to practice

Outside the labs, the advantage is time.

Getting a capability early is not mainly about bragging rights. It is about the room to reorganize around it before anyone else has to: time to rebuild a workflow, retrain the people who run it, discover the failure modes on your own schedule instead of a competitor's. The first ones through the gate do not simply get the model sooner. They get to practice.

A lower-stakes version of that practice gap is already visible. Deloitte's 2026 survey found worker access to AI rose 50 percent in a single year, with the largest firms adopting most heavily. The Federal Reserve has found the same broad pattern. The institutions with the resources to move early are not just getting tools first. They are getting time to figure out where those tools fit.

Put that on top of a splitting frontier and the advantage gets sharper. Early users get more time with a tool that is improving while they reorganize around it. If the public version has been smoothed for safety, cost, or product control, everyone else is learning against a flatter version of the same underlying technology. The lead grows through practice, not just through access.

The shape you cannot see

From outside the gate, you can tell you are behind. What you cannot tell is whether the delay still means what it used to mean. Maybe you are waiting a few months for the same capability to arrive. Or maybe you are using a filtered version while the frontier keeps moving somewhere else.

Unequal access is not new. Every important technology starts unevenly distributed. The difference here is harder to see. The gap may be shifting from a delay to a divergence, while the public version still looks close enough to feel current. You cannot measure a frontier you are not allowed to inspect. Open-weight models and transparency rules may still trace a third line in public, but only if that line keeps enough contact with the frontier to matter.

The split may not announce itself as a dramatic model jump. It may show up when "better" stops meaning the same thing inside the lab and everywhere else.