Frontier labs are making a massive bet on scalable supervision. One part is human supervision: experts writing, ranking, correcting, and annotating model outputs so the system can absorb human judgment at scale. Another part is verifiable supervision: tasks where success can be checked by an external signal, whether code passes its tests, a proof checks out, or an agent completes a reproducible workflow in a simulated environment.

The bet is that enough human judgment data and verifiable rewards can push models toward superintelligence escape velocity. My guess today is that the models hit a ceiling before they generalize their way out of some of the limitations inherent in that recipe.

The Problem with Scaling Human Supervision

Start with the human side. Whether experts are writing demonstrations, ranking outputs, or correcting mistakes, that supervision can only encode the knowledge, calibration, and blind spots of the annotator pool. The resulting models tend to mode-collapse toward consensus, damping down the rare, but correct reasoning. Stack that on top of sticky pretraining misconceptions and you get diminishing returns and no way to certify answers to questions humans can't already adjudicate. This hurts most in low-feedback, contested domains such as macroeconomics, grand strategy, and historical causation.

Where Verifiable Rewards Help and Where They Do Not

RLVR avoids some of that because the reward can come from the world itself. The proof verifies, the code passes, the task completes. But it has its own constraint. It works best where the task is not only verifiable but cheap to replay, reset, and run at enormous scale. A great deal of consequential judgment does not look like that. In investing, history, science, and business-building, the relevant feedback may take years to arrive. None of this proves the current approach fails. Longer contexts, tools, real-world deployment, better continual learning—maybe that's enough to bridge the gap. But the bridge from being excellent at millions of reproducible tasks to reliably noticing when the prevailing frame is wrong, seeking out the right unfamiliar evidence, and learning durably from ambiguous feedback is still a major empirical bet.

2008 Financial Crisis and the Mortgage Market

In 2005, if you'd asked most experts what the odds were that the major banks would melt down because of their exposure to the U.S. housing market, they would have said next to impossible. Ask Michael Burry and you'd have gotten a very different answer. He went through the loan-level detail on subprime mortgage pools, saw what would happen when the teaser rates reset, and started buying credit default swaps against the bonds, a bet almost nobody else wanted to make (Lewis, 2010).

Here's the thing, a model would find one part of this easy and one part hard. Hand it the data and tell it to look, and I think an AI gets to Burry's conclusion pretty quickly. The hard part is deciding that this particular corner of the mortgage market was the thing to obsess over in the first place while every authority was signaling things were fine.

The Civil War: Discovering That the Accepted Frame Was Incomplete

Thirty or forty years ago if you asked historians what caused the Civil War, you would have got real disagreement: states' rights, tariffs, economics, culture, and slavery all in the argument (Woods, 2012). The modern consensus is much clearer: slavery and its expansion were central. Part of how the field got there is worth discussing.

The historian Joseph T. Glatthaar spent years on a statistical study of Lee's Army of Northern Virginia, drawing a random sample of 600 soldiers and running their service records, census entries, pension files, and obituaries (Derewicz, n.d.; Glatthaar, 2008). One finding challenged an assumption core to the argument against slavery as the primary cause. The Confederate rank and file were mostly poor whites with no personal stake in slavery. Glatthaar found that in 1860, while 25% of Southern households owned slaves, 44% of his soldiers' households did (Glatthaar, 2008). Slavery's hold on the South ran through kinship, aspiration, and the household economy, well past the actual slaveholders.

The advance didn't come from a new fact dropping out of the sky with a label attached. It came from creativity; reframing the question and asking how slavery sustained support among people who didn't own slaves, then going to look for evidence in records nobody had bothered to aggregate that way. Moving the consensus took decades. A model trained on a century of conflicting historians and conflicting human feedback carries all of that contradiction in its weights. Could it decide the accepted frame was incomplete and go looking for the missing mechanism before the field agreed? That's the capability I'm unsure about.

The Counterexample: Overcoming a Mountain of Wrong Information with Reasoning

Here's a counter to my argument. There's a number in optics called the Abbe number. It's a function of a material's chromatic dispersion, the colored fringing you notice on high-contrast edges when you pixel-peep a photo. Go searching for advice on choosing the clearest sunglass lenses and you'll find a strong, confident consensus that a high Abbe number is critical for clarity. Marketing leans on it hard.

That consensus is mostly wrong. Past a fairly low bar, the lens material isn't what's limiting clarity. People can't tell the difference. The human eye has its own substantial chromatic aberration and the visual system largely works around it (Campbell & Gubisch, 1967; Thibos et al., 1990).

Ask a current model at low reasoning effort how important Abbe number is for sunglass clarity on a scale of one to ten, and in my own little testing it says eight. Turn the reasoning up, let it work through the actual optics, and it comes back with a two. There's a mountain of shallow, repeated, wrong content online, and more inference-time reasoning lets the model climb over its own bad prior and land where the 1980s and 1990s literature already was. This is one small example, but overturning that bad consensus is what I said the models would struggle with. That said, optics has fixed physics and a settled literature to reason toward, which the 2008 mortgage market and an open historical question do not.

So Where Does That Leave It

None of this proves the labs are wrong. Longer context windows, better tools, real-world deployment that generates fresh feedback, self-distillation, some genuinely new form of continual learning—any of those could bridge the gap I'm describing. Maybe the real bet is different from how I've framed it. Maybe scaling current methods is what lets researchers discover architectures that would otherwise have taken decades. So maybe the current recipe doesn't get us to exit velocity—but it builds the engine that does. IDK. We'll see.

References

Campbell, F. W., & Gubisch, R. W. (1967). The effect of chromatic aberration on visual acuity. The Journal of Physiology, 192(2), 345–358.

Derewicz, M. (n.d.). Life under Lee. Endeavors. https://endeavors.unc.edu/life_under_lee

Glatthaar, J. T. (2008). General Lee's army: From victory to collapse. Free Press.

Lewis, M. (2010). The big short: Inside the doomsday machine. W. W. Norton.

Thibos, L. N., Bradley, A., Still, D. L., Zhang, X., & Howarth, P. A. (1990). Theory and measurement of ocular chromatic aberration. Vision Research, 30(1), 33–49.

Woods, M. E. (2012). What twenty-first-century historians have said about the causes of disunion: A Civil War sesquicentennial review of the recent literature. Journal of American History, 99(2), 415–439.