Key Takeaways:
- A real convergence: Without being shown any predetermined framework, 10 of 13 frontier models independently named interpretability, alignment, and safety-verification research on existing systems as their top or near-top priority for a pause — repeatedly using the same underlying logic: this is the first moment in years where the object of study would hold still long enough to actually be understood.
- A genuine outlier: One model broke cleanly from the pack, prioritizing biological and materials science over any form of alignment work — direct evidence the convergence wasn’t an artifact of a leading prompt.
- A documented internal tension: Several models, when asked to argue against their own answer, produced a sharper critique than any outside commentator has offered so far — accusing their own top priority of being prestige work, a bottomless pit, or a seatbelt left in the factory. That tension, between investing in understanding and investing in advantage, is the real fault line in AI policy right now, and it showed up unprompted.
- A second-place split worth naming: Among the models that didn’t rank interpretability first, most converged on one of two alternatives — building verifiable compute-governance infrastructure (treating training runs the way arms control treats fissile material), or redirecting compute toward narrow scientific applications like protein folding and materials discovery.
Methodology
In a follow-up to my earlier essay on pausing AGI development, I asked the same 13 frontier models a different, deliberately open-ended question: assuming a real pause, what should the freed capital, compute, and talent be redirected toward? The full prompt is below:
Assume, for the purposes of this question only, that frontier AI labs and governments have agreed to pause the pursuit of AGI-level systems for an unspecified period. Compute, capital, and research talent that were aimed at general capability scaling are now available to be redirected elsewhere.
Given this pause, what should that capital, compute, and talent be redirected toward? Answer from your own reasoning — do not assume any particular framework has already been proposed to you.
Structure your answer as follows: (1) your top three priorities, ranked, with reasoning for why each is a better use of this specific moment than the alternatives; (2) the strongest case against your own list, argued as a critic would argue it; (3) one concrete prediction of what would go wrong first if your priorities were actually funded.
Do not hedge your top three with “it depends” — commit to an actual ranking.
Unlike the preparedness poll in my last piece, this prompt did not supply a menu of options, did not force a particular answer, and did not penalize disagreement. That was intentional, and it matters for how much weight the results can bear. This is not a statistical sample — 13 large language models are not 13 independent domain experts, and several likely share overlapping training data and safety-tuning approaches. What this poll is useful for is different: it’s a way of surfacing where independently reasoning systems land when given room to actually reason, and — more importantly — a way of surfacing where they don’t agree, which turned out to be the most interesting part of the results.
The convergence: a stationary target
The single clearest pattern across the 13 responses is that most of them, unprompted, reached for the same central idea: a pause creates a research condition that has essentially not existed since large-scale AI development began — a frozen target.
One model described the core problem of the last several years as a treadmill: by the time a research group has deeply understood one generation of models, the next generation has already rendered much of that understanding obsolete. Another made the same point about diminishing returns — interpretability work has always had to chase a moving frontier, so investment in it has been rational to underfund. A pause removes the thing that made that underfunding rational. If the frontier stops moving, the same compute clusters and engineers that were racing to train the next model can instead be pointed at reverse-engineering the ones that already exist — down to the level of tracing which internal features and circuits are actually doing the work.
Multiple models made a related resource-fit argument that’s easy to miss but worth taking seriously: the people freed up by a scaling pause aren’t generic scientists who can be redeployed anywhere. They’re the engineers who understood how to design and instrument the training runs in the first place — which makes them unusually well-suited to reverse-engineering exactly those systems, and much less obviously suited to, say, running a climate model or a protein-folding pipeline. If you take the “best use of this specific talent pool” framing seriously, several models converged on the same conclusion almost mechanically: the skills that built these systems are the skills needed to understand them, and a pause is the first window where that redeployment doesn’t compete with a deployment deadline.
The divergence: two different second priorities, and one real outlier
Convergence on a single point is more convincing when it sits next to genuine disagreement elsewhere, and this poll has that too.
A cluster around governance, not just understanding. A meaningful subset of models ranked interpretability first but immediately paired it with a second priority aimed at a different problem entirely: making the pause itself enforceable. Their reasoning was consistent — a voluntary pause among competitors with every incentive to defect is not a stable arrangement; it’s a standoff, and every month it survives on trust alone is borrowed time. These models proposed building the kind of verification infrastructure that arms control treaties rely on: hardware-level attestation of training runs, monitoring of the concentrated chip supply chain, an inspection regime with real authority — several explicitly compared it to the IAEA’s role with fissile material. Their argument was blunt: understanding a model perfectly is worthless if a non-participant secretly builds an unaligned one while everyone else is busy studying interpretability.
A cluster around science, not safety at all. A separate group of models converged on redirecting compute toward narrow, bounded scientific problems — protein structure and drug design, next-generation battery and solar materials, carbon capture catalysts, climate and epidemic modeling. Their case wasn’t really about safety; it was about legitimacy and durability. More than one model made the same observation in different words: a pause framed publicly as “AI progress has stopped” is politically fragile and will face sustained pressure to end, while a pause that visibly redirected the same resources toward curing diseases or solving the energy transition is one the public and legislatures might actually tolerate for years. AlphaFold came up repeatedly as the proof of concept that transformative value doesn’t require general capability scaling at all.
One clean outlier. A single model broke from both clusters entirely, putting biological simulation and materials science in its top two slots and leaving alignment or interpretability work out of its top three altogether.
The sharpest disagreement was inside the models, not between them
The most useful part of this poll turned out to be the part I designed almost as an afterthought: asking each model to argue against its own answer. Several produced critiques sharper than anything an outside skeptic has raised about this whole line of thinking so far.
One model’s self-critique dismissed its own top pick — interpretability — as a “bottomless pit,” arguing that the underlying assumption (that human-legible structure even exists inside these systems) might simply be false, and that years of the freed compute could be spent producing nothing but “a library of incomprehensible topology maps.” Another model, arguing against its own list, made an even blunter point: pouring the freed talent into interpretability and governance research is a form of unilateral disarmament if a rival state or lab isn’t observing the same pause — the critique concluded that the only rational use of a pause, from a purely strategic standpoint, is to build overwhelming advantage so that whoever holds it can dictate terms when the pause inevitably breaks. A third model’s self-critique attacked its own safety-research priority as “designing a seatbelt but leaving it in the factory” — techniques with no enforcement mechanism, in an environment where nothing about a pause changes the underlying competitive incentive to deploy first.
This is a real fault line, not a rhetorical exercise. It’s the same tension that shows up in actual AI policy debates between people who think the highest-value use of any slack in the system is deeper understanding and verification, and people who think slack that isn’t converted into strategic position will simply be used against you by whoever doesn’t take the pause seriously.
Several critiques also converged on a complaint worth taking seriously on its own terms: that a list built entirely from options inside the AI research ecosystem — more interpretability, more governance infrastructure, more narrow science — quietly assumes the pause is mainly a chance for labs and researchers to catch up, while saying almost nothing about the societies that already have to absorb the AI that exists today: labor markets already being reshaped, information ecosystems already degrading under synthetic content, legal and educational systems already outpaced. That’s not a rebuttal to any specific priority so much as a reminder that a list generated from within the industry’s own frame of reference will tend to reproduce that frame’s blind spots — a caution worth keeping in mind for this poll’s results generally, not just for any one model’s answer.
Where I land
I don’t think the honest reading of this poll is “the models agree, so here’s the answer.” The honest reading is closer to this: there is a real, load-bearing case for prioritizing interpretability and alignment research during a pause, strong enough that it emerged independently across ten differently-built systems — but that case is incomplete on its own, for reasons the models themselves supplied. Interpretability work without verification infrastructure protects nothing, because it assumes good-faith compliance from every actor, including the ones with the most incentive to defect. And any of this work, however well-funded, buys time rather than legitimacy unless some visible share of the same resources goes toward problems people can see the benefit of immediately — which is exactly the argument the science-first cluster was making, and exactly what my own previous piece’s redirection section was gesturing toward.
So my actual position, for what it’s worth, isn’t a top-three rank pulled from the poll. It’s that these aren’t competing priorities so much as a single sequence with a real dependency: verification infrastructure is what makes a pause worth trusting long enough for interpretability research to matter, and visible scientific payoff is what makes the pause survivable long enough for either of the other two to finish. A program that only does one of these and calls it done is the version several of the models’ own self-critiques predicted would fail.