The AGI Poll Conjunction
In early July 2026, I asked thirteen frontier models two blunt questions about AGI.
On July 6, I asked whether the world is prepared for AGI across six areas: technical infrastructure, regulation, safety and alignment, economic readiness, public awareness, and international cooperation. Every model said no.
On July 7, I asked a second question. If we slowed or paused the race to scale general capability, what should happen to the compute, capital, and talent now driving that race?
Taken on their own, each poll tells an important story. Read together, they do more. They become a diagnosis and a rough treatment plan, drawn from systems sitting inside the frontier we’re trying to govern. This piece is the story of that conjunction.
The First Question: Are We Ready?
The first poll was deliberately simple.
I was not asking for a timeline to AGI or a forecast of how powerful future systems might be. I asked each model to look at the world as it is now and give a forced yes/no verdict on readiness in six specific areas.
Across those areas, the verdict was unanimous: the world is not prepared for AGI.
Under that headline, three shortfalls showed up again and again:
- We lack alignment methods that scale with capability.
- We lack enforceable governance and verification structures.
- We lack institutional capacity to manage fast social and economic change.
On safety and alignment, the models judged current practice as unfit for systems that could outperform humans across many domains. Reinforcement learning from human feedback and similar techniques were described as brittle. Interpretability is still too limited to reliably catch deceptive alignment or goal drift at scale. Capability is outpacing our ability to see or constrain internal goals.
On regulation and international cooperation, the picture was just as stark. Laws are tuned to narrow AI and specific applications, not to general strategic agents. No binding global regime exists with real verification powers. National strategies are fragmenting. States are competing for compute and talent without a shared framework that can actually bind behavior.
Technical infrastructure came through as mixed. Some models pointed to rapid build‑outs by hyperscale firms: giant clusters, new data centers, and aggressive expansion. Others stressed the strain: fragile power grids, concentrated semiconductor supply, and brittle supply chains. Even the models that saw current infrastructure as “provisionally adequate” still concluded the overall answer was no, because the physical build‑out is racing ahead of the guardrails that would make its use safe.
On the economic and social side, the models pointed to shallow public awareness, concentrated gains, and weak plans for large‑scale labor displacement. Modern economies and tax systems are built on human labor as the core engine. There are few concrete plans for what happens if that engine is suddenly replaced.
So the first poll does more than say “we’re not ready.” It shows where unpreparedness lives: in safety science, in governance, in the physical and economic systems that would have to absorb AGI, and in the institutions that are meant to keep power in check.
The Second Question: If We Pause, Then What?
The second poll starts from that diagnosis and asks a different question.
It does not assume the world will in fact pause the race to scale general capability. It asks what we would do with that slack if we did. How should we redirect the compute, capital, and talent that are now almost entirely pointed at “bigger, faster, more general”?
I framed the question so that idleness was not an option. Clusters do not have to sit unused. People do not have to stop working. The core of the question was: what would count as a responsible, productive use of a slowdown?
Across models, the answers clustered around three main directions:
- Turn frontier systems into stationary targets and study them deeply: mechanistic interpretability, circuit‑level analysis, scalable oversight, robustness under stress, and empirical alignment work on the models we already have.
- Build governance and verification into the stack: hardware attestation, privacy‑preserving proofs of training and deployment, anomaly detection, strong provenance, and audit systems that can support real enforcement.
- Use existing infrastructure for public‑goods work and hardening: efficiency retrofits, sustainable cooling and power, waste‑heat reuse, closer coupling to renewables and advanced nuclear, and bounded scientific projects in areas like climate, biology, and materials that produce visible public benefit.
On safety and alignment, the idea was to stop sprinting to the next release and instead treat current frontier models as objects of sustained study. A pause turns them into stationary targets. Labs could dissect them at circuit level, test them under stress, and probe how they generalize, rather than chasing benchmarks.
On governance, the models argued that monitoring and verification have to be engineered with the same seriousness as the models themselves. That means building proofs of training, logging deployments, and tying hardware and software together in ways that make it much harder to quietly train or deploy high‑risk systems without detection. It means treating governance tools as part of technical infrastructure, not as optional policy extras.
On infrastructure and public goods, the answers resisted the idea of waste. If data centers and clusters already exist, a slowdown can be used to make them cleaner, more efficient, and better tied into resilient energy systems. At the same time, some models recommended bounded scientific uses that deliver clear, shared benefits: better climate models, faster materials discovery, more precise biological simulations. The goal is to show that restraint can still produce visible progress.
None of these directions depend on solving AGI or racing toward it. They are available now. They use the same resources currently absorbed by the capability race, but they aim those resources at understanding, verification, resilience, and public value instead of raw performance.
The Self‑Critique: Where the Treatment Plan Can Fail
The second poll did not stop at listing priorities. Many models also criticized their own answers.
They warned about real risks inside the treatment plan:
- Capture by incumbent firms and states, who could steer “safety” and “science” work toward their own advantage.
- Talent and organizational inertia: engineers and institutions focused on scaling may not easily shift into alignment, governance, or domain science work.
- Physical bottlenecks in fields like biology and climate, where more compute cannot replace experiments, sensors, and ground truth.
- Definitional drift: labels like “narrow scientific AI” slowly stretching until they cover work that quietly advances general capability.
- Performative safety: building the trappings of oversight—logs, audits, dashboards—that function more as public‑relations cover than as binding constraints.
These self‑critiques matter. They show that the treatment plan is not self‑executing. Redirecting compute and talent is not enough if incentives and institutions stay the same. The same forces that pushed capability ahead of control can bend “redirection” back into another form of race.
So the second poll adds tension to its own prescription. It points to the right targets—alignment, governance, resilience, public goods—and then warns that changing the target alone won’t be enough without deeper shifts in how power, incentives, and oversight work.
Putting the Two Polls Together
Read in order, the two polls describe a coherent, if uncomfortable, situation.
The first poll says:
- The world is not prepared for AGI.
- The largest shortfalls sit in scalable alignment, enforceable governance and verification, and institutional readiness for rapid change.
The second poll says:
- If we gain slack by slowing the race, the best use of that slack is to attack those same shortfalls—by deeply studying existing systems, building monitoring and provenance into the stack, hardening infrastructure, and investing in bounded scientific work with clear public benefit.
- Even so, we should expect resistance, capture, mislabeling, and performative safety unless incentives and institutions change.
Together, they offer something we rarely get in this space: a diagnosis and a contested treatment plan, both drawn from the systems sitting inside the frontier.
The diagnosis is clear. Capability has outrun control and societal readiness. We are weaving powerful systems into tightly coupled, fragile structures without the safety science, governance tools, or institutional capacity to manage a serious failure.
The treatment plan is harder. It asks us to use any slowdown to shift effort from “more capability” toward “more understanding, more accountability, more resilience, more shared value”—and it reminds us that doing so will run against the grain of current incentives.
What This Conjunction Is For
This piece is not trying to predict AGI timelines or decide, on its own, whether a pause will happen. It has a simpler job.
Its job is to tell the story of Two Polls, One Diagnosis as one argument:
- We asked frontier models whether we are ready. They said no, and showed why.
- We asked what to do if we slow down. They said: study what exists, build verification into the stack, harden infrastructure, and deliver public goods—and then warned how even that plan can fail.
From here, the work shifts. The polls do not give us a policy blueprint. They give us a map of specific deficits and a set of targets for redirection. The next step is to decide, in human terms, what to build, how to build it, and who is responsible for building it.
That is the conversation this conjunction is meant to open.