The ASI transition: naive, or naive squared?
Vitalik Buterin read the plan for surviving superintelligence, read its critics, and refused to join either side — because he says nobody honestly knows which world we are in. That refusal is the most useful thing anyone has said about AI this month, and it points at what is actually worth owning.
@VitalikButerin · @DKokotajlo · @eli_lifland · @binarybits · @adi_baradwaj · @RyanGreenblatt · @banteg · @ylecun · @zooko
The chain of the argument
Daniel Kokotajlo and Eli Lifland — the authors of the AI 2027 forecast, which predicted AI takeover or irreversible power concentration — published AI 2040: Plan A, a plan for international coordination toward a good outcome instead. Every scenario in it has superintelligence arriving by 2040 unless an extraordinary effort stops it.
The detractors arrived fast. Timothy B. Lee called the whole thing so implausible he didn't know where to start, and named the real problem: an epistemic chasm between people who think superintelligence implies near-omnipotence and people who don't. Adi Baradwaj sharpened it into a specific charge — a motte-and-bailey between "Big-S Superintelligence," which is omnipotent by definition, and "small-s superintelligence," the plausible end state of current research. banteg put the same objection in the register it deserves: draw a line up a scaling graph, the line reaches heaven, conclude capitalism ends in 2031.
Vitalik's move was to refuse the frame. He points out that the detractors are not being symmetrically skeptical: they call the plan naive about human coordination, but see no naivety in assuming the ASI transition just goes fine by default, no power-concentration risk in ASI itself, and no problem with humanity's hard power going to zero. That stance is coherent in a world where AI is normal technology. It is incoherent in a world where superintelligence lands by 2030. And he says he does not know which world this is.
But currently, I see zero plans for how to deal with an ASI transition that are not naive. Perhaps humanity is stuck with a choice between naive and naive squared.
The two sides
The people who forecast takeover now publish a positive plan instead of a warning — and it mandates open source and spreads the kill-switch across several actors rather than one.
@DKokotajloThe plan exists to force alternatives into the open: treat superintelligence with the seriousness it deserves, and let rival plans compete.
@eli_liflandThe asymmetry does the work. If AI is merely normal technology, Plan A's downside is modest and most of it can wait. If ASI by 2040 is even plausible, the upside dominates. You don't need confidence — you need the bet to be lopsided.
@RyanGreenblattThe definition settles it: if a machine does every task better than a human plus a machine, there is no economic reason to pay a human. Saying "there will still be jobs" is not optimism about ASI — it is disbelief in ASI.
@VitalikButerinThe plan is implausible enough that there is no obvious place to begin. The disagreement isn't about policy, it's an epistemic chasm about whether superintelligence implies near-omnipotence at all.
@binarybitsA motte-and-bailey is being run between two different words. "Big-S Superintelligence" is omnipotent by definition; "small-s superintelligence" is where current research might land. The argument only works if you quietly swap them.
@adi_baradwajThe reasoning is a scaling graph extrapolated into eschatology — by the same people who spent a decade explaining that central planning fails, except when they are the planners.
@bantegA third position that refuses both: the danger is not the machine, it is a handful of firms owning proprietary assistants. The answer is open-source foundation models, not a coordinated pause that hands the same firms the switch.
@ylecunThe exchange, in their words
Vitalik's actual proposal, and why it is not a compromise
The part that got summarized as "Vitalik proposes a pause deal" is subtler than that. Borrowing from Robin Hanson, he argues the winning deal is one both camps would accept from their present beliefs, for opposite reasons: pre-agree on concrete triggers — a super-pandemic, unemployment above 25%, something involving slaughterbots — and pre-agree that crossing them moves everyone toward a slowdown. The skeptics sign because they expect the triggers never to fire. The worriers sign because they expect they will. Nobody has to be converted first, and when the triggers hit or don't, one worldview has earned the right to be believed.
He also draws a distinction almost nobody else does: pause buttons are not interchangeable. A mechanism that lets 2–5 actors trigger a global compute winter — everyone stops — is a categorically different object from one that lets 1–5 actors selectively disenfranchise the people they dislike while exempting themselves. The first is a circuit breaker. The second is a throne. Plan A picked the first, and he credits it for that.
And then he undercuts himself on purpose: maybe this is also naive. Actually, probably it's naive. The essay's whole force comes from a man refusing to pretend he has resolved an uncertainty he has not resolved.
Why this is an investment question, not a philosophy seminar
Because the argument identifies something rare: a set of assets whose thesis does not depend on who wins.
Vitalik's d/acc list is not a mood. It is a shopping list — formal verification, cryptography, secure and open hardware, defensive biotech, food and basic resource security, public epistemics, non-power-concentrating physical security. His claim is that these are worth building in both worlds. In the normal-technology world they are ordinary infrastructure for a society running on software it cannot audit. In the ASI world they are the only things standing between us and a single owner of everything.
Notice that most of that list is what crypto has spent a decade building badly and is finally building properly: verifiable computation, open hardware, cryptographic attestation, decentralized training, credible neutrality as a product feature. And the capital has started to notice — Vincent Weisser raised $130M at a $1B valuation to build an open superintelligence stack, on the explicit thesis that pre-training concentrated frontier AI in a few labs and reinforcement learning breaks that concentration open again.
The two camps disagree about the timeline. They agree, without quite noticing, about the failure mode: power concentration. Yann LeCun says the biggest risk is a few firms owning the assistants. Kokotajlo's original forecast said the risk is irreversible concentration. Vitalik says ASI itself is a massive power concentrator. The doomers and the skeptics are pointing at the same thing from opposite ends.
The TT desk thoughts
Do not trade the timeline. That is Vitalik's entire point and it is correct: nobody — not Kokotajlo, not LeCun, not the people mocking the scaling graph — has an edge on when or whether ASI arrives, and any position sized on a 2030 date is a position sized on a coin flip someone told you was a forecast.
Trade the invariant instead. The one claim both camps concede is that concentrated control of AI is the thing that goes wrong — they only disagree about whether the concentrator is a machine or a boardroom. That makes anti-concentration infrastructure the rare trade that pays in both worlds, and it is precisely the d/acc list: verifiable compute, open model weights and the rails to serve them, formal verification, cryptographic attestation of what model ran on what data, decentralized training. In the normal world these get bought because auditability becomes a compliance requirement. In the ASI world they get bought because they are the last line. You do not need to know which world you are in to own them — which is the only kind of conviction available here.
What to watch, concretely: whether open-weight frontier models keep pace with closed ones (LeCun's thesis lives or dies here); whether verifiable-inference and attestation projects get real usage rather than testnet applause; and whether anyone actually pre-commits to Vitalik's triggers. The last one is the tell. If serious labs and governments start naming the conditions under which they would slow down, the coordination world is real and the infrastructure gets bid. If nobody will name a trigger, the honest read is that the race is the only plan there is — and then you want to own the things that survive a race, not the things that assume a treaty.
Keep reading
Why AI agents need crypto rails · Emerging tech — the full framework · All notes