Superintelligence Strategy explained: the Hendrycks, Schmidt and Wang paper and MAIM

What is the Superintelligence Strategy paper? It is a 2025 proposal by Dan Hendrycks, Eric Schmidt and Alexandr Wang for how nations should handle the race to superintelligence, which they define as "AI vastly better than humans at nearly all cognitive tasks." It was submitted to arXiv on March 7, 2025 and revised on April 14, 2025. Its best-known idea is Mutual Assured AI Malfunction, or MAIM.

Why explain it now? The phrase "superintelligence" has moved from research essays into US government titles, including the September 29, 2026 executive order we covered in our explainer on the Super Intelligence order. Ideas written for that moment still shape the arguments, and this paper is cited more than most.

Who wrote it

The authors are Dan Hendrycks, director of the Center for AI Safety; Eric Schmidt, former Google chief executive and chairman; and Alexandr Wang, founder and chief executive of Scale AI, according to the project's website. The mix matters. One author comes from safety research, one from big tech and one from AI infrastructure.

The three pillars

The paper organizes strategy under three headings.

Deterrence. The authors propose MAIM, described in the paper as "a deterrence regime resembling nuclear mutual assured destruction (MAD) where any state's aggressive bid for unilateral AI dominance is met with preventive sabotage by rivals." The tools range from cyberattacks to, in the extreme, strikes on data centers. The argument is that a state racing alone toward dominance would trigger a response before it got there.

Nonproliferation. This means tracking where advanced AI chips are and keeping them from rogue actors who could use AI as a weapon. The model is how nuclear materials are controlled.

Competitiveness. This means making sure a nation has domestic access to chips through manufacturing capacity, and strengthening its economy and military using AI responsibly.

The authors contend that sabotaging destabilizing AI projects is relatively straightforward, which they say makes MAIM already the operative strategic reality for the major AI powers. That is a strong claim, and it is the one critics attack first.

What MAIM assumes

MAIM rests on four assumptions you can check one at a time.

1. Red lines can be seen. A state must be able to tell when a rival crosses into dangerous territory, such as automated AI research. 2. Sabotage works. Preventive attacks must set a rival back enough to matter. 3. Threats are credible. Each side must believe the other will act. 4. Stakes are shared. Both sides must prefer mutual restraint to a bet on winning.

The critique

On April 11, 2025, David Abecassis of the Machine Intelligence Research Institute's technical governance team published a response, "Refining MAIM." It says the framework "struggles with unclear and unmonitorable red lines for AI development, questionable threat credibility (particularly the capability of sabotage to deny progress), and a volatile deterrence calculus driven by immense stakes and uncertainty."

In detail, the critique argues that red lines around automated AI research are hard to monitor because progress is rapid and incremental, that sabotage gives delay and not denial, and that the perceived stakes create overwhelming motivation to pursue superintelligence regardless. Its proposed direction is to deter earlier in the chain, at advanced chip fabrication or compute concentration, instead of trying to monitor ongoing AI development.

That last suggestion is close to the paper's own nonproliferation pillar, so the disagreement is partly about emphasis. MIRI wants the weight on the chokepoints, while the authors put MAIM at the center.

Where it shows up now

Three live debates borrow from this framework. First, the argument over automating AI research, covered in our piece on the intelligence explosion paper, echoes the paper's red-line idea. Second, export controls on chips, the nonproliferation pillar in practice, remain central to US policy. Third, legislative proposals to prohibit superintelligence development, such as the Sanders and Casar bill, rest on the assumption that restraint is possible, which is the question MAIM tries to answer with deterrence.

We did not interview the authors or critics, and the paper's strategic claims are predictions, not findings.

Why a 2025 paper still matters

The paper is more than a year old, and the policy language has moved. The White House published an executive order titled "Inaugurating The Era Of Super Intelligence" on September 29, 2026, according to its presidential actions page. We have not verified whether that order draws on this paper, and nothing we opened says it does. What the paper supplies is vocabulary: deterrence, nonproliferation and competitiveness are now standard headings in the debate, whoever uses them. That is why it keeps being cited even by people who reject MAIM itself.

How to read it

Treat the paper as a structured argument and test it with questions. Who monitors the red lines? What happens when a state says it has not crossed one? Is sabotage a credible response when the target is a software capability and not a missile?

Our position: the paper is valuable because it names the problem as a strategic one, and its three pillars are a usable checklist. We would not accept the claim that MAIM is already the operating reality. Nothing we opened shows an actual sabotage deterrence episode, and the critique that red lines cannot be monitored is, so far, unanswered by evidence. Watch for governments that adopt compute reporting, since that is the piece every side supports.