Dear Aventine Readers, 

This week we dive into Substack, where a team of AI safety researchers published a road map called Plan A that they believe will lead to optimal AI outcomes. The multistep plan, once you learn the details, does not scream realistic. (The last step is essentially to give AI the nuclear codes.) But it forces a useful thought exercise about how we, collectively, get from where we are now to a world in which AI is a powerfully useful but also benevolent, incorruptible force. There are plenty of critics, and we include their voices too. But reading the plan and dismissing it forces the question: What would be better? So read it in that spirit and let us know your thoughts. 

Further thoughts from Substack: 

Thanks as usual for reading, 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

You might remember AI 2027, a project published in April 2025 that imagined the near-term impact of rapidly escalating AI capabilities. It was put together by a team of well-known AI safety researchers working together as the AI Futures Project, and it has held up pretty well. It predicted the dramatic rise of AI coding agents in early 2026 and described a national security panic that — if you squint — resembles this summer’s flap around Anthropic's Mythos. The same-ish team is back with a new analysis called AI 2040: Plan A, announced on the group’s Substack earlier this month. 

While AI 2027 was a series of predictions, AI 2040 is a highly ambitious road map for what the US should do to ensure optimal outcomes from AI and avoid existential catastrophe. To convey a sense of what level of catastrophe the team imagines, it is telling that the first step they prescribe is to slow down AI development as soon as possible. Scott Alexander, a highly influential thinker and writer about AI, philosophy and futurism was one of the AI 2027 co-authors and is now serving in an advisory role for AI 2040. He describes the current state of affairs on his own Substack as “on track for a poorly-controlled intelligence explosion that either ends the world or dooms it to permanent techno-oligarchy.” But thanks to some extra foresight and determination, he writes, there is a future in which the US “makes only good choices” if it follows the report’s year-by-year plan of action. 

Very briefly, the future plotted out in Plan A goes roughly as follows. In the very near term, multiple "Mythos moments" convince Washington that the AI race against China is progressing too quickly and that the rival nation could soon gain “a dangerous strategic advantage” over the US. The US proposes a joint regulatory regime to China, which China accepts partly because such a proposal implicitly treats the two countries as equals in the AI race, thus flattering China, which is actually still a little behind. The deal is designed not to depend on mutual trust: There is joint control of chip supply, shared knowledge of chip locations, mutually audited data centers in neutral territories and so on. Through top-down government regulation, AI capabilities are then deliberately slowed in each country. After that, AI is scaled up to be only as intelligent as top human geniuses by the mid-2030s. At that point, further advancement is paused while the genius-level AIs solve safety (or alignment) problems, ensuring that future forms of AI are built to be incapable of functioning in ways that could harm humans. As all this happens, Plan A anticipates significant economic upsides from the genius-level AIs, which fund a healthy universal basic income for all citizens. Only then, once safety concerns are put to rest, is AI allowed to accelerate once more. At this point, given the fallibility of humans, Plan A believes that AI should have final say over issues of grave international consequence. “The idea is that the US president or the paramount leader of China still controls day-to-day decisions,” Alexander explains in his Substack, “but if someone tries to pull off a coup and launch the nukes for no reason, then an aligned AI controls the nuclear missiles and it says no." 

If you think Plan A seems a bit pie-in-the-sky, that is in some ways the point: to be the opening salvo in a conversation that implicitly forces the question, “What’s your plan, then?” And its authors — well known in AI communities — command enough attention to get such a conversation started. Already, there is plenty of criticism online. Richard Ngo, who provided advice to the AI Futures Project but disagrees with parts of the published version of AI 2040, wrote about some of the main problems with the project on his Substack, Mind The Future. One of his big objections is the default assumption that fast takeoff — by which an artificial intelligence model rapidly transitions from roughly human-level intelligence to vastly superhuman intelligence in a matter of weeks, days or even hours — will have a dramatic impact on society. He argues that there is a huge gulf between the benchmark performance of AI and its diffusion into real-world settings (with limited real benefit so far), making the idea of requiring a slowdown questionable.

Tom Davidson, meanwhile, argues that the scenario's own enforcement mechanism doesn't hold up: Pausing software progress while compute keeps scaling, an assumption of Plan A, means that if the deal ever breaks down, the resulting sprint to superintelligence would be far faster and more dangerous than if no one had paused at all. And criticism rounded up by Zvi Mowshowitz on his Substack, Don't Worry About the Vase, includes the argument that the proposals in Plan A violate the spirit of the First and Fourth Amendments and constitute a recipe for authoritarianism. Meanwhile, Nicholas Decker, on Homo Economicus, contends that pausing AI research to solve the alignment problem is a bit like giving up flying planes to make sure they never crash: “I am doubtful that we can align via theorizing about a technology which does not yet exist,” he wrote.

Obviously there’s never going to be uniform agreement over something like this. But the response has for the most part ignored the mere fact of such a document — an arguably actionable plan put forward by respected voices in AI safety on how humans and AI can safely co-exist. Like it or not, Plan A begs the questions: Should we have such a plan? And if so, what should it be? 

A Framework for Frontier AI and the Dawning of a New Age, from Demis Hassabis

AI governance is also on the mind of Demis Hassabis, the CEO of Google DeepMind, who has taken to Substack for the first time to lay out his vision for how AI should be regulated during what he describes as “the dawning of a new age for humanity." The US, he believes, should take on a watchdog role with global reach. In this capacity, it would screen advanced models and potentially coordinate an industry slowdown if risk starts to build. The US, he argues, could "establish a new Standards Body modeled on a federally overseen public-private partnership or self-regulatory organization, much like the Financial Industry Regulatory Authority (FINRA)." Initially, he suggests, frontier models would be shared with this new regulatory body on a voluntary basis, but before long that would change and all models would have to pass a review to be deployed in the US. Other AI CEOs, including Sam Altman and Elon Musk, praised the idea. Elsewhere on Substack, though, the proposal has been criticized as light on detail and underpowered as a solution.

Making Our Own Luck, from Macroscience

There are two main approaches to funding science on a national scale: Target the most pressing problems and fund them aggressively, accepting that a top-down bet is high risk; or spread resources among more projects, which hedges any single bet. In reality, agencies like the National Science Foundation and the National Institutes of Health blend the two, but the mix still requires tough, subjective choices. In this post, Kris Willis, who researches how scientific progress occurs, explores whether it's possible to take a more data-led approach to funding. Using biomedicine as a test case, she and her collaborators grouped 18 million papers published between 1994 and 2017 into topics and found distinct publication patterns leading to 21 known breakthroughs. Such breakthroughs were preceded by a burst of newly published papers that built on each other, pulled in work from other fields and collectively contributed to a significant advancement in our understanding of the subject. Willis shows that this pattern emerges an average of five years before significant scientific discoveries and up to 30 years before such discoveries win a major prize like the Nobel. She argues that early funding for research exhibiting such patterns could be a more effective way to deploy funds than current methods and proposes an experiment to prove it. She suggests identifying signals from recent years, funding half of the identified areas at random, and comparing results between the two groups at five and ten years — a method she estimates could work as well in physics or materials science as in biology. This, she argues, may be one of the best shots we have at "making our own luck."

Medicine is Moving From Calendars to Clocks, from Ground Truths

In this post, physician and researcher Eric Topol explains a new Nature Medicine paper he co-authored with neurology professor Tony Wyss-Coray on "biological aging clocks." They argue that medicine is shifting from measuring age by the calendar to measuring it by biological markers at the organ or even cellular level. Research has already shown that we don't age gradually but experience aging in waves. We're now starting to understand why, and what that means. To start with, a person’s organs don't all age at the same rate — at any given time, one organ might function as older or younger than another — and the pace at which organs age varies from person to person. Fast-aging organs correlate with disease in that organ, and organ age reflects lifestyle factors like drinking, smoking, education level and physical activity. Newer research suggests that we can track aging at the cellular level as well, based on an analysis of blood proteins. Accelerated aging of a specific kind of brain cell known as an astrocyte, for instance, carries a 12.6-fold higher Alzheimer's risk compared with people whose astrocytes were aging more slowly than the norm. And two systems — the brain and immune system — appear to act as master regulators; when they age at a normal pace, people tend to live longer even when other organs age quickly. These techniques for measuring aging aren't in routine clinical practice yet, but Topol expects that to change soon. In the near future, he believes, physicians will look at a person’s rate of aging to gauge overall health and researchers will develop interventions to try to reverse it. (For more from Dr. Topol, he speaks with Aventine about his latest book in this episode of our podcast; transcript here.)  

TokenBudgeting: Our Conversations with Enterprises on Token Spend, from SemiAnalysis

The era of "tokenmaxxing," when employees at companies like Meta and Salesforce were encouraged to use AI models as intensely as possible, has already given way to more careful budgeting of AI use. This post by SemiAnalysis, a tech research company, provides some fascinating insights into what these changing norms around AI spending look like inside US companies, based on industry analysis and more than 50 conversations with people inside large businesses. Newly established budgets for individual employees vary wildly with no consensus figure: from $250 a month per employee at an aerospace firm, to $2,000 at Workday and Stripe, up to tens of thousands for senior staff at some firms. In some companies, there are guidelines rather than hard limits. Meanwhile, one US airline ties token allocation to project-level revenue forecasts, treating AI like travel expenses. And it’s worth noting that the headline figures are by no means the norm: While the top-spending 1 percent of companies might be paying as much as $90,000 per employee per year, the median customer according to the SemiAnalysis data spends just $136 per employee on AI a year. For a typical Fortune 500 company it's under $100. 

The AI Superforecasters Are Here, from Astral Codex Ten

One of our January predictions for 2026 was that AI would help us make better high-stakes decisions. Scott Alexander, mentioned in the story about Plan A above and who writes Astral Codex Ten, explores how that's playing out and finds that AI is increasingly reaching parity with human superforecasters — people who are able to make unusually accurate predictions about complex real-life events. The Metaculus Cup, held three times a year, is a forecasting tournament in which competitors answer questions like, "Who will win such-and-such election?" or "Will such-and-such country attack such-and-such country?" In the most recently completed Cup, which ended in May, humans took the top two spots, but AIs came in third, eighth and tenth out of over 1,200 contestants. The question now, it seems, is how quickly AI models might surpass humans at making forecasts, and here opinions are split: Some think models will plateau around human level, others think they will keep improving and consistently surpass human forecasters. (For what it's worth, human Metaculus forecasters predict a 12 percent chance that a bot will win this summer's Metaculus Cup and a 98 percent chance one will win sometime before 2030.) But Alexander argues it might not matter if AI doesn’t keep getting better: Even AIs that are no better than top humans could change things dramatically, because highly accurate human forecasters are scarce, slow and expensive. Having computerized equivalents on tap could help us make better informed decisions about practically everything.

Why American ambulance rides are so expensive, from David Oks

An ambulance ride in the US is often billed at more than $10,000 for a journey measured in minutes. David Oks explains that the reason for the exorbitant charge isn't the greed of ambulance operators, many of which survive on razor-thin margins. The problem is a structural mismatch between the cost of keeping a fleet of ambulances ready around the clock and the way the US healthcare system pays for such a service. Most wealthy nations fund ambulance service as an ongoing need — like road repairs and infrastructure — through upfront payments from every taxpayer or household. This ensures that everyone can get a near-free ambulance ride when they need one. In the US, ambulance service is paid for on a per-ride basis due to the way it was classified in the bill that introduced Medicare in 1965. The issue is that most rides covered by government-provided health insurance lose ambulance companies money: Medicare pays about $329 for a ride that costs around $2,673; Medicaid pays even less. Uninsured citizens rarely pay the difference, so the privately insured carry the entire system. And because of the nature of the business, insurers can't steer patients to specific — more economical — ambulances, so they have no leverage to bring costs down. That means roughly 80 percent of rides are billed out-of-network, producing large bills. Oks argues that without a more evenly distributed approach to paying, that's unlikely to change.

What Would You Do if Labor Was Free?, from It Can Think!

Here's a nice little post imagining the many good things robots could deliver to society. While much of the world frets about robots usurping jobs, leaving humans without purpose or means of support, roboticist Chris Paxton imagines a robot-driven future in which: "Litter would vanish. Public spaces could be maintained, by robots, at the level of a boutique hotel." He imagines a world where "homes can be de-cluttered and automatically organized, their contents entered into a contextual database, instantly queryable and searchable." And he suggests that "with ‘free’ robotic labor available at scale, we might repair old shoes or upgrade old phones instead of throwing them out, reupholster old furniture, and patch worn clothes instead of consigning them to a landfill." It’s definitely a more positive spin than most of us are used to hearing, but that scenario — like the robot job apocalypse — is likely a very long way off.