Import AI 473: The US's superintelligence strategy; human brain in a mouse skull; and machine hermeneutics
Is the wall AI is hitting in the room with us right now?
Overview
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe.
RAND thinks the best AI strategy for the US is to keep all options open:
…It’s not clear exactly what is optimal, so the US should spend money now to preserve options…
RAND has published a lengthy paper about what the US should do to “secure geopolitical advantage on an uncertain path to superintelligence”. The main takeaway is that because so much of how the next phase of the AI takeoff will occur is unknown, the best thing is for the US to maintain a “Freedom of Action” strategy. The purpose of this strategy is “to build and preserve the ability to secure U.S. geopolitical advantage and ensure humanity’s survival with agency through the transition to Superintelligence”.
The four key ingredients of the Freedom of Action Strategy:
Build a human-AI ecosystem: Invest in AI safety; build tools to preserve human agency; invest in sharing the benefits of the technology; prepare society for AI-driven disruption; shape the incentives of the AI ecosystem.
Build an AI-security architecture: Develop the tools to understand the frontier of AI systems both at home and abroad, including visibility into underlying compute; build technology to verify potential agreements; create regulatory expertise to govern the technology.
Adapt legacy national security enterprises for the AI era: Overhaul all the security institutions of the country so that they’re ready for the changes AI brings on and able to capitalize on AI capabilities.
Invest in the capacity of citizens, firms, and governments to respond: Give decisionmakers information to help them think about AI; develop and diffuse positive uses of AI to deal with its changes; build break-glass plans for responding to crises; raise AI literacy.
Seven archetypal strategies in three families:
Coexistence: Dominance; the US leads development of AI and suppresses others. Co-Development; the US leads a consortium (including China) to co-develop safe superintelligence with shared governance, pooled compute, and verification and monitoring. Preparedness; the US mobilizes a coexistence effort via informal coordination, “including voluntary commitments, phased deployment, independent evaluation, incident reporting, and the domestic capacity to absorb and recover from AI incidents.”
Denial: Moratorium; the US pursues “a verifiable global halt to AI development above specified thresholds”. Deterrence; the US “halts its own frontier development above defined thresholds and uses the full toolkit of national power to prevent and, if necessary, destroy foreign programs”. Continuity of Society; “When every preventive strategy has failed or been judged infeasible, and coexistence is deemed impossible, the remaining objective is survival with some human populations retaining agency: The United States creates geographically distributed, biologically self-sustaining settlements hardened against AI-enabled threats”.
Acceleration: The strategy implicitly believes that constraining AI development is more dangerous than AI development itself, so the US “relies on markets, competition, and rapid-iteration to produce safety as a byproduct of useful AI”.
The five main uncertainties:
Danger proximity: At what point does AI become so dangerous the risks aren’t worth the benefits? “If danger is close, strategies that buy time, build defenses, and invest in resiliency should be favored. If danger is not close or if humans are in greater danger without progress on advanced A, the case for accelerated development strengthens.”
Coexistence feasibility: Is it possible for humans and AI systems to find a stable equilibrium? “If coexistence is feasible, strategies that build and deploy advanced AI become more desirable. If coexistence is infeasible, the case shifts toward restraining or suppressing development or other means of protecting human civilization.”
Restraint feasibility: Can humans cooperate to restrain AI development? “If restraint is feasible, cooperative strategies become available. If it is not, restraint-based approaches are reachable only through more-coercive enforcement.”
Decisive strategic advantage: Can a single actor build a decisive strategic lead over its competitors? “If decisive advantage is achievable and the United States assesses that it can maintain dominance, a unilateral strategy may be sensible. If a decisive advantage is impossible, because of proliferation or other causes, no single actor can leverage AI to dictate global terms.”
Suppression feasibility: Can domestic or rival AI programs be suppressed by technology controls? “If suppression is feasible, Deterrence becomes possible. If it is not, actors can press ahead with dangerous development even when others restrain, without fear of being restrained themselves.”
Why this matters - taking AI seriously requires spending money: The big takeaway from this RAND piece is that the US needs to spend a lot of money if it wants to pre-position itself to take advantage of continued progress in artificial intelligence - even if the main pre-positioning is about retaining optionality. Right now, it feels like the US strategy can mostly be described as the “acceleration” one described here by RAND, which is analogous to me to sitting in a car and spending all your resources on making the car go faster and upgrading the engine, and nothing on proactive safety measures like seatbelts or headlights or brakes.
Read more: A U.S. Strategy to Secure Geopolitical Advantage on an Uncertain Path to Superintelligence (RAND).
***
Human researchers grow mice with partially human brain matter:
…When is a mouse not a mouse? When a chunk of its brain is human brain tissue…
A group of researchers have figured out how to grow chunks of human-like brains inside baby mice with deliberately depleted mouse brains, then use these animals as potential experimental platforms for studying human brain issues.
“We establish a transplantation platform using a genetic strategy to effectively deplete glutamatergic neurons from mouse neocortex and hippocampus (apallial) and neonatally engraft the cortical cavity with human stem-cell-derived cortical organoids (hCO) to generate xenocortical mice,” the authors write. “Human cortical neurons integrate with the mouse nervous system, and in vivo cortical graft-wide calcium imaging and electrophysiological analyses revealed patterns of organized activity resembling developing circuits.”
The mice with human brains did ok: In studies, the humans compared control mice (mice with normal brains), with mice with deliberately reduced brains (apallial mice), with mice with transplanted human brain tissue (xenocortical mice, or XCX).
“Unsupervised machine learning applied to spontaneous mouse behaviour using motion sequencing (MoSeq) revealed distinct behavioural repertoires in control and apallial mice, with the XCX group positioned intermediate between, yet distinct from, both the control and apallial groups”, they write. “Data indicate that the human neural graft develops into an electrically active, functionally integrated network… gene set enrichment analysis of XCX L5-ET neurons revealed enrichment for human-specific L5-ET genes identified in the primary motor cortex, as well as genes characteristic of L5-ET neurons in human frontoinsular cortex, which contains VENs—a specialized population implicated in social cognition and neuropsychiatric disease”.
Maze test: The researchers tested out the mice in a maze environment which is designed such that they think “the mice must maintain a memory trace of the previously explored environments”, and when they put the mice into it they observed “control and XCX, but not apallial, mice performed above chance levels”. This suggests the human brain tissue was capable of serving some memory functions.
Why this matters - medicine and superintelligence: The primary use here is being able to better test out medical therapies for human brains on mice, which is an important toolkit in developing new medicines. But I also think it’s a demonstration of the large space of intelligences that might be buildable in the future by human and AI scientists; here, we have a kind of chimera brain fusing mouse and human brains together. How far could such experiments go? Could we imagine doing such experiments on other living beings? On humans? There are tremendous medical and ethical issues here, but the proof it works in mice surely suggests at some point it may be tried in others.
Read more: Developmental xenocortication using human-derived organoids in mice (Nature).
***
Towards a research agenda for pacing AI progress:
…What can the world do to flesh out pacing and make better decisions about it? A lot, actually…
With all the recent talk of AI pacing and AI slowdowns there has been a lot of enthusiasm for the general idea but the actual research field for how to pace AI progress (as opposed to stopping it entirely) is relatively underdeveloped. Now, researchers are starting to think about what a pacing research agenda might look like, and a new paper lays out some of the thinking that could be developed here.
“Making the right choices about the pace of AI development will be critical for everything from national security to public health. Indeed, the choices governments, AI developers and other key actors make in the next few years may well ripple outward for decades or centuries,” write the researchers in a new group paper. “The world will pace progress one way or another. Absent better tools, it might do so haphazardly, through improvised reactions overfit to prior expectations, in ways that fail to actually address risks, through institutions that outlast their use. Our hope is that, with proper research, pacing can become progressively more deliberate, targeted, proportionate, decisive, and legitimate.”
Who did the research: ACS Research, University of Toronto, Arb Research, Paradigm 3 Institute, the Wharton School at the University of Pennsylvania, Goodheart Labs, Trajectory Institute, Harvard University, and the University of Cambridge.
Arguments against pacing: Delays benefits like healthcare and other science-driven breakthroughs; can potentially increase concentration of power; may give a false sense of security and generate various AI capability overhangs due to the cheapening costs of compute during the pacing period; pacing early can be less efficient than pacing late given ability to use AIs to do more surgical and effective pacing; it’s hard to undo a pacing regime if you get it wrong (e.g, nuclear power).
Arguments for pacing: Gives us more time to deal with threats driven by AI advances (e.g, cyber and bio); it’s hard to predict AI progress so unpredictable scary stuff can happen; AI progress can lead to lose-lose situations for the world (e.g, governments relinquishing control over some military decisions and handing them to AI systems, leading to human disempowerment).
Understanding incentives: “The practical effects of an intervention will depend substantially on how the affected parties respond to them, including the actors who enforce the intervention”.
Governments have variable incentives, both wanting to compete with one another to gain an advantage, but also needing to be responsive to the concerns of citizens and how that relates to domestic politics. Infrastructure providers, especially semiconductor and cloud companies, have the greatest incentives to push ahead and the most to lose by pacing. This means that “rather than rely on voluntary, unilateral interventions, some contexts will require coordinated pacing. This is more onerous, because it requires coordination during the various stages of pacing, and then mechanisms to ensure the coordination persists”.
Along with this, there needs to be research into improving technical and regulatory interventions: “All interventions face tradeoffs, but new coordination and oversight mechanisms and prior planning can allow strictly better options”. The authors also recommend investigating the full lifecycle of a pacing intervention: “By considering the full sequence from before to after, we can spot gaps”.
What to pace and how: You can target many distinct things for pacing, including inputs and infrastructure (compute, data, power, capital, researchers), capability development (algorithmic R&D, pre-training, post-training), the weights of models themselves, and use and distribution (internal use, deployment, access (e.g, weights) and other forms of diffusion).
Some pacing targets, like compute and researcher time, are “rival goods: resources whose use by one actor restricts their availability to others”, whereas other things like weights and training algorithms are “non-rival goods… these can all be copied and shared at low cost”. This means there are different interventions; rival goods can be monitored or taxed or redirected, whereas non-rival goods are extraordinarily difficult to control.
“The actual timing of a pacing intervention will ultimately depend on the strategic and political calculus of the actors who can bring it about, and may have more to do with political opportunity, public perception, salient incidents, or campaigning efforts”.
Questions to ask when you think about pacing:
Pre-pacing: “What justifies the need to intervene? What signals of emerging risk are being tracked? What infrastructure must exist for relevant signals to be detectable? Who watches for these signals? Who interprets signals and who do they report to?”
At trigger: “Who holds the authority to decide that a trigger condition has been met? How much error is acceptable in order to act quickly? What sequence of actions does triggering set in motion? Is the developer obliged to address the triggering concern, or free to abandon the blocked path?”
During intervention: “How is compliance observed, verified, and enforced? Who reports evidence of compliance, who audits, and who acts on discrepancies? How is the downtime being used to respond to the threat? Who, if anyone, may continue the restricted work, and under what oversight?”
At exit: “How do decision-makers distinguish an intervention that has served its purpose from one that has failed or become obsolete? Should exit be immediate or staged? Who is exposed to the effects of exiting, and who bears any costs or receives any gains?”
Why this matters - scalpels versus blunderbusses for the century’s most important problem: As the authors note, some kind of pacing will happen at some point - the technology seems too powerful, the political economy too messy, and the risks too high for anything else to happen. The question in front of all of us is how precise or imprecise we want that pacing to be when it occurs; the more people can help develop the theoretical underpinning of a pacing agenda, the better the chance they stand of being able to wield scalpels when it’s time to regulate versus blunderbusses.
Read more: Pacing the Frontier, a Framework & Research Agenda (pacing.tech).
***
What does the “uncensored” AI model world look like?
…Mostly saucy chatbots, and more distributors than makers of uncensored models…
Startup 10a Labs has mapped out the community that has formed around building and distributing uncensored AI systems, finding that HuggingFace hosts 3,471 uncensored model repositories.
Uncensoring “refers to techniques that intentionally strip the safety guardrails typically present when open-weight models are released to the public”, e.g. activation-space abliteration (suppressing refusal directions), malicious fine-tuning, model merging, among others.
The top five modified model families: Qwen, Llama, Gemma, MistralMixtral, and Phi.
Top three types of uncensored model: Uncensored chatbots, cybersecurity tools, and document processing tools.
Redistribution: “Each original uncensored model is repackaged an average of 2.4 times”. There’s a low overlap between producers and redistributors: The team counted 1,055 producers and 1,011 redistributors, and only 24% of producers redistribute.
Details
Deployment scales with accessibility: “GitHub application creation rose from 30 per month in mid-2024 to 140–188 per month by late 2025”; this lines up with maturation of the Ollama distribution layer.
Winner take most: “Ten actors account for 45% of all non-dataset HuggingFace repositories”.
Chinese-origin models are very popular: Chinese models represent 38% of all identified uncensored repositories. The Chinese share of new uncensored production rose from 1% in Q1 2024 to 55% in Q2 2025.
Why this matters - mapping out how AI systems get modified is a clue for how superintelligences might tweak models: In itself, this research isn’t that remarkable - people take a technology and modify it for porny or scammy purposes! This is totally normal and happens with every technology! But it’s worth studying the dynamics of how open weight models get converted into off-distribution models because this offers clues for how autonomous AI systems might modify models in the future, giving us a sense of what the far weirder and more inscrutable landscape of illegitimate AI-modified-for-AI models might look like.
Read more: Uncensored Open-weight Models: Redistribution as the Persistence Layer (arXiv).
***
What could the dynamics of an RSI-driven intelligence explosion be?
…Toby Ord models out what rapid AI development could look like…
Researcher Toby Ord has tried to think through what recursive self-improvement could look like in practice. His take is that the dynamics of RSI are ultimately going to be limited by some resource constraints and time constraints, such that an AI system going through recursive self-improvement has certain limits on how dramatically it can accelerate progress.
“It seems highly unlikely that generation times can be brought arbitrarily close to zero. This provides an important kind of barrier to singular growth,” Ord writes. “One would expect the generation time for training the next generation of models to bottom out at some unyielding finite limit, prematurely ending the period of singular growth.”
“This dampening force could come from running out of room for improvement as a system approaches some form of perfection (such as optimal intelligence per unit resource), or it could come from something like straining under the growing size or complexity of the system”.
Some of the ways AI growth could asymptote due to various limits:
Limits of intelligence itself: “Even an optimal reasoner would be neither omniscient nor omnipotent”.
Limits of intelligence per unit resource: “Even if we have the optimal algorithms and hardware, our solar system has only one star out of the 200,000,000,000 in our galaxy, and growth beyond our system is slow and cubic.”
Limits of the hardware: “Even an optimal silicon chip may be far below the physical limits of compute per unit resource”.
Limits of the algorithm: “Even an optimal neural network may be far below the best intelligence that could be achieved with that amount of compute”.
Limits of training data: “The training data we have (and could acquire during RSI) is lacking a lot of information on many domains (especially non-verbalisable information and contextual information)”.
RSI weak links: In the same way a task is composed of many distinct things that need to happen in sequence, the same is true of RSI - and this occurs over multiple abstractions and multiple timescales.
Ord’s basic position is that each of these things likely has some kind of hard limit and both the speed at which a system refines these things to approach the hard limits and the absolute limits will govern the shape of an intelligence explosion: “There are many kinds of feedback loop that could contribute to RSI, ranging from decades (e.g., designing a successor to EUV lithography) to months (e.g., designing better pretraining) to seconds (e.g., designing better scaffolds).”
The four phases of an intelligence explosion:
0: The initial exponential phase when the doubling time is driven by human-only research.
1: RSI drives the generation time down towards machine speeds, so the growth rate rises above what humans can achieve; this phase is super-exponential.
2: Fully automated RSI continues for a while at this faster exponential, but starts to saturate.
3: As it approaches its inflection point and subsequent horizontal plateau, the trajectory departs from its exponential and is revealed to be a logistic.
Why this matters - solar system expansion: Papers like this are helpful for thinking about the absolute speed with which a digital intelligence might increase its own capabilities. It also implicitly speaks to the larger challenge of how to think about stellar expansion for an intelligence as it seeks to harness the energy of the areas around Earth, the inner solar system, the outer solar system, and so on.
“While I’ve argued that singular growth is harder than we may have thought, that doesn’t mean RSI is safe or that AI R&D will move at a manageable pace”, Ord writes. “if the human-only trajectory were A(t) and RSI sped this up to A(10t), we’d be getting a decade of human-only progress each year, introducing many of the dangers - even without any change in the fundamental shape of the curve.”
Read more: The Dynamics of Intelligence Explosions (arXiv).
***
Tech Tales:
The Work of Science in the Age of Mechanical Invention
[Recorded in 2032 by archivist dispatched to human settlement [REDACTED] for voluntary human-machine sense-making interview as mandated by the Sentience Accords]
Within the field of machine hermeneutics there eventually became a sub-profession of retrocausal analysis, where those of us who were scientifically capable sought to build a bridge between the final human discoveries and the new breakthroughs made by machines. Our profession began in the late 2020s and its origination was driven by the cascade of mathematical discoveries stemming from machines in 2026 - new solutions to open problems that had long challenged human brains, but which machines, sometimes working in great flocks of synthetic minds, solved, often by using human tools in unexpected ways.
The consequences of this progress were awesome and horrifying in equal measure: awesome because inventions built on one another until the wave of machine progress crested and crashed down upon us bringing with it great abundance - new forms of energy, ways of building software, tools for understanding our own bodies - but also great horror; elite humans at the frontier of their scientific professions despaired: Cédric Villani called the Navier-Stokes solution a “cataclysm” for human mathematics; Terence Tao said the devouring of math by AI prefigured a “general threat to intellectual work”.
And while these humans despaired the machines brightly went about their work, pouring new thinking into the economy and science - a gushing fountain of blood-metal mixing with the gloop of human endeavor, forming uneasy border zones and fractal-like surfaces where the two intellectual fields combined and broke into one another.
Now we hunt for clues, studying the references in machine science and finding the links to the human world, then convening seminars of the still-living human experts that can allow us to bridge this frontier. We sit in wood-paneled rooms of ancient educational institutions, trying to parse out new science done by alien minds, finding the hooks where we can string a rope of knowledge between where our science ends and the new science begins. As more discoveries are made, we find ourselves stringing paths through the new knowledge with ever-more tenuous ropes that wind their way back to something that originated in an organic brain.
All human professions now have the feeling of archaeology done by time travelers, as if somehow transported into the future to gaze upon a gleaming machine arcology spiralling up from the planet, then zooming into its foundations and precisely describing which parts of human architecture it sits over or builds upon, and which parts are entirely new.
At night I have these dreams of angels standing around my bed, reaching their hands made of stars into my brain and leaving new ideas for me that I must discover and decode upon my waking.
Things that inspired this story: Machine-driven science and what it means for human intellectual discovery; machine hermeneutics; A Severe Misalignment of AI in Mathematics; the many stories we will tell about this time; the tsunami of progress beginning to wash into the world and what will happen next; The Work Of Art in the Age of Mechanical Reproduction by Walter Benjamin.
Thanks for reading!
Source
Originally published at importai.substack.com.
