Import AI 473: The… | Jack Clark from Import AI Summary

1. Key Themes

The US lacks a coherent strategy for superintelligence and is defaulting to "acceleration" by omission

RAND's framework argues that the government should be spending heavily to preserve strategic optionality, not simply racing forward. Jack Clark's own read is blunt: the US isn't really executing a deliberate strategy — it's accelerating by default while neglecting safety infrastructure.

"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."

"Pacing" AI progress is an emerging, underdeveloped policy discipline — not the same as stopping AI

A new multi-institution research agenda treats the rate of AI progress as a design variable to be actively managed (sped up, slowed down, or redirected), rather than a binary stop/go decision. This is a distinct and more nuanced framing than typical "pause AI" debates.

"The world will pace progress one way or another. Absent better tools, it might do so haphazardly... Our hope is that, with proper research, pacing can become progressively more deliberate, targeted, proportionate, decisive, and legitimate."

Recursive self-improvement (RSI) is bounded by physical and resource constraints, not infinite

Toby Ord's modeling pushes back on unconstrained "intelligence explosion" narratives, arguing that generation times, hardware, algorithms, and training data all have hard limits that will force any explosive growth curve into a logistic (S-curve) shape rather than a true singularity.

"It seems highly unlikely that generation times can be brought arbitrarily close to zero. This provides an important kind of barrier to singular growth... 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."

Open-weight model "uncensoring" is a maturing shadow ecosystem, increasingly dominated by Chinese-origin base models

A mapping of the uncensored-model community shows rapid growth, concentration among a few distributors, and a dramatic geographic shift toward Chinese-origin models — a proxy for how ungoverned modification of frontier open models could scale.

"The Chinese share of new uncensored production rose from 1% in Q1 2024 to 55% in Q2 2025." "Ten actors account for 45% of all non-dataset HuggingFace repositories."

2. Contrarian Perspectives

The US's real-world AI posture is riskier than its stated ambition, because "acceleration" masks the absence of a strategy. Most public commentary treats US AI policy as a deliberate race-to-win strategy; Clark's contrarian framing is that it is actually an absence of investment in safety/optionality dressed up as competitive urgency.

"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."

Intelligence explosions will not be unbounded "singularities" — against the popular assumption of near-infinite runaway growth. Ord's contrarian, quantitatively-grounded take is that even full RSI automation will saturate and plateau due to physical limits, though he cautions this doesn't make the transition safe.

"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... 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."

Pacing AI development too early could be counterproductive — against the "slow down now" consensus among safety advocates. The research agenda highlights that early pacing may be less efficient than pacing later, when AI itself can be used to pace more surgically, and that premature restraint can create false security and capability overhangs.

"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)."

3. Companies Identified

RAND — Policy think tank. Mentioned as author of the superintelligence strategy framework used to structure the newsletter's lead story. "RAND has published a lengthy paper about what the US should do to 'secure geopolitical advantage on an uncertain path to superintelligence'."
10a Labs — Startup studying the uncensored open-weight model ecosystem. Mentioned as the source of the mapping data on uncensored AI 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."
HuggingFace — Model hosting platform. Cited as the primary hub where uncensored model repositories concentrate. "HuggingFace hosts 3,471 uncensored model repositories."
Ollama — Model distribution layer. Cited as an enabler of accelerating uncensored model deployment. "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.
Qwen, Llama, Gemma, Mistral/Mixtral, Phi — Open-weight model families. Identified as the most frequently "uncensored"/modified base models. "The top five modified model families: Qwen, Llama, Gemma, MistralMixtral, and Phi."

4. People Identified

Toby Ord — Researcher/philosopher. Mentioned for his formal modeling of recursive self-improvement dynamics and intelligence explosion limits. "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."
Cédric Villani — Mathematician. Cited (in the closing fiction piece) as reacting with alarm to AI's mathematical breakthroughs. "Cédric Villani called the Navier-Stokes solution a 'cataclysm' for human mathematics."
Terence Tao — Mathematician. Cited alongside Villani as warning about AI's encroachment on intellectual work. "Terence Tao said the devouring of math by AI prefigured a 'general threat to intellectual work'."

5. Operating Insights

Infrastructure providers (compute/cloud/semiconductors) are the actors with the strongest incentive to resist any pacing or slowdown regime, meaning coordination efforts aimed at governance must specifically account for their resistance rather than assuming voluntary compliance.

"Infrastructure providers, especially semiconductor and cloud companies, have the greatest incentives to push ahead and the most to lose by pacing... rather than rely on voluntary, unilateral interventions, some contexts will require coordinated pacing."

Distinguish "rival" vs. "non-rival" resources when designing any AI governance or competitive strategy — compute and talent can be metered or taxed, but weights and algorithms, once created, are nearly impossible to contain, which should shape where founders and policymakers focus control efforts.

"Some pacing targets, like compute and researcher time, are 'rival goods... whereas other things like weights and training algorithms are 'non-rival goods... these can all be copied and shared at low cost'."

Redistribution, not production, is the dominant value-capture layer in the uncensored-model ecosystem — a structural pattern (few producers, many redistributors) that operators building in adjacent "modification/distribution" layers of AI infrastructure should study.

"Each original uncensored model is repackaged an average of 2.4 times... only 24% of producers redistribute."

6. Overlooked Insights

The human-brain-tissue-in-mouse research (xenocortical mice) is being framed by Clark as an early proof point for a much broader design space of hybrid/chimeric intelligences — a signal for biotech and AI convergence plays that's easy to dismiss as pure medical research but has speculative implications for future substrate experimentation.

"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... surely suggests at some point it may be tried in others."

RAND's "Co-Development" archetype — a US-led consortium that explicitly includes China in shared governance of superintelligence — is a notable, easy-to-miss policy option buried among more attention-grabbing "Dominance" and "Deterrence" strategies, and represents a real alternative to the current adversarial framing of US-China AI competition.

"Co-Development; the US leads a consortium (including China) to co-develop safe superintelligence with shared governance, pooled compute, and verification and monitoring."