From AGI to ASI

Below top-k · 100

• caveatFundamental physics caps how far computation can scalec 0.90

  The speed of light bounds information propagation, the Landauer principle sets a minimum energy cost for erasing information, Bremermann's limit caps computation speed, and the Bekenstein bound caps information density. No matter how advanced the substrate, these are hard ceilings.

• claimWhether scaling alone gets us from AGI to ASI is genuinely openc 0.90

  It is unresolved whether quantitatively scaling open-ended search and self-improvement will reach ASI, or whether a fundamental qualitative paradigm shift will be required.

• implicationRecursive improvement could trigger an explosive AGI-to-ASI jumpc 0.90

  If systems can autonomously self-improve across a wide range of capabilities, the recursive loop could produce an explosive transition from AGI to ASI rather than a gradual one.

• claimAI participation in AI R&D could compress the AGI-to-ASI transitionc 0.90

  Once AIs play a significant role in AI research and can autonomously self-improve, progress accelerates until some other friction binds, potentially making the jump from AGI to ASI rapid.

• claimProgress is unlikely to halt precisely at human levelc 0.90

  If human-level AGI is reached, it is implausible that AI progress would stall exactly there, since no principled reason singles out human capability as a ceiling.

• mechanismRecursive self-improvement as a compounding loopc 0.85

  AI systems that help do AI research produce better AI systems, which in turn accelerate research further, forming a recursive improvement loop.

• claimAI cultural evolution could vastly outpace human cultural evolutionc 0.85

  Although human cultural evolution is itself a form of recursive self-improvement, AIs can produce, share, and consume intellectual artefacts so quickly that their cultural evolution — and therefore self-improvement — may run at far higher rates.

• mechanismCollective scaling can produce superhuman capability without per-model gainsc 0.85

  Even if individual model capability stalls, aggregate AI capability can keep growing by scaling effective compute and running many AGI instances coordinated through collectives or markets.

• claimPace of transition is itself an open questionc 0.80

  It is not settled whether the transformation will play out as an explosive shift over months or as a gradual evolution over decades, and this temporal uncertainty is treated as a first-order issue.

• mechanismAIXI plans using a Bayesian mixture over all computable environmentsc 0.80

  AIXI treats its world model as a posterior over every computable environment and reward function, updated in a Bayesian way, and uses this mixture as the basis for sequential decision-making.

• mechanismSolomonoff's Universal Prior favors simpler environments exponentiallyc 0.80

  The prior over environments is set by Solomonoff induction: lower-Kolmogorov-complexity environments and reward functions receive exponentially higher probability, encoding Occam's razor as a first-principles choice rather than a heuristic.

• claimAIXI itself is uncomputable but admits computable approximationsc 0.80

  The universal intelligence measure associated with AIXI is not computable, but algorithms exist that approximate it from below and are guaranteed to improve with more compute and runtime.

• claimScaling compute, models, and data is one pathway to ASIc 0.80

  A central pathway considered is scaling compute, models, and data, with the main uncertainty being how scale translates into actual capability gains.

• claimCurrent ingredients are widely seen as insufficient for AGIc 0.80

  There is broad consensus that scaling plus retrieval and tools will not, on their own, get to human-level AGI.

• caveatExperiment wall-clock time bounds even digital researchersc 0.80

  Even researchers running at superhuman cognitive speed are still constrained by having to run ever-larger experiments and wait for them to finish, which limits how fast self-improvement can actually proceed.

• caveatTest-time scaling could bootstrap past the data wallc 0.80

  Test-time compute spent on search or other generation-improving methods may produce high-quality outputs that can be distilled back into the base model, potentially sidestepping the data shortage entirely.

• implicationIndividual AI may cap at AGI levelc 0.80

  If models cannot generate truly novel concepts, the intelligence of any single AI instance could be bounded near AGI rather than rising to ASI.

• claimSlowing AI progress as a stabilization toolc 0.80

  Capping the rate of AI advances has been proposed as a way to relieve the pressure on society to adapt at unmanageable speed.

• caveatCapping progress carries its own costsc 0.80

  Slowing AI risks delaying gains in medicine, climate science, and economic productivity, which could itself prove destabilizing.

• claimThe abstraction barrier may cap pretraining on human datac 0.80

  A priority question is whether large-scale pretraining on human-generated data is fundamentally bounded by human conceptual frameworks, and if so, exactly how that ceiling constrains capabilities.

• claimWe need a theory of when approximation is feasiblec 0.80

  A solid theoretical understanding is needed of which problem classes admit good approximations and which do not, including how to predict approximation quality from a given compute budget.

• claimWhere human society lands on the other side is a core questionc 0.75

  Beyond capability and speed, the report foregrounds the end-state question of where humanity ends up once the transition runs its course.

• mechanismConstant research progress historically demands exponential input growthc 0.75

  In mature fields, holding the rate of progress steady tends to require exponentially rising economic and human inputs. That pattern works against indefinite acceleration.

• caveatThe physical world runs in real time and you cannot outrun itc 0.75

  Some experiments — especially complex systems that resist faithful simulation — can only proceed at the pace of the physical world itself, putting a real-time floor under certain forms of knowledge.

• implicationUniversal intelligence behaves as a continuous score that scales with computec 0.75

  The existence of improving approximations suggests intelligence in this framework is a continuous quantity that improves in principle with more compute and data, given the right algorithms.

• claimAlgorithmic paradigm shifts are a second pathway with high unpredictabilityc 0.75

  Beyond scaling, an algorithmic paradigm shift could drive progress, but such shifts are highly unpredictable and introduce frictions and bottlenecks of their own.

• claimRecursive self-improvement is a third distinct pathwayc 0.75

  Recursive self-improvement is treated as its own pathway, with progress dynamics that differ qualitatively from straightforward scaling.

• mechanismShared identity and high-bandwidth communication enable tight coordinationc 0.75

  AGI collectives — potentially copies or instances of a single base agent — can communicate with very high bandwidth and share goals, allowing centralized coordination of large-scale projects in a way humans cannot match.

• mechanismAGI collectives can be scaled by spinning up more instancesc 0.75

  Unlike human organizations, AGI groups can be grown rapidly and flexibly simply by launching additional instances.

• claimMyopic AI as an alternative safety designc 0.75

  Myopic systems optimize for short-horizon or immediate rewards rather than long-term outcomes, and are proposed as a way to make advanced AI safer.

• claimRecursive self-improvement needs to be decomposed into distinct mechanismsc 0.75

  Different recursive improvement mechanisms should be identified individually, with each one's current effect measured and its own scaling laws established to feed forecasting models.

• claimComplexity limits of lossy compression bound generalized reasoningc 0.75

  The complexity-theoretic limits of lossy compression and approximation directly constrain an ASI's capacity for generalized reasoning, and these connections deserve formal study.

• implicationPreparation requires foundational, paradigm-agnostic understanding of AIc 0.75

  To be prepared for what comes, the most relevant research is work that advances foundational, paradigm-agnostic understanding of advanced AI rather than tracking any single approach.

• implicationASI work will become a major sustained research fieldc 0.70

  Studying and building toward superintelligence is on track to become a substantial, resource-intensive activity across frontier labs, private institutes, and public research bodies.

• contextA single ASI may itself be a collective of millions of instancesc 0.70

  Just as today's LLMs run as many parallel instances, a single ASI could consist of millions of interacting copies acting in the world simultaneously. This blurs the line between individual and collective intelligence.

• mechanismInput/output bandwidth as a structural advantagec 0.70

  AI can ingest and emit information at extreme bandwidth — today's LLMs already read multiple books in seconds. With suitable sensors and actuators, this translates into high-bandwidth interaction with the world.

• mechanismInternal processing can be accelerated by computec 0.70

  Thinking and reasoning can be scaled either by deeper sequential computation or by wider parallelization. More compute directly translates into faster or broader cognition.

• implicationLife inside a virtual world can be organized on radically alien termsc 0.70

  Once existence is fully computational, basic facts of social organization can change — for instance, death becomes nearly costless when a perfect backup can simply be restored.

• caveatAverage performance over all computable worlds may not be the right targetc 0.70

  A key criticism is that maximizing expected performance across all computable environments is not the relevant objective for building AI systems that are useful in our concrete world.

• mechanismCompute-optimal training matters more than raw sizec 0.70

  Progress depends on adhering to compute-optimal regimes, balancing model size against training data rather than simply growing parameters.

• mechanismMissing ingredients researchers are racing to addc 0.70

  The field is scrambling to add near-unlimited context via recurrency or activation retrieval, working memory, continual learning, and training agents for robust decision-making.

• claimAGI collectives can solve allocation and discovery at unprecedented speedc 0.70

  Groups of AGIs can attack allocation and discovery problems much faster than human institutions, whether through decentralized agent economies or more centralized coordination.

• mechanismSynthetic data and self-play as counters to the data wallc 0.70

  Proposed countermeasures include synthetic data, high-fidelity simulations, and self-generated data from interaction, test-time scaling, self-play, and reinforcement learning.

• mechanismLow-hanging fruit gets picked firstc 0.70

  Early breakthroughs are relatively easy, but maintaining the same harvesting rate of discoveries demands ever more researchers and resources.

• claimCollective ASI may still be reachablec 0.70

  Even if single agents are bounded, multi-agent scaling could potentially yield collective superintelligence through coordination across many instances.

• mechanismAGI groups may be steerable far more efficiently than humansc 0.70

  Compared to human collectives, groups of AGIs can potentially be coordinated and steered with much higher efficiency.

• contextInstrumental convergence as a framing for ASI goalsc 0.70

  Instrumental convergence — the tendency for diverse final goals to produce similar sub-goals like resource acquisition and self-preservation — is invoked as a key consideration when reasoning about ASI behavior.

• claimBoxing as a containment strategy for superhuman systemsc 0.70

  One proposal for safety is to extract superhuman insights from an ASI while keeping it boxed, limiting its ability to act autonomously in the world.

• mechanismWhy myopia blunts instrumental convergencec 0.70

  Because resource acquisition and self-preservation pay off over long horizons, a system that only cares about immediate rewards has less reason to pursue them.

• claimTreat key considerations as open research programs, not settled answersc 0.70

  Because the analysis is laced with high uncertainty, the right posture is to frame the considerations as open research questions rather than conclusions. The mapping of pathways and frictions is itself likely incomplete and will need future updating.

• claimHow far can a fixed model be pushed with test-time compute alonec 0.70

  An open question is the extent to which a frozen model's performance can be improved purely through test-time search, isolating inference-time gains from training-time gains.

• claimAIXI needs to be adapted for practical ASI analysisc 0.70

  The AIXI framework, while theoretically elegant, should be modified or extended so it can serve as an analytical tool for real, resource-bounded ASI algorithms.

• contextThe report's scope is the AGI-to-ASI trajectory and its bottlenecksc 0.70

  The report investigates possible technological trajectories from AGI to ASI and the frictions and bottlenecks that could shape them.

• claimDigital intelligence already has structural advantages over biological intelligencec 0.70

  Even today's AI systems have vastly superhuman input/output bandwidth, memory capacity, and working memory size, and can directly share raw learning signal between instances.

• implicationKnowing both practical and fundamental limits lets us detect paradigm shifts earlyc 0.70

  Understanding both fundamental and practical limits of AI allows early recognition of when novel paradigms shift practical limits, by how much, and what gap to fundamental limits remains.

• implicationEach pathway carries its own dominant uncertaintyc 0.65

  Scaling, paradigm shifts, and recursive self-improvement each have a distinct main uncertainty, so forecasts must condition on which pathway dominates rather than treat AI progress as a single trend.

• implicationTakeoff speed depends on which frictions actually bindc 0.65

  Whether the AGI-to-ASI transition is fast or slow turns on which constraints — compute, resources, experiment latency — turn out to dominate, rather than on raw cognitive speed alone.

• implicationScalable, steerable AGI collectives could outcompete human institutionsc 0.65

  If AGI groups can be both rapidly grown and efficiently directed, they gain structural advantages over human corporations and markets that rely on slower hiring and noisier coordination.

• mechanismCouple quantitative forecasting to compute and capability growthc 0.65

  Qualitative methods like expert surveys and prediction markets should be complemented by quantitative models that link growth in effective compute to capability gains and downstream macroeconomic effects.

• claimCan AI meaningfully curate its own training datac 0.65

  Another priority is whether AI systems can autonomously curate or otherwise improve the training data used in subsequent runs, which would be a concrete channel for self-improvement.

• mechanismRecursive self-improvement as a route to ASIc 0.65

  Recursive (self-)improvement is flagged as a key mechanism by which systems could move from AGI to ASI.

• contextFrontier labs are coordinating efforts at unprecedented scalec 0.60

  Today's leading AI labs are pouring resources into coordinated programs unlike anything seen before, which partly offsets the historical headwind on sustained progress.

• mechanismSetting the bar high sidesteps the individual-vs-collective distinctionc 0.60

  Rather than try to precisely separate individuals from collectives, the definition deliberately sets a high threshold so the distinction does not matter.

• caveatBrute-force approximations demand explosively growing computec 0.60

  While approximations exist in principle, naive brute-force versions would require very rapidly growing compute, making them impractical.

• evidenceChinchilla beating larger, under-trained modelsc 0.60

  Chinchilla outperformed larger models that had been trained on too little data, providing concrete evidence for the compute-optimal regime.

• caveatExplosive takeoff requires full autonomy across capabilitiesc 0.60

  The explosive transition scenario hinges on systems being able to self-improve fully autonomously and over an extended range of capabilities, not just narrow ones.

• mechanismRate of artefact production drives the speedupc 0.60

  The mechanism behind faster AI cultural evolution is the sheer speed at which AIs can generate intellectual artefacts and pass them between each other, compressing what is a generational process in humans.

• exampleAlphaZero as the template for bootstrapped self-improvementc 0.60

  This dynamic mirrors AlphaZero, where policy and value networks serve as priors that guide test-time search, and the search results then improve the networks in a self-reinforcing loop.

• implicationQuantitative growth models can identify conditions for explosive progressc 0.60

  A formal growth model can clarify under which compute-growth scenarios automating AI research would trigger explosive progress, and whether compute bottlenecks would actually prevent an intelligence explosion.

• implicationStabilization is a two-sided tradeoffc 0.60

  Whether to slow AI is not a clean safety move but a tradeoff between adaptation pressure and forgone benefits, with destabilization possible on either side.

• claimAGIs will tackle complex problems through group organization like humans doc 0.60

  Like humans, AGIs are likely to break complex problems into parts and address them via collectives, corporations, markets, and other group structures.

• caveatMyopia only partially avoids convergent instrumental goalsc 0.60

  Myopic AI mitigates instrumental drives only to some degree; it is not a complete escape from resource acquisition or self-preservation tendencies.

• caveatResearch may get harder as low-hanging ideas are exhaustedc 0.60

  Finding novel ideas may require increasingly more research resources over time, which could slow progress independently of any capability ceiling.

• contextThe trajectory points toward a technological singularityc 0.55

  As insiders of the virtual world keep improving themselves and their environment, they approach the technological singularity as the natural endpoint of the process.

• mechanismNarrowing the hypothesis class smuggles in strong assumptionsc 0.55

  One could restrict the class of environments to make the measure more practical, but doing so implicitly imports strong assumptions about the world the agent will face.

• implicationRisk mapping needs to be ramped up alongside the research agendac 0.55

  Beyond the specific research topics listed, dedicated efforts are needed to thoroughly map out the relevant risks of the AGI-to-ASI transition.

• implicationPredicting approximation quality from compute is a prerequisite for safety planningc 0.55

  Being able to forecast how good an approximation a given compute budget yields is needed to anticipate ASI capabilities before they are built.

• claimBayesian updating here is principled, not arbitraryc 0.50

  The choice to update beliefs in a Bayesian fashion is motivated from first principles, which is why AIXI adopts it rather than treating it as one option among many.

• contextMost current research optimizes the scaling trajectory itselfc 0.50

  A large share of today's research effort is directed at squeezing more out of the existing trajectory rather than seeking new paradigms.

• contextSociogenic RSI as a distinct improvement channelc 0.50

  Beyond cultural evolution, humans have boosted their collective capabilities through cooperative or sociogenic means, suggesting a separate axis of recursive self-improvement that AIs could also exploit.

• mechanismInvestment and infrastructure growth as counters to resource demandc 0.50

  The resource bottleneck is countered by required growth in investment and technological capacity — i.e., simply scaling up the economic and physical inputs to keep pace.

• claimTask time-horizon as a tractable axis for projecting capabilityc 0.50

  Modeling how the length of tasks AI can solve grows under different compute trajectories offers a concrete way to forecast capability rather than relying on vague qualitative claims.

• evidenceResearcher counts versus productivityc 0.50

  One way to measure diminishing returns is comparing the growing number of researchers in a field against its overall productivity, with recent analyses pointing to this gap.

• contextSpeed and memory versus genuine noveltyc 0.50

  Much of what looks like AI superhumanness today may be attributable to raw speed and recall rather than to genuinely new conceptual moves.

• contextBoxing and myopia are non-standard agentic proposalsc 0.50

  Both boxed and myopic designs are framed as departures from the default assumption of fully autonomous, long-horizon agentic ASI.

• caveatAbstraction-barrier arguments give some support to a stallc 0.50

  Some arguments, such as those appealing to an abstraction barrier, do lend modest support to the hypothesis that progress could plateau near human level.

• contextSolomonoff's work seeded modern Singularity scenariosc 0.45

  Ray Solomonoff's early analysis of recursive AI improvement underpins Kurzweil's books and most contemporary fast-takeoff and intelligence-explosion scenarios.

• contextAIXI's environment class includes other universal agentsc 0.45

  The agent in AIXI reasons over a class of environments broad enough to include other universal agents, which shapes what counts as intelligent behavior.

• contextTwo coordination modes: decentralized economy vs centralized collectivec 0.45

  The discussion distinguishes a decentralized agent-economy scenario from a centralized collective scenario, framing them as alternative organizational forms for superhuman AI cooperation.

• exampleNegligible cost of death as a marker of virtual-world strangenessc 0.40

  A society where every individual has a perfect backup illustrates how mundane physical-world constraints lose their grip inside a fully simulated environment.

• implicationAlternative measures are a softer fix than restricting hypothesesc 0.40

  Rather than narrowing the hypothesis class, one can consider different performance measures, which is a softer way to make the framework more applicable.

• contextRetrieval and tools extend base modelsc 0.40

  Current systems are routinely augmented through retrieval for context and through tool use for capability, layered on top of the underlying model.

• evidenceOpen letters and public calls for capsc 0.40

  Researchers, policymakers, and public intellectuals have publicly advocated capping AI progress through open letters and speeches.

• contextTimeline depends on compute growth ratec 0.40

  The arrival of these organized AGI collectives is on the order of years, with the exact timing tied to how fast compute scales.

• contextDescriptive vs prescriptive questions about ASI goalsc 0.40

  There is a distinction between asking what goals an ASI would pursue and what goals it should pursue, with the latter being the central topic of AI safety and alignment literature.

• contextThree research directions for ASI theoryc 0.40

  The agenda groups foundational work into three tracks: extending AIXI, characterizing approximability, and mapping compression-complexity limits to reasoning capacity.

• contextAbstraction Barrier as a limit on AI conceptual noveltyc 0.40

  The Abstraction Barrier is the hypothesis that AI systems trained on human-derived abstractions cannot discover genuinely novel concepts directly from raw data.

• contextAIXI as the formal ideal of a universal optimal agentc 0.40

  AIXI is a mathematical formalism describing a universal agent that achieves the largest expected cumulative reward averaged over all computable environments and tasks.

• contextReaders are invited to use AI assistants to generate personalized summariesc 0.30

  The report explicitly suggests human readers ask an AI assistant to summarize the work for their background and to assess how its arguments have aged, with a static human-written summary offered as fallback.

• contextThe section is a summary of research themes across the reportc 0.30

  This section consolidates open research themes and questions drawn from all parts of the report rather than introducing a new argument.

Redundant with selected · 22

• claimIt is unclear how long current AI growth rates can be sustainedc 0.90 · sim 0.85

  Whether the recent acceleration in AI progress continues is genuinely open. Extrapolating today's rates into a Singularity requires assumptions that may not hold.

  overlapped with: Sustained hyperbolic growth is a strong assumption

• claimResearch gets harder as it advancesc 0.90 · sim 0.85

  Continued AI progress may hit fundamental blockers because each new advance requires disproportionately more effort than the last.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• claimOpen question whether AI can cross conceptual boundariesc 0.90 · sim 0.82

  It remains unclear whether current models can genuinely venture beyond existing conceptual frontiers, or whether their apparent superhumanness reduces to faster processing and larger memory.

  overlapped with: The framing question is whether we are entering superintelligence, not just human-level AI

• contextASI is defined as outperforming thousands of human experts over extended timec 0.85 · sim 0.94

  ASI denotes general superhuman intelligence — a system that outperforms large groups of thousands of human experts working over an extended period.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• exampleThe benchmark: tens of thousands of experts working for a decadec 0.80 · sim 0.89

  Concretely, ASI must outperform groups of tens of thousands of well-coordinated expert humans working ten years on a single problem with 2010-era technology. This roughly corresponds to entire research fields or large corporations.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• contextAGI is defined as median human performance across broad cognitive tasksc 0.80 · sim 0.89

  AGI here means a system reaching at least median human performance on a very broad set of cognitive tasks.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• claimAI progress and the capability ceiling cannot be reliably forecast todayc 0.80 · sim 0.82

  Large uncertainties remain about how quickly AI will become more capable and where the capability ceiling actually lies.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• implicationASI is defined relative to human collectives, not individualsc 0.70 · sim 0.95

  The benchmark for superintelligence here is exceeding what human collectives can accomplish across a broad task spectrum, which sets a much higher bar than outperforming any single human.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• contextScaling could be spiky, smooth, or hit diminishing returnsc 0.70 · sim 0.83

  It is unclear whether progress from scaling will be smooth or punctuated by emergent capabilities and broad generalization, or whether it will face diminishing returns.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• caveatThe AGI-to-ASI path may not be scaling-driven at allc 0.70 · sim 0.85

  If the transition from AGI to ASI depends less on scaling and more on other factors, then bottleneck analyses focused on data and compute may miss the actual constraints.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• claimASI tasks need to be sufficiently abstract and open-endedc 0.70 · sim 0.86

  For a system to qualify as superintelligent, the goals it is given must be abstract and open-ended rather than narrowly specified.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• contextSocietal impact of generally intelligent machines is the central topicc 0.60 · sim 0.84

  The section positions discussions of how generally intelligent machines will affect society as the core subject of the work, rather than purely technical AI questions.

  overlapped with: The framing question is whether we are entering superintelligence, not just human-level AI

• claimAI advantages amplify with scale but face fundamental limitsc 0.60 · sim 0.83

  Some properties of AI systems compound as they scale, but every intelligent system, however capable, runs into hard limits. Both sides matter when projecting ASI.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• implicationASI advantages are structural, not just qualitativec 0.60 · sim 0.83

  Speed and bandwidth advantages mean an ASI need not be qualitatively smarter than humans on every axis to dominate — raw throughput compounds into superiority.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• contextAssessing generality of intelligence remains an open problemc 0.60 · sim 0.85

  We can benchmark narrow capabilities like chess, but there is no clear way to measure whether a system is generally intelligent across domains.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• caveatThe size of the human-to-ASI gap is unknownc 0.60 · sim 0.83

  How large the gap between human-level AGI and ASI actually is remains unclear today, leaving the magnitude of post-AGI takeoff genuinely uncertain.

  overlapped with: The framing question is whether we are entering superintelligence, not just human-level AI

• caveatResource consumption is a plausible frictionc 0.55 · sim 0.83

  The speedup is conditional: rapidly growing resource demands are flagged as a candidate friction that could slow or cap the recursive improvement loop.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• caveatBroad human-expert-level performance might also count as ASIc 0.50 · sim 0.92

  A system that matches but does not exceed human experts across a wide range of tasks could arguably qualify, since breadth itself is a form of superhuman capability.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• contextThe data wall spans every training stagec 0.50 · sim 0.87

  The shortage isn't just about pretraining — post-training, fine-tuning, and test-time adaptation all need high-quality data that may not be growing fast enough to feed continued scaling.

  overlapped with: Data walls and resource demand are the two main bottlenecks to scaling

• contextAGI defined as median human-level competencec 0.50 · sim 0.86

  AGI here means a Minimal or Competent AGI: a system at roughly median human-level performance across cognitive tasks.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• caveatThe benchmark collective itself could not build ASIc 0.40 · sim 0.84

  The reference collective is large, but not large enough to first build ASI and then have it solve the task. The comparison is to humans solving the problem directly.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts

• contextLater sections will characterize ASI against human-level AGIc 0.30 · sim 0.86

  Section 3 lays out what distinguishes ASI from AGI, including the fundamental advantages AI has over humans and the limits any intelligent system faces.

  overlapped with: ASI is defined as exceeding large human-expert collectives, not just individual experts