From AGI to ASI

The report starts with three drivers of effective compute: better hardware, more investment, and more efficient algorithms. The authors use an illustrative combined rate near 10× per year. Continued growth would allow larger training runs, more model instances, faster inference, and more test-time search.

Capability gains need not track compute gains. Diminishing returns could turn exponential compute into slow progress. Better algorithms or AI-assisted research could push the other way. A strong feedback loop between improved AI and faster AI R&D could produce super-exponential growth until physical, economic, or technical frictions slow it.

Historical evidence supports rapid growth within measured regimes. Epoch AI estimates frontier training compute grew roughly 4–5× per year during much of the deep-learning era. Published estimates of algorithmic efficiency vary by task and metric. Neither result establishes a durable 10× effective-compute rate.

The authors therefore reject one fixed forecast. They call for several quantitative models, frequent updates, and direct measurement of compute, efficiency, automation, and capability.

AGI means roughly median human performance across most cognitive tasks. ASI means broad performance beyond large, well-coordinated expert groups. The comparison is system-level: a collective of ordinary human-level agents could qualify if the group becomes broadly superhuman.

The paper uses the Legg-Hutter measure as a formal reference. It defines intelligence as expected performance across computable tasks, weighted toward simpler environments. This gives the authors a continuum from weak systems through AGI and ASI toward a theoretical maximum.

Digital intelligence also scales differently from biological intelligence. Models can be copied, sped up, run in parallel, moved across hardware, given large memories, and connected through high-bandwidth channels. More compute can expand each advantage.

The formal measure has limits. Its value depends on the selected universal machine and task weighting, and it cannot be computed exactly. Hardware gains can also hit latency, memory-bandwidth, energy, and coordination constraints. The framework grounds the terms without turning them into directly measurable thresholds.

AIXI is the report's theoretical endpoint. It combines Solomonoff induction with sequential decision-making: maintain a weighted mixture over computable environments, update it with observations, and choose actions that maximize expected reward.

The construction captures prediction, credit assignment, exploration, and planning in one agent. It also shows why ASI would remain bounded. No agent escapes the speed of light, finite energy, incomplete observations, computational complexity, real-time experiments, or logical limits.

Exact AIXI is incomputable. Its optimality depends on the chosen prior, reward function, horizon, and environment class. Practical variants restrict compute, time, or hypotheses and lose parts of the original guarantee. Current neural systems do not provide a demonstrated engineering path to AIXI. The framework supplies a limit and a language for constraints.

The first route extends current practice: train larger or better-allocated models on more data, add test-time compute, and run more instances. Even if one AGI stays near human level, millions of faster copies might form a broadly superhuman system.

Scaling laws support predictable loss reductions over tested model families. The Chinchilla experiments also showed that model size and training data must grow together: a smaller, better-trained model beat much larger undertrained systems at the same compute budget.

Those results cover finite ranges and specific evaluations. They do not show that scaling reaches AGI or ASI. Some apparent emergent abilities also disappear under continuous metrics, which means benchmark jumps can reflect scoring choices rather than sudden new competence.

The route depends on sustained data, energy, chips, capital, and algorithmic efficiency. It also depends on group capability rising faster than coordination costs when population scaling replaces individual-model gains.

The authors separate likely extensions of current systems from deeper algorithmic shifts. Near-term extensions include retrieval, tool use, test-time search, external memory, continual learning, interactive training, longer sequence processing, and learned world models.

These mechanisms already exist in restricted forms. Retrieval-augmented generation connects a model to updateable documents. Neural Turing Machines add differentiable external memory. Mamba processes sequences with linear scaling in sequence length and reported strong results up to million-token tests.

The evidence is narrower than several phrases in the paper. Retrieval does not provide perfect recall. Linear-time sequence processing does not create unlimited context. Memory systems can retrieve bad evidence or lose information over long horizons. None of these systems establishes progress from current models to AGI.

True architectural or optimization shifts are harder to forecast because they change the constraints being forecast. The paper treats them as a live route with little basis for timing.

Recursive improvement includes better model code, hardware, data, and division of labor. An improved system contributes to the next improvement, which may strengthen the following system.

Narrow loops already work. AlphaGo Zero used self-play and search to generate stronger training targets, then distilled them into its policy and value networks. AlphaEvolve combines language-model proposals with automated evaluators and evolutionary search; Google DeepMind reports uses in algorithms, chip design, data centers, and AI training. AI Scientist systems can propose experiments, run code, and draft papers in bounded machine-learning settings.

These systems rely on clean objectives, automated scoring, human problem selection, and deployment review. The AI Scientist paper also judged outputs partly through an automated reviewer. Physical science adds experimental equipment, real-time delays, safety controls, and replication.

Current evidence shows AI accelerating pieces of R&D. It does not show sustained autonomous improvement across open-ended AI research. The feedback could stall when verification, experiments, data quality, compute, or research difficulty grows faster than the improvements.

The fourth route moves the unit of analysis from one model to a group. AI agents could form companies, markets, research teams, or self-organizing networks. Specialization and delegation might let the group solve problems beyond any member.

Group-agency theory explains how an organization can develop beliefs, goals, and actions that are not reducible to one participant. Current multi-agent experiments provide small examples of delegation and coordination gains.

They also show failure. Agents drift from the problem, copy errors, settle into poor equilibria, collude in auctions, or lose performance as the population grows. Results depend on task structure, communication, incentives, topology, and the quality of the component models.

More instances therefore supply compute and labor. They do not guarantee more intelligence. The paper identifies multi-agent scaling laws as an open research problem.

The report lists bottlenecks across the whole system. High-quality human text is finite. Synthetic data can degrade under repeated naive training. Energy, memory bandwidth, chips, fabrication, data centers, capital, and suitable sites may fail to scale at the required rate. AI research may get harder as easy gains disappear.

Current architectures may also lack continual learning, robust agency, grounded concept discovery, or reliable operation in the physical world. Automated research faces long experiments and weak feedback. Large agent groups add communication, incentive, and governance costs. Regulation or public resistance can slow deployment.

Each friction has a possible counterforce. Search and self-play can produce useful synthetic targets. Better algorithms reduce compute and data requirements. AI can help design chips and infrastructure. Simulation can accelerate some experiments. Investment can rise when deployment returns rise.

The evidence does not fix the net effect. Text-data exhaustion estimates depend on model and dataset assumptions. Model collapse appears under particular synthetic-data loops and is avoided in settings with strong verification. Energy and bandwidth are real constraints, though neither has been shown to be a permanent ceiling. The paper's “abstraction barrier”—that models trained on human concepts may fail to discover new grounded concepts—remains a hypothesis.

Its policy discussion also needs current context. The EU's compute threshold triggers systemic-risk duties rather than a blanket license for frontier training. The United States revoked Executive Order 14110 in January 2025 and shifted federal policy toward faster development with separate security controls.

The remarks section tests the framework against harder questions. More compute can improve search, yet restricted model classes and bad priors can impose limits that brute force cannot remove. Groups of AGIs may bypass some individual limits, depending on the task.

ASI capability cannot be predicted cleanly from complexity theory. Hard exact solutions may allow strong approximations, and finding the approximation can require running the computation. Scaling laws and benchmark stitching give empirical forecasts, though human-level benchmarks saturate and private tests can hide contamination or overfitting.

Several details need correction or qualification. Complexity results for generalized chess do not prove that fixed chess can never be played perfectly by a finite machine. Benchmark saturation claims depend on the benchmark version and date. Formal work on instrumental convergence, corrigibility, safe interruption, and knowledge-seeking agents relies on defined objectives and simplified environments. It does not guarantee the behavior of frontier systems.

The paper also conflates two systems. Lu et al.'s 2024 AI Scientist is an autonomous research agent. Yoshua Bengio and collaborators' 2025 Scientist AI proposal is the non-agentic world-model and oracle design discussed in the safety passage.

The authors propose work on seven fronts: scaling frictions, quantitative forecasting, post-human benchmarking, recursive-improvement dynamics, multi-agent scaling, superintelligence theory, and safety with social effects.

They want forecasts that join compute, algorithms, automation, and economics, then update as measurements change. Epoch AI's GATE model does this in a scenario sandbox. Its documentation says the outputs are not Epoch's forecasts and depend on chosen assumptions.

METR's task-completion time horizon offers one longitudinal capability measure. Its 2025 study found an approximately seven-month historical doubling on a bounded set of software and reasoning tasks, with clear external-validity limits. The 2026 International AI Safety Report also describes capability evaluation as fragmented, vulnerable to contamination, and weak at predicting real-world use.

The paper's agenda matches these gaps. It also assumes alignment is solved well enough to study post-AGI capability in isolation. The authors call that a strong working assumption. Unsafe or uncontrollable systems could themselves block research automation and deployment.

The authors think progress is unlikely to stop exactly at human level. They see two more plausible outcomes: current methods plateau before AGI, or AGI continues into weak ASI through individual improvement, more compute, or organized populations. Recursive improvement could make the transition faster.

They give this judgment low confidence. A survey of 2,778 AI researchers found wide timeline dispersion, with a 50% aggregate date of 2047 for unaided machines outperforming humans in every task and 2116 for full automation of all occupations.

Multi-agent evidence also cuts both ways. Some coordination methods improve bounded tasks. AgentsNet found that some frontier models performed well in small networks and worsened as the network grew.

The paper closes with a scenario, not a point forecast: movement from AGI into ASI within one or two decades is possible enough to justify preparation. Its strongest recommendation is continuous measurement across several trajectories.

Appendix A restates the definitions, four routes, bottlenecks, and research questions in the same order. It keeps the uncertainty visible. Scaling may continue or hit data and resource limits. New algorithms may change efficiency. Recursive loops may accelerate or stall. Agent groups may amplify capability or coordination failure.

The appendix is the authors' approved compression of the report. It does not supply separate evidence. Every mechanism and forecast retains the qualifications attached to the corresponding main section.