Four Paths from AGI to ASI and the Six Walls That Could Halt Progress

Three Core Concepts

The paper defines AGI as "competent AGI" at median human level across most cognitive tasks, noting that current models like AlphaFold and AlphaGo excel in narrow domains but are not yet general. ASI is described as an intelligence that can reliably outperform a coordinated team of tens of thousands of top experts over a decade, potentially composed of millions of instances. UAI (AIXI) is presented as the theoretical upper bound on intelligence, optimal but incomputable, analogous to thermodynamics.

Digital Intelligence’s Innate Advantages

The authors list six advantages of digital over biological intelligence: (1) I/O speed – large language models can read multiple books in seconds; (2) internal processing speed – faster serial or parallel computation; (3) working memory – far larger than human capacity; (4) substrate independence – AI can migrate across hardware; (5) lossless copying – both code and memory states can be duplicated; (6) high‑bandwidth experience sharing – data streams can be stored, replayed, and shared among agents.

A counter‑intuitive drawback is that humans, constrained by low communication bandwidth, develop deep internal models, whereas AI’s high‑bandwidth I/O may reduce the need for such abstractions.

Limits on ASI

The paper enumerates physical and theoretical constraints: speed of light, Landauer’s principle, Bremermann’s limit, Bekenstein bound, real‑time operation, physical manufacturability, cognitive limits from measurement precision, computational complexity (P vs NP vs PSPACE), and logical limits (Gödel’s incompleteness, halting problem). These limits do not directly predict specific capabilities such as curing aging or building Dyson spheres.

AIXI Framework: Theoretical Upper Bound

AIXI models an agent interacting with an unknown computable environment, updating Bayesian posteriors over all computable hypotheses, solving action under uncertainty, credit assignment, and exploration‑exploitation trade‑offs. Its optimality under the Legg‑Hutter score is acknowledged as incomputable, but lower‑bound algorithms can approximate it with more compute.

The authors argue that modern large‑scale pre‑training can be seen as a resource‑bounded approximation to universal compression, suggesting a possible path toward ASI without fundamental theoretical obstacles, though the argument remains tentative.

Four Paths from AGI to ASI

1. Continuous Scaling

Effective compute growth is broken into hardware improvements (~1.5× per year, 60 years of Moore’s law), investment growth (~2.5× per year over the last decade), and algorithmic efficiency gains (~3× per year, with some estimates up to 6×). Combined, these yield a conservative ~10× annual increase in effective compute.

Assuming AGI’s base model plateaus while compute continues to grow 10× yearly, the number of AGI instances could rise from 1 000 to 10 000 in one year and to 100 million in five years, raising the question of whether quantitative growth triggers qualitative change.

Potential frictions include the looming exhaustion of high‑quality text data within the decade, and the need for synthetic data or distillation, which carries degradation risk.

2. Paradigm Shift

The current paradigm relies on massive supervised pre‑training of Transformers followed by instruction‑tuning, RL‑tuning, and test‑time extensions (chain‑of‑thought, retrieval‑augmented generation, tool use). The authors distinguish “evolution” (incremental advances such as infinite context, working memory, continual learning) from “transformation” (radical new architectures like spiking networks, neuromorphic hardware, simulation‑based learning).

They stress that true paradigm shifts are hard to predict but must be considered, as they could dramatically alter the scaling trajectory.

3. Recursive Self‑Improvement

Four modes of recursive improvement are outlined: genetic (code and hardware blueprints), cultural (data‑driven self‑enhancement, automated dataset generation, recursive distillation), cooperative (division of labor among agents), and algorithmic (neural architecture search, automated hyper‑parameter tuning, AI‑assisted chip design). Existing examples include FunSearch and AlphaEvolve.

Theoretical work such as Schmidhuber’s Gödel machines and Christiano’s iterated amplification is cited, while noting practical limits like resource explosion for physical experiments.

4. Multi‑Agent Collaboration

Superintelligence may emerge from millions of AGI agents cooperating, either through coordinated design or self‑organized market dynamics, forming an AI economy whose collective intelligence exceeds any single agent.

The authors admit that understanding emergent dynamics in large multi‑agent systems remains a major open research area.

Six Potential Walls

1. Data Wall – high‑quality text may run out within a decade, with synthetic data posing degradation risks.

2. Resource Wall – limits in energy, chips, rare earths, and infrastructure constrain compute growth.

3. Paradigm Ceiling – the Transformer‑pre‑train‑fine‑tune pipeline may hit a fundamental ceiling before reaching AGI.

4. Research‑Effort Wall – sustaining exponential research progress likely requires exponential economic investment.

5. Abstraction Wall – AI may struggle to discover concepts beyond those present in human data, tied to the “embodiment factor” hypothesis.

6. Deliberate Slow‑down – regulation, accidents, or public backlash could intentionally curb development.

Interesting Detail

The paper itself invites readers to use an AI assistant to generate customized summaries, providing explicit prompting instructions and deliberately refusing to be fully summarized by static human text.

Core Judgments

AGI emergence is expected to be a series of incremental transformations rather than a single breakthrough, requiring global interdisciplinary coordination.

Measuring, modeling, and forecasting AI progress will become a distinct research field demanding sustained funding from labs, private institutes, and public agencies.

The central tension lies in how much AI can accelerate broader scientific progress versus how much the required research effort and economic input can offset that acceleration; both forces will coexist in a race‑like dynamic.

Even a superintelligent ASI will remain bounded by physical laws, computational complexity, and logical limits, making precise predictions of its ultimate capabilities highly uncertain.