pith. sign in

arxiv: 2606.12683 · v1 · pith:XUKB3HHInew · submitted 2026-06-10 · 💻 cs.AI · cs.CY· cs.LG

From AGI to ASI

Tim Genewein , Matija Franklin , Alexander Lerchner , Laurent Orseau , Samuel Albanie , Adam Bales , Cole Wyeth , Stephanie Chan
show 6 more authors
Iason Gabriel Joel Z. Leibo Allan Dafoe Marcus Hutter Thore Graepel Shane Legg
This is my paper

Pith reviewed 2026-06-27 09:43 UTC · model grok-4.3

classification 💻 cs.AI cs.CYcs.LG
keywords artificial general intelligenceartificial superintelligenceAI development pathwaysrecursive improvementmulti-agent collectivesAI societal impactstechnological acceleration
0
0 comments X

The pith

Achieving AGI could lead to a series of AI-driven societal changes rather than one single leap to superintelligence.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates how machine intelligence might progress after human-level AGI is reached, focusing on the path to ASI. It identifies four pathways for this development and examines frictions that could influence their speed and form. The central suggestion is that uncertainties allow for continued acceleration of AI progress, which would mean society experiences repeated transformations enabled by advances across science and technology instead of one abrupt shift. A reader would care because this framing changes how preparation for future AI might need to be structured over time.

Core claim

After defining ASI as a system more intelligent and capable than large organizations of humans, the report lays out four pathways from AGI: scaling existing systems, shifts to new AI paradigms, recursive self-improvement, and emergence within large multi-agent collectives. It then covers possible frictions and bottlenecks on these paths. Large uncertainties prevent ruling out accelerated progress, which would replace the picture of one transformative step change with a series of changes driven by AI-enabled breakthroughs in many areas.

What carries the argument

The four pathways from AGI to ASI together with the frictions and bottlenecks that may affect their pace and shape.

If this is right

  • Frictions along the pathways could slow or reshape the arrival of ASI in concrete ways.
  • Concrete open research questions on the size of those frictions need answers to assess timelines.
  • AI progress might keep accelerating in the near term across multiple domains.
  • Preparation for AI impacts requires a broad interdisciplinary effort rather than focus on a single event.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Advances enabled by AGI in one scientific area could feed back into faster progress along another pathway.
  • Policy and planning might need to treat technological change as continuous rather than preparing for one discrete jump.
  • Early indicators from current scaling trends could help test which pathway is gaining the most traction.

Load-bearing premise

The four pathways and the listed frictions are the main factors determining the speed and shape of post-AGI development.

What would settle it

Clear evidence after an AGI is built that one pathway reaches ASI rapidly with negligible friction, producing a single abrupt societal transformation instead of repeated changes.

read the original abstract

Over the last decade, building human-level artificial general intelligence has moved from far-fetched speculation to being a concrete next-decade target for many of the largest AI organisations. Achieving this goal would have profound and far-reaching impacts on human society, which raises many complex questions for the decade ahead. This report investigates how AI itself might continue to develop in a post-AGI world along the continuum of machine intelligence. The endpoint of this continuum, Universal AI, is theoretically well understood, which provides some formal grounding for the main focus of this report: the transition from human-level AGI to artificial general superintelligence, which, intuitively, can be understood as a system that is more intelligent and cognitively capable than large organisations of humans. After characterizing ASI, the report discusses four potential pathways from AGI to ASI: scaling AGI, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi-agent collectives. The report then discusses possible frictions and bottlenecks along these pathways. Determining whether the impact of these frictions will be negligible or substantial raises a number of concrete open research questions. Due to large uncertainties for predicting ASI progress, it cannot be ruled out that AI progress might continue to accelerate over the next years. This could imply that the image of a single transformative step change, caused by the introduction of human-level AGI into our society, could be inaccurate. More apt might be the prospect of a series of transformative societal changes caused by AI-enabled progress and breakthroughs across many areas of science and technology. Preparing for this prospect requires a massively interdisciplinary endeavour of global scope and interest.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript characterizes artificial superintelligence (ASI) as systems more capable than large human organizations and examines the transition from AGI along four pathways—scaling AGI, AI paradigm shifts, recursive improvement, and emergence from large-scale multi-agent collectives—while enumerating possible frictions. It concludes that post-AGI progress may consist of a series of transformative societal changes driven by AI-enabled breakthroughs rather than a single step change, and frames the discussion as raising concrete open research questions.

Significance. The report supplies a structured enumeration of pathways and frictions that could usefully organize future work on AI progress forecasting and governance, provided the listed factors are shown to be representative. Its explicit framing of claims as open questions rather than predictions is a strength for a speculative topic.

major comments (1)
  1. [Abstract] Abstract: the premise that the four listed pathways and frictions constitute the main determinants of post-AGI development speed and shape is stated without comparative justification, literature synthesis, or argument that other factors (e.g., regulatory, economic, or hardware constraints) are secondary; this assumption directly supports both the open-question list and the claim that multiple transformative changes are more likely than a single step.
minor comments (2)
  1. The manuscript would benefit from a summary table or explicit subsection listing each pathway together with its associated frictions to improve readability.
  2. The claim that 'Universal AI' is 'theoretically well understood' requires at least one specific citation to the relevant formal results.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment and recommendation of minor revision. The single major comment concerns scoping and justification in the abstract; we address it directly below and will revise accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the premise that the four listed pathways and frictions constitute the main determinants of post-AGI development speed and shape is stated without comparative justification, literature synthesis, or argument that other factors (e.g., regulatory, economic, or hardware constraints) are secondary; this assumption directly supports both the open-question list and the claim that multiple transformative changes are more likely than a single step.

    Authors: We agree that the abstract could more clearly delimit its scope. The manuscript presents the four pathways as prominent routes discussed in the AI progress literature rather than as an exhaustive or proven set of main determinants; the text frames all claims as open questions precisely to avoid overstatement. Nevertheless, to eliminate any possible implication of primacy, the revised abstract will explicitly note that these pathways are selected for focused analysis, that external factors (regulatory, economic, hardware) can modulate their impact and are partially captured under frictions, and that the discussion of multiple transformative changes is offered as one plausible outcome under continued acceleration rather than a direct consequence of the listed pathways being exhaustive. No comparative literature synthesis will be added to the abstract itself, as it is a concise summary, but the body already situates the pathways within existing discourse. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

This is a high-level speculative discussion paper with no equations, derivations, fitted parameters, or formal claims that reduce to self-defined inputs. The four pathways and frictions are framed as topics for open research questions rather than derived results, and the text does not invoke self-citations as load-bearing uniqueness theorems or ansatzes. The analysis is self-contained as forward-looking commentary grounded in external concepts like Universal AI without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The report contains no mathematical derivations, fitted constants, or new postulated entities; its content rests on domain assumptions about the plausibility of listed pathways and the existence of identifiable frictions.

axioms (1)
  • domain assumption The four enumerated pathways and associated frictions constitute the primary determinants of post-AGI progress speed and character.
    Invoked in the abstract when the authors state that determining the impact of these frictions raises concrete open research questions.

pith-pipeline@v0.9.1-grok · 5858 in / 1319 out tokens · 17095 ms · 2026-06-27T09:43:43.741198+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MAS-PromptBench: When Does Prompt Optimization Improve Multi-Agent LLM Systems?

    cs.LG 2026-06 unverdicted novelty 6.0

    A new benchmark study finds that prompt optimization can deliver significant gains in multi-agent LLM systems but its effectiveness varies strongly with task, workflow, communication protocol, and team size.

Reference graph

Works this paper leans on

16 extracted references · 7 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    URLhttps://arxiv.org/abs/1711.05541. S. Armstrong, A. Sandberg, and N. Bostrom. Thinking inside the box: Controlling and using an Oracle AI.Minds and Machines, 22(4):299–324, 2012. doi: 10.1007/s11023-012-9282-2. L. Aschenbrenner. Situational awareness, 2024. URLhttps://situational-awareness.ai/. Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jo...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/s11023-012-9282-2 2012

  2. [2]

    URLhttps://arxiv.org/abs/1810.08575. M. Cohen, B. Vellambi, and M. Hutter. Asymptotically unambitious artificial general intelligence. Proceedings of the AAAI Conference on Artificial Intelligence, 34:2467–2476, 2020. doi: 10.1609/ aaai.v34i03.5628. J. P. Crutchfield. Between order and chaos.Nature Physics, 8:17–24, 2012. A. Dafoe. On technological determ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1177/0162243915579283 2020

  3. [3]

    URLhttp://www.hutter1.net/ai/uaibook2.htm. M. Janner, Y. Du, J. Tenenbaum, and S. Levine. Planning with diffusion for flexible behavior synthesis. InProceedings of the 39th International Conference on Machine Learning, pages 9902–9915, 2022. D. Janzing. Is there a physically universal cellular automaton or hamiltonian?, 2010. URLhttps: //arxiv.org/abs/100...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2021.naacl-main.324 2022

  4. [4]

    ISBN 978-0241625248. S. Legg.Machine super intelligence. Phd thesis, University of Lugano, June 2008. S. Legg and M. Hutter. Universal Intelligence: A Definition of Machine Intelligence.Minds and Machines, 17(4):391–444, 2007a. doi: 10.1007/s11023-007-9079-x. S. Legg and M. Hutter. A collection of definitions of intelligence.Frontiers in Artificial Intell...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/s11023-007-9079-x 2008

  5. [5]

    URLhttps://philarchive.org/rec/LERTAF. X. Leroy. Formal verification of a realistic compiler.Communications of the ACM, 52(7):107–115, 2009. P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. InAdvance...

    work page doi:10.48550/arxiv.2311.00176 2009

  6. [6]

    O’Brien, Carrie J

    URLhttps://arxiv.org/abs/2110.10819. 48 From AGI to ASI D. Paglieri, B. Cupiał, S. Coward, U. Piterbarg, M. Wolczyk, A. Khan, E. Pignatelli, Ł. Kuciński, L. Pinto, R. Fergus, J. N. Foerster, J. Parker-Holder, and T. Rocktäschel. Balrog: Benchmarking agentic llm and vlm reasoning on games. In13th International Conference on Learning Representations, 2025. ...

    work page doi:10.1145/3586183.3606763 2025

  7. [7]

    URLhttps://intelligence.org/files/Corrigibility.pdf. L. Soldaini, R. Kinney, A. Bhagia, D. Schwenk, D. Atkinson, R. Authur, B. Bogin, K. Chandu, J. Dumas, Y. Elazar, V. Hofmann, A. H. Jha, S. Kumar, L. Lucy, X. Lyu, N. Lambert, I. Magnusson, J. Mor- rison, N. Muennighoff, A. Naik, C. Nam, M. E. Peters, A. Ravichander, K. Richardson, Z. Shen, E. Strubell, ...

    arXiv 2024

  8. [8]

    53 From AGI to ASI A

    doi: 10.26599/cvm.2025.9450460. 53 From AGI to ASI A. Summary This report investigates possible technological trajectories from AGI to ASI, and discusses potential frictions and bottlenecks along these trajectories. In the report, AGI denotes a system that reaches at least median human performance on a very broad set of cognitive tasks. ASI, in contrast, ...

    work page doi:10.26599/cvm.2025.9450460 2025

  9. [9]

    Scaling of compute, models & data.Exponential scaling may continue for a number of years, as it has over the last decade and more

  10. [10]

    Algorithmic paradigm shifts.More data-, compute-, or energy-efficient algorithms and archi- tectures, as well as learning paradigms, may be discovered

  11. [11]

    Recursive(self-)improvement.AI systems may significantly, or even fully automate AIresearch and development, leading to a self-accelerating cycle of AI progress

  12. [12]

    We can only see a short distance ahead, but we can see plenty there that needs to be done

    ASI via group agent formation.AI collectives may become much more intelligent than its individual members. Scaling group size by running more instances is straightforward. While today the pathway of scaling (models & data) seems most promising to deliver progress, it is unclear how long exponential growth rates can be sustained economically and in terms o...

    2024

  13. [13]

    Business-as-usual

    Scaling compute, models & data.“Business-as-usual” scaling of model size and data to train on, that is, a continuation of what enabled the current AI breakthroughs. Exponential growth of these two factors implies exponentially increasing compute and energy demands—which may potentially be alleviated by exponentially increasing hard- and software efficienc...

  14. [14]

    get harder

    Algorithmic paradigm shifts.If scaling hits its limits (e.g., economic limits, or diminishing returns), further progress may require sharp deviations from today’s paradigm of pretraining a large base model, plus post-training, and test-time scaling & scaffolding. What these new paradigms may be and how their energy-, compute-, and data-demands are is hard...

  15. [15]

    explosive

    Recursive (self-) improvement.If AI can significantly speed up AI research and development, or even fully automate it, this could lead to recursive improvements where AI enabled R&D leads to better, faster, and cheaper AI, which will speed up AI R&D even more, and so on. Hypothetically this could lead to self-accelerating progress dynamics and an “explosi...

  16. [16]

    individual

    ASI via group agent formation.It may be possible to increase the collective intelligence of groupsofAIsmoreeasilythanimproving“individual”modelintelligence, similarlytohowgroups of humans can achieve more intellectually than individuals (typically through parallelization and diversity of skills & thinking). At the moment it is unclear for which kinds of p...