Pathways to AGI

Gordon Fletcher; Saomai Vu Khan

arxiv: 2605.06029 · v1 · submitted 2026-05-07 · 💻 cs.AI

Pathways to AGI

Gordon Fletcher , Saomai Vu Khan This is my paper

Pith reviewed 2026-05-08 10:43 UTC · model grok-4.3

classification 💻 cs.AI

keywords AGIpathwaysgenerative AIleverage nodessocio-technical systemsalternative modelscommercial AIdecision points

0 comments

The pith

AGI is conceptually and definitionally problematic, so its pursuit must account for the social and economic conditions shaping current AI tools.

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

The paper explores five questions about the pathways leading to dominant generative AI from a perspective that questions assumptions of inevitability. It traces critical decision points and alternative branches across different model types to understand what shaped today's tools. This analysis leads to suggestions for development programs that could approach advanced capabilities while emphasizing transparency and sustainability. A sympathetic reader would care because it reframes AGI not as an inevitable outcome but as one contingent on choices that can be examined and redirected.

Core claim

From the critical software studies view, the current situation with AI is not inevitable, and AGI itself is problematic conceptually and definitionally. This leads to examining the closeness of linkage between commercial AI development and prevailing circumstances, identifying pathways, leverage nodes, and alternative possibilities that can inform recommendations for more balanced socio-technical approaches to AGI-adjacent capabilities.

What carries the argument

The analysis of pathways and leverage nodes, where leverage nodes are decision points with small changes that produce large downstream effects and dead ends represent alternative possibilities that did not become dominant.

If this is right

Pathways to current generative AI tools can be traced through their capabilities, product forms, and adoption patterns.
Decision points act as leverage nodes with large effects, while dead ends show missed alternatives.
Trajectories differ between proprietary frontier models, open-weight models, and domain-specific or sovereign models.
Alternative projects from key nodes have varying states of success, stall, failure, or absorption.
Socio-technical programs can be proposed to move toward AGI-adjacent capability with requirements for transparency, moderation, wellbeing, and sustainable business models.

Where Pith is reading between the lines

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

Understanding these historical branches could inform policy decisions on AI governance to favor certain models over others.
The emphasis on avoiding inevitability assumptions might extend to questioning whether AGI is the right goal altogether.
Linking AI development to broader social structures suggests that changes in economic systems could alter future AI trajectories.

Load-bearing premise

That assumptions about the inevitability of the current AI situation can and should be avoided in order to critically examine its development.

What would settle it

Historical evidence showing that all current AI capabilities arose without significant influence from specific decision points or that no viable alternative paths existed would challenge the paper's conditional conclusions on future directions.

Figures

Figures reproduced from arXiv: 2605.06029 by Gordon Fletcher, Saomai Vu Khan.

**Figure 1.** Figure 1: The OpenAI/ChatGPT pathway view at source ↗

**Figure 2.** Figure 2: Anthropic’s Claude critical pathway view at source ↗

**Figure 3.** Figure 3: Google’s Gemini critical pathway view at source ↗

**Figure 4.** Figure 4: xAI’s Grok critical pathway view at source ↗

**Figure 5.** Figure 5: Microsoft’s critical pathway view at source ↗

**Figure 6.** Figure 6: Network of actors involved in the development of Large Language Models (LLMs) In this diagram nodes represent individuals and organisations. The edges represent relationships (such as an employment relationship, a research collaboration or an investment). Node size reflects connectivity (the number of attached edges), and colour shows the entity type. Edges are directed and weighted. The development of the… view at source ↗

read the original abstract

Our focus are five related questions that stem from a critical software studies perspective. Underpinning this view is the acknowledged need to avoid assumptions regarding the inevitability of the current situation relating to AI. What we need to see is the closeness of the linkage between current commercial AI development and our prevailing social, political and economic circumstances. This does mean that the perspectives presented here are done so critically and conditionally. Most importantly, Artificial General Intelligence (AGI) is seen as being problematic both conceptually and definitionally. This conditioning of any view regarding AGI does lead the discussion in specific directions and to certain conclusions regarding the future. However, adopting this perspective enables the work to offer some final recommendations. We set out to ask the following questions, 1. What are the critical pathways that produced the current dominant generative AI tools (capabilities, product forms, adoption patterns)? 2. Which decision points acted as leverage nodes (small changes that had large downstream effects), and which dead ends reveal alternative possibilities that did not become dominant? 3. How do pathways differ across three foundational-model trajectories such as the frontier proprietary models, open-weight models or specific domain and sovereign models? 4. Which alternative projects branched from key leverage nodes, what is their current state, and why did some succeed, stall, fail or become absorbed? 5. Based on this analysis, what socio-technical development programmes could plausibly move toward AGI-adjacent capability while meeting requirements for transparency, moderation, wellbeing and sustainable business models?

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.

Desk Editor's Note private letter to a colleague

This paper uses a critical software studies lens to question AGI's conceptual basis and map contingent pathways in AI development, but it adds no new data, models, or technical findings.

read the letter

The core point is that the authors frame AGI as problematic both conceptually and definitionally, then use that stance to ask five questions about how current generative AI tools emerged, what leverage points shaped them, how trajectories differ across proprietary, open-weight, and domain models, what alternatives branched off, and what socio-technical programs could steer toward capability with more transparency and sustainability. They stress the need to drop assumptions of inevitability and to see the tight links between commercial AI and prevailing economic conditions. This conditional approach leads directly to their recommendations without claiming to predict the future. The paper does a solid job of structuring the discussion around historical contingency and decision nodes rather than treating progress as fixed or linear. It keeps the analysis perspectival and avoids overclaiming, which makes the five questions useful as prompts for thinking about alternatives that got sidelined. The internal logic holds together on its own terms. The main limitation is the absence of concrete evidence or case studies to support the pathways described. The abstract stays at a high level, so claims about specific leverage nodes or stalled projects rest on interpretive framing rather than documented examples or data. Without those details the piece functions more as a set of guiding questions than a fully substantiated analysis, and soundness cannot be judged fully until the full arguments and sources are checked. This is aimed at readers already working in AI ethics, governance, or science and technology studies who want structured ways to discuss alternatives. Technical researchers focused on mechanisms or benchmarks will find little to engage with. It deserves peer review in a venue that handles conceptual and critical work, because the framing is coherent and the questions are worth testing against real histories even if the current version needs more grounding.

Referee Report

2 major / 4 minor

Summary. The manuscript adopts a critical software studies perspective to interrogate the socio-technical pathways that produced current dominant generative AI tools. It treats AGI as conceptually and definitionally problematic, rejects assumptions of inevitability, and structures its analysis around five questions concerning historical pathways, leverage nodes and dead ends, differences across proprietary/open-weight/domain-specific trajectories, the fate of alternative projects, and the derivation of socio-technical development programmes that could achieve AGI-adjacent capabilities while satisfying transparency, moderation, wellbeing and sustainability criteria.

Significance. If the interpretive historical analysis is substantiated with concrete case material, the paper could usefully reframe AGI discourse as contingent rather than teleological, thereby opening space for policy and research interventions that prioritise socio-technical values over unchecked scaling. Its conditional, non-inevitabilist stance is a strength that distinguishes it from much of the existing AGI literature.

major comments (2)

The foundational claim that AGI is 'problematic both conceptually and definitionally' (Abstract) is asserted without engagement with specific technical definitions or debates in the AGI literature (e.g., those centred on generality, human-level performance, or benchmark suites). Because this claim conditions the direction of all five questions and the final recommendations, the absence of such engagement weakens the load-bearing premise of the argument.
Question 5 and the concluding recommendations section: the socio-technical programmes are presented as plausible outcomes of the pathway analysis, yet the manuscript does not supply an explicit mapping showing how particular leverage nodes, dead ends, or trajectory differences identified in questions 1-4 directly support the proposed programmes. This missing linkage is central to the paper's claim to offer actionable conclusions.

minor comments (4)

Abstract: 'Our focus are five' is grammatically incorrect; revise to 'We focus on five' or 'Our focus is on five'.
Abstract: the sentence 'This does mean that the perspectives presented here are done so critically and conditionally' is awkward and redundant; a clearer phrasing would be 'This implies that the perspectives presented here are critical and conditional.'
The term 'critical software studies perspective' is used without an early definition or key references; adding a brief characterisation in the introduction would improve accessibility for an AI readership.
The five questions are enumerated but the manuscript would benefit from a short roadmap paragraph indicating which sections address each question.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each of the major comments in turn below, indicating where revisions will be made to strengthen the paper.

read point-by-point responses

Referee: The foundational claim that AGI is 'problematic both conceptually and definitionally' (Abstract) is asserted without engagement with specific technical definitions or debates in the AGI literature (e.g., those centred on generality, human-level performance, or benchmark suites). Because this claim conditions the direction of all five questions and the final recommendations, the absence of such engagement weakens the load-bearing premise of the argument.

Authors: While our analysis adopts a critical software studies perspective that inherently questions the conceptual foundations of AGI as a socio-technical construct rather than a purely technical one, we recognize that the abstract presents this claim without direct reference to specific technical debates. To address this, we will expand the introduction section to briefly engage with prominent AGI definitions from the literature, such as those focusing on generality across tasks, human-level performance on benchmarks, and related discussions. This addition will clarify how our socio-technical critique builds upon and diverges from these technical framings, thereby reinforcing the premise without altering the paper's core direction. revision: yes
Referee: Question 5 and the concluding recommendations section: the socio-technical programmes are presented as plausible outcomes of the pathway analysis, yet the manuscript does not supply an explicit mapping showing how particular leverage nodes, dead ends, or trajectory differences identified in questions 1-4 directly support the proposed programmes. This missing linkage is central to the paper's claim to offer actionable conclusions.

Authors: We agree that making the linkages explicit would enhance the coherence and actionability of our conclusions. In the revised version, we will include a new subsection following the analysis of questions 1-4 that explicitly maps each element of the proposed socio-technical development programmes to specific leverage nodes, dead ends, and trajectory differences identified earlier. This will demonstrate how, for instance, certain open-weight trajectories inform recommendations for transparency-focused programmes. revision: yes

Circularity Check

0 steps flagged

No significant circularity in interpretive critical analysis

full rationale

The paper advances a conditional, perspectival argument from a critical software studies viewpoint, structured around five explicitly posed questions about AI pathways, leverage nodes, and socio-technical alternatives. It states upfront that AGI is treated as conceptually and definitionally problematic and that assumptions of inevitability are to be avoided, but these are framing choices rather than derivations. No equations, fitted parameters, predictions, or load-bearing self-citations appear; the analysis rests on historical interpretation and internal coherence of the questions themselves. The central claims do not reduce to prior inputs by construction, satisfying the default expectation of non-circularity for non-mathematical interpretive work.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central perspective rests on a small set of domain assumptions drawn from critical software studies rather than new postulates or fitted quantities.

axioms (2)

domain assumption The current situation relating to AI is not inevitable and is closely linked to prevailing social, political and economic circumstances.
Explicitly stated as the underpinning view that conditions all subsequent analysis.
domain assumption AGI is problematic both conceptually and definitionally.
Presented as the key conditioning premise that shapes the directions and conclusions of the work.

pith-pipeline@v0.9.0 · 8597 in / 1361 out tokens · 55886 ms · 2026-05-08T10:43:03.982800+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

Code Red

The consequence was a structural shift in the search advertising market and Google was forced into a panic that would force Bard too quickly to market (and ironically produce reputational damage) followed by the Gemini demonstration. The alternative pathway suggests that if Google had decided to deploy LaMDA in mid-2021, in a similar form to the ChatGPT o...

work page 2021
[2]

This event typifies what happens when an organisation whose competitive advantage is based on research rigour is forced to operate at the pace that a consumer culture demands. Google’s misstep is a symptom of the organisational mismatch between how and what it does - previously without significant competition - and what being in a competitive situation re...

work page 2023
[3]

anti-woke

the Pause Giant AI Experiments letter signed and immediately contradicted by the founding of a competing lab and the 2024 lawsuit are all pivotal in evolution of Grok (Future of Life Institute, 2023; Musk v. Altman et al., 2024a, 2024b). All the features in the original grievance then shape every subsequent decision with the positioning of the tool as “an...

work page 2024
[4]

Chris Olah was at Google Brain from 2015 to 2018 before joining OpenAI and then subsequently co-founding Anthropic

Barret Zoph, Luke Metz, Sam Schoenholz and Liam Fedus were all Google Brain researchers before joining OpenAI. Chris Olah was at Google Brain from 2015 to 2018 before joining OpenAI and then subsequently co-founding Anthropic. His mechanistic interpretability research is the foundation for Anthropic's safety methodology. What happened and was learned at G...

work page 2015
[5]

The crisis of November 2023, when Sam Altman was briefly removed as CEO for five days before being reinstated, was ostensibly about governance

Schulman's personal pathway confirms a wider pattern in which the individuals who bear the most concentrated knowledge regarding the foundational approaches to alignment and post-training do not remain linked to any single organisation over the duration of their most productive years. The crisis of November 2023, when Sam Altman was briefly removed as CEO...

work page 2023
[6]

This new company has the stated aim of building a safe superintelligence before commercialising any capabilities

Ilya Sutskever, who had voted with the board to remove Altman but then signed the letter supporting his reinstatement, left OpenAI in June 2024 to co-found Safe Superintelligence (SSI). This new company has the stated aim of building a safe superintelligence before commercialising any capabilities. Lilian Weng, who had held roles including head of applied...

work page 2024
[7]

These departures are not separate unrelated incidents. The November 2023 crisis exposed the fundamental unresolved tension in OpenAI between the rapid pace of commercialisation, supported by Altman, and a more cautious orientation toward capability and safety research. Those departing were most closely associated with the latter position. The impact was t...

work page 2023
[8]

Leike joined Anthropic, and consolidated a pattern in which OpenAI's most safety-focused researchers migrate toward Anthropic

His subsequent public statement that OpenAI's safety culture and processes had not kept pace with its product and capabilities teams was the most direct public statement regarding these internal tensions (Milmo, 2024). Leike joined Anthropic, and consolidated a pattern in which OpenAI's most safety-focused researchers migrate toward Anthropic. 6c Thinking...

work page 2024
[9]

before co-founding xAI from May 2023 to July

work page 2023
[10]

Kyle Kosic, who had been an OpenAI engineer from April 2021, co-founded xAI in May 2023, then returned to OpenAI in May

work page 2021
[11]

The xAI personnel network reveals a distinct structural constraint

The pattern here from OpenAI to xAI and returning to OpenAI reflects a common dynamic bounding from the excitement of founding a new venture against the pull of superior resources and scale at the original organisation. The xAI personnel network reveals a distinct structural constraint. xAI's founding group do not constitute a coherent intellectual progra...

work page doi:10.1017/s0033291720000835 2023
[12]

Minds and Machines , year =

Elon Musk reportedly left OpenAI's board in 2018 after Sam Altman and other cofounders rejected his plan to run the company. Business Insider. https://www.businessinsider.com/elon-musk-reportedly-tried-lead-openai-left-after-founders-objected-2023-3 Jin, H., & Dang, S. (2023, April 18). Elon Musk says he will launch rival to Microsoft-backed ChatGPT. Reut...

work page doi:10.1007/s11023-007-9079-x 2018
[13]

State Department switches to OpenAI as US agencies start phasing out Anthropic. Reuters. https://www.reuters.com/business/us-treasury-ending-all-use-anthropic-products-says-bessent-2026-03-02 Senju, A. (2012). Spontaneous theory of mind and its absence in autism spectrum disorders. The Neuroscientist, 18(2), 108–113. https://doi.org/10.1177/10738584103972...

work page doi:10.1177/1073858410397208 2026

[1] [1]

Code Red

The consequence was a structural shift in the search advertising market and Google was forced into a panic that would force Bard too quickly to market (and ironically produce reputational damage) followed by the Gemini demonstration. The alternative pathway suggests that if Google had decided to deploy LaMDA in mid-2021, in a similar form to the ChatGPT o...

work page 2021

[2] [2]

This event typifies what happens when an organisation whose competitive advantage is based on research rigour is forced to operate at the pace that a consumer culture demands. Google’s misstep is a symptom of the organisational mismatch between how and what it does - previously without significant competition - and what being in a competitive situation re...

work page 2023

[3] [3]

anti-woke

the Pause Giant AI Experiments letter signed and immediately contradicted by the founding of a competing lab and the 2024 lawsuit are all pivotal in evolution of Grok (Future of Life Institute, 2023; Musk v. Altman et al., 2024a, 2024b). All the features in the original grievance then shape every subsequent decision with the positioning of the tool as “an...

work page 2024

[4] [4]

Chris Olah was at Google Brain from 2015 to 2018 before joining OpenAI and then subsequently co-founding Anthropic

Barret Zoph, Luke Metz, Sam Schoenholz and Liam Fedus were all Google Brain researchers before joining OpenAI. Chris Olah was at Google Brain from 2015 to 2018 before joining OpenAI and then subsequently co-founding Anthropic. His mechanistic interpretability research is the foundation for Anthropic's safety methodology. What happened and was learned at G...

work page 2015

[5] [5]

The crisis of November 2023, when Sam Altman was briefly removed as CEO for five days before being reinstated, was ostensibly about governance

Schulman's personal pathway confirms a wider pattern in which the individuals who bear the most concentrated knowledge regarding the foundational approaches to alignment and post-training do not remain linked to any single organisation over the duration of their most productive years. The crisis of November 2023, when Sam Altman was briefly removed as CEO...

work page 2023

[6] [6]

This new company has the stated aim of building a safe superintelligence before commercialising any capabilities

Ilya Sutskever, who had voted with the board to remove Altman but then signed the letter supporting his reinstatement, left OpenAI in June 2024 to co-found Safe Superintelligence (SSI). This new company has the stated aim of building a safe superintelligence before commercialising any capabilities. Lilian Weng, who had held roles including head of applied...

work page 2024

[7] [7]

These departures are not separate unrelated incidents. The November 2023 crisis exposed the fundamental unresolved tension in OpenAI between the rapid pace of commercialisation, supported by Altman, and a more cautious orientation toward capability and safety research. Those departing were most closely associated with the latter position. The impact was t...

work page 2023

[8] [8]

Leike joined Anthropic, and consolidated a pattern in which OpenAI's most safety-focused researchers migrate toward Anthropic

His subsequent public statement that OpenAI's safety culture and processes had not kept pace with its product and capabilities teams was the most direct public statement regarding these internal tensions (Milmo, 2024). Leike joined Anthropic, and consolidated a pattern in which OpenAI's most safety-focused researchers migrate toward Anthropic. 6c Thinking...

work page 2024

[9] [9]

before co-founding xAI from May 2023 to July

work page 2023

[10] [10]

Kyle Kosic, who had been an OpenAI engineer from April 2021, co-founded xAI in May 2023, then returned to OpenAI in May

work page 2021

[11] [11]

The xAI personnel network reveals a distinct structural constraint

The pattern here from OpenAI to xAI and returning to OpenAI reflects a common dynamic bounding from the excitement of founding a new venture against the pull of superior resources and scale at the original organisation. The xAI personnel network reveals a distinct structural constraint. xAI's founding group do not constitute a coherent intellectual progra...

work page doi:10.1017/s0033291720000835 2023

[12] [12]

Minds and Machines , year =

Elon Musk reportedly left OpenAI's board in 2018 after Sam Altman and other cofounders rejected his plan to run the company. Business Insider. https://www.businessinsider.com/elon-musk-reportedly-tried-lead-openai-left-after-founders-objected-2023-3 Jin, H., & Dang, S. (2023, April 18). Elon Musk says he will launch rival to Microsoft-backed ChatGPT. Reut...

work page doi:10.1007/s11023-007-9079-x 2018

[13] [13]

State Department switches to OpenAI as US agencies start phasing out Anthropic. Reuters. https://www.reuters.com/business/us-treasury-ending-all-use-anthropic-products-says-bessent-2026-03-02 Senju, A. (2012). Spontaneous theory of mind and its absence in autism spectrum disorders. The Neuroscientist, 18(2), 108–113. https://doi.org/10.1177/10738584103972...

work page doi:10.1177/1073858410397208 2026