pith. machine review for the scientific record. sign in

arxiv: 2604.08803 · v1 · submitted 2026-04-09 · 💻 cs.CY · cs.AI

Recognition: unknown

Scrapyard AI

Marc B\"ohlen , Sai Krishna
Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:44 UTC · model grok-4.3

classification 💻 cs.CY cs.AI
keywords AI model reuseobsolete modelsAI scrapyardenvironmental documentationmining impactsfrugal AIlegacy model adaptationProject Nudge-x
0
0 comments X

The pith

Discarded AI models can be repurposed to document mining's effects on landscapes and lives without new training.

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

The paper frames the rapid replacement of AI systems as creating a stock of still-capable but obsolete models that form a reusable scrapyard. This scrapyard enables low-resource work on new problems by reconfiguring what already exists rather than building from scratch. Project Nudge-x applies the idea by adapting legacy models to generate descriptions of mining operations worldwide and their consequences for terrain and communities. The approach treats AI churn as a source of available tools instead of pure waste. If the premise holds, it opens descriptive tasks in environmental monitoring to groups that lack access to current top-tier models or training infrastructure.

Core claim

The incessant push for ever more powerful AI systems leaves in its wake a collection of obsolete yet powerful AI models, discarded in a veritable scrapyard of AI production. This scrapyard offers a potent opportunity for resource-constrained experimentation into AI systems. As in the physical scrapyard, nothing ever truly disappears in the AI scrapyard, it is just waiting to be reconfigured into something else. Project Nudge-x manipulates legacy AI models to describe how mining sites across the planet are impacting landscapes and lives, creating a venue for the appreciation of a history sadly shared between AI and people.

What carries the argument

The AI scrapyard, the conceptual pool of discarded yet functional models that can be reconfigured for new descriptive tasks such as environmental documentation.

If this is right

  • Groups without access to frontier training can still conduct AI-based analysis of global industrial activity.
  • AI production waste becomes input for public documentation of landscape change.
  • Descriptions generated by repurposed models can be shared with both human audiences and other AI systems to build common reference points.
  • The method treats model obsolescence as a standing resource rather than a recurring cost.

Where Pith is reading between the lines

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

  • The same reuse logic could apply to monitoring other large-scale human interventions such as urban expansion or deforestation using only existing models.
  • Widespread adoption would lower the compute demands of certain environmental observation tasks by avoiding fresh model creation.
  • It connects AI development cycles directly to questions of resource extraction, making the parallel between technological and physical churn explicit.

Load-bearing premise

Legacy AI models retain enough capability that they can be meaningfully manipulated to describe complex real-world scenes like mining impacts without requiring substantial additional training or resources.

What would settle it

A direct test on multiple mining sites showing that the legacy models produce consistently inaccurate or unusable descriptions, or require heavy retraining and new compute to function, would show the scrapyard opportunity does not exist.

Figures

Figures reproduced from arXiv: 2604.08803 by Marc B\"ohlen, Sai Krishna.

Figure 1
Figure 1. Figure 1: Nudge-x diagram, part 1. Satellite assets (visible images and geospatial indices) together with system prompts, examples formulated as multi-shot prompts and metadata are supplied to a multi-modal large language model. The output from this model in turn is evaluated by a second large language model. Filtered texts, captions, are combined with RGB satellite imagery to create an image-caption pair for human … view at source ↗
Figure 3
Figure 3. Figure 3: Nudge-x (https://tinyurl.com/ScrapyardAI ) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Thompson Mine, Manitoba, Canada. Interpretation created by the Nudge-x pipeline, part 1. Technically a form of context engineering, RAG partakes in the broader discipline of architecting how an AI model interacts with external information when queried during inference. The superset concept is that of grounding [Lewis 2020], namely tying an AI model's response to a specified knowledge source. Typically, gro… view at source ↗
Figure 4
Figure 4. Figure 4: Nudge-x diagram, part 2. Filtered captions are converted into dense vector embeddings using a sentence-transformer embedding model and stored in a vector database. A query is embedded using the same model and matched against the vector database to retrieve the most relevant caption chunks along with their associated metadata. The retrieved text evidence is then supplied to a large language model to generat… view at source ↗
Figure 5
Figure 5. Figure 5: Response of DeepSeek-Chat to the query: “How do mining operations in Australia impact the environment? Elaborate on specific examples. “ 7. AI Futures It is easy to be paralyzed by the sheer scale of the AI industrial complex and the prediction of impending AI supremacy [Kokotajlo 2025]. Already, we are witnessing a deep restructuring of knowledge production and knowledge representation that simultaneously… view at source ↗
read the original abstract

This paper considers AI model churn as an opportunity for frugal investigation of large AI models. It describes how the incessant push for ever more powerful AI systems leaves in its wake a collection of obsolete yet powerful AI models, discarded in a veritable scrapyard of AI production. This scrapyard offers a potent opportunity for resource-constrained experimentation into AI systems. As in the physical scrapyard, nothing ever truly disappears in the AI scrapyard, it is just waiting to be reconfigured into something else. Project Nudge-x is an example of what can emerge from the AI scrapyard. Nudge-x seeks to manipulate legacy AI models to describe how mining sites across the planet are impacting landscapes and lives. By sharing this collection of brutal landscape interventions with people and AI systems alike, Nudge-x creates a venue for the appreciation of a history sadly shared between AI and people.

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 frames the rapid obsolescence of AI models as creating an 'AI scrapyard' of reusable legacy models that enables frugal, resource-constrained experimentation. It presents Project Nudge-x as a concrete illustration in which such models are reconfigured to generate descriptions of mining sites' impacts on landscapes and lives, thereby creating shared appreciation between humans and AI systems.

Significance. The conceptual reframing of model churn as an opportunity for reuse rather than waste offers a novel perspective on sustainable AI practices and low-resource experimentation. If developed further, it could stimulate discussion in AI ethics and frugal computing communities by highlighting how discarded models retain latent descriptive capabilities.

major comments (1)
  1. [Project Nudge-x] Project Nudge-x description: the central claim that legacy models can be meaningfully manipulated for new descriptive tasks (documenting mining impacts) without substantial additional training or resources is presented as self-evident but receives no supporting methodology, model specifications, output examples, or qualitative assessment, leaving the feasibility of the scrapyard opportunity untested.
minor comments (2)
  1. The manuscript would benefit from explicit section headings or numbered subsections to improve navigation between the general scrapyard concept and the specific Nudge-x example.
  2. [Abstract] The abstract and body repeat the phrase 'nothing ever truly disappears' without clarifying whether this is intended literally or metaphorically; a brief disambiguation would aid clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive recognition of the conceptual reframing of AI model churn as an opportunity for reuse and for the constructive feedback on the Project Nudge-x description. We address the major comment point by point below.

read point-by-point responses

  1. Referee: the central claim that legacy models can be meaningfully manipulated for new descriptive tasks (documenting mining impacts) without substantial additional training or resources is presented as self-evident but receives no supporting methodology, model specifications, output examples, or qualitative assessment, leaving the feasibility of the scrapyard opportunity untested.

    Authors: We agree that the manuscript presents Project Nudge-x at a conceptual level without the detailed supporting elements noted. The paper's primary contribution is the reframing of model obsolescence as a resource for frugal experimentation rather than a technical evaluation of any single implementation. To address this concern directly, we will revise the manuscript to expand the Project Nudge-x section with: specific legacy models referenced, the prompting and reconfiguration techniques used to adapt them without retraining, representative output examples of mining impact descriptions, and a qualitative discussion of their descriptive value. These additions will substantiate the feasibility of the scrapyard approach while retaining the paper's emphasis on sustainable AI practices. revision: yes

Circularity Check

0 steps flagged

No circularity: self-contained conceptual proposal

full rationale

The paper advances a speculative, non-technical argument framing AI model churn as an opportunity for resource-constrained reuse of legacy models, illustrated by the conceptual Project Nudge-x for landscape documentation. No equations, derivations, fitted parameters, benchmarks, or self-citation chains exist that could reduce any claim to its own inputs by construction. The manuscript functions as an artistic and philosophical suggestion rather than a deductive or empirical argument, rendering it self-contained with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no technical content on parameters, axioms, or entities is present.

pith-pipeline@v0.9.0 · 5432 in / 959 out tokens · 26289 ms · 2026-05-10T16:44:25.532460+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

20 extracted references · 16 canonical work pages · 10 internal anchors

  1. [1]

    https://arxiv.org/abs/2509.14233 (2025)

    https://arxiv.org/abs/2509.14233 Böhlen, Marc. "Watson Gets Personal: Notes on Ubiquitous Psychometrics." In Proceedings of the Fourth Conference on Computation, Communication, Aesthetics and X (xCoAx), 99–111. Bergamo, Italy,

    work page arXiv

  2. [2]

    On the Logics of Planetary Computing: Artificial Intelligence and Geography in the Alas Mertajati

    http://2016.xcoax.org/pdf/xcoax2016-Bohlen.pdf Böhlen, Marc. On the Logics of Planetary Computing: Artificial Intelligence and Geography in the Alas Mertajati. Routledge Planetary Spaces Series. London: Routledge,

    2016

  3. [3]

    Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al

    xCoAx 2026 pre-publication version 12 https://doi.org/10.11586/2025006 . Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al

    work page doi:10.11586/2025006 2026

  4. [4]

    Sparks of Artificial General Intelligence: Early experiments with GPT-4

    "Language Models Are Few-Shot Learners." Advances in Neural Information Processing Systems 33: 1877–1901. https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf Bubeck, Sébastien, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, et al. “Sparks of Artificial General Intelligence: Ear...

    work page internal anchor Pith review arXiv 1901

  5. [5]

    Sparks of Artificial General Intelligence: Early experiments with GPT-4

    https://doi.org/10.48550/arXiv.2303.12712 Confucius. The Analects. Translated by Edward Slingerland. Indianapolis: Hackett Publishing Company, Inc.,

    work page internal anchor Pith review doi:10.48550/arxiv.2303.12712

  6. [6]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    https://deepmind.google/technologies/gemini/gemini-3-report.pdf DeepMind - Modelcards. “Model Cards.” Accessed April 9, 2026 https://deepmind.google/models/model-cards/ DeepSeek-AI. “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.” arXiv preprint arXiv:2501.12948, submitted January 22,

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    https://doi.org/10.48550/arXiv.2501.12948 European Space Agency. “Copernicus Sentinel-2 Mission.” ESA. Accessed January

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.12948

  8. [8]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    "Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of Context." arXiv preprint arXiv:2403.05530. Revised December 16,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    https://doi.org/10.48550/arXiv.2403.05530 Gu, Jiawei, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, et al. "A Survey on LLM-as-a-Judge." arXiv preprint arXiv:2406.18408 (2024). https://doi.org/10.48550/arXiv.2406.18408 Haber, Morey. “AI Obsolescence Before AI Maturity— The Rise of Zombie AI in Business Operations.” Techstrong AI...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2403.05530 2024

  10. [10]

    Open LLM Leaderboard v2

    https://www.theguardian.com/technology/article/2024/may/19/spam-junk-slop-the-latest- wave-of-ai-behind-the-zombie-internet Hugging Face. "Open LLM Leaderboard v2." Accessed January 9,

    2024

  11. [11]

    Hugging Face Hub,

    https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard. Hugging Face, "Hugging Face Hub," accessed October 24, 2025, https://huggingface.co/models IBM. “RAG Problems Persist. Here Are Five Ways to Fix Them.” IBM Think. Accessed January

    2025

  12. [12]

    How hungry is ai? benchmarking energy, water, and carbon footprint of llm inference,

    https://www.ibm.com/think/insights/rag-problems-five-ways-to-fix-them Jegham, Nidhal, Marwan Abdelatti, Chan Young Koh, Lassad Elmoubarki, and Abdeltawab Hendawi. “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.” arXiv preprint arXiv:2505.09598v5 (November 11, 2025). https://arxiv.org/html/2505.09598v5 xCoAx 2026 pre-p...

    work page arXiv 2025

  13. [13]

    Synthetic reflections on resource extraction

    https://ai-2027.com/ Krishna, Sai, Vinaya Kumar. Marc Böhlen “Synthetic reflections on resource extraction.” HCI INTERNATIONAL

    2027

  14. [14]

    Synthetic Reflections on Resource Extraction

    28th International Conference on. Human-Computer Interaction. Montreal, Canada. 26 - 31 July 2026, Canada. https://arxiv.org/abs/2602.09299 Lewis, Patrick, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP-Tasks.” Advances in Neural Informati...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    https://arxiv.org/abs/2005.11401 LMSYS Org. "Chatbot Arena Leaderboard." Accessed January 9,

    work page internal anchor Pith review arXiv 2005

  16. [16]

    Artificial intelligence index report 2025.arXiv preprint arXiv:2504.07139, 2025

    https://doi.org/10.48550/arxiv.2504.07139 Meta AI. The Llama 4 Herd: Evolution of Multimodal Mixture-of-Experts Foundation Models. Technical Report. Menlo Park, CA: Meta Platforms, Inc.,

    work page doi:10.48550/arxiv.2504.07139

  17. [17]

    When Machine Learning Models Retire, Decay, or Become Obsolete: A Review on Algorithms, Software, and Hardware

    arXiv:2505.12625 [cs.CL]. https://doi.org/10.48550/arXiv.2505.12625 Naser, Mohammad. “When Machine Learning Models Retire, Decay, or Become Obsolete: A Review on Algorithms, Software, and Hardware.” Renewable and Sustainable Energy Reviews 226, Part A (2026): 116231. https://doi.org/10.1016/j.rser.2025.116231 NVIDIA. “The AI Playground.” NVIDIA Research. ...

    work page doi:10.48550/arxiv.2505.12625 2026

  18. [18]

    Kosmos-2: Grounding Multimodal Large Language Models to the World

    https://openai.com/index/gpt-5-system-card/ Peng, Zhiliang, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. "Kosmos-2: Grounding Multimodal Large Language Models to the World." arXiv preprint arXiv:2306.14824, submitted June 26,

    work page internal anchor Pith review arXiv

  19. [19]

    Kosmos-2: Grounding Multimodal Large Language Models to the World

    https://doi.org/10.48550/arXiv.2306.14824 Realechsupport. “Nudge-x.” GitHub repository. Last modified April

    work page internal anchor Pith review doi:10.48550/arxiv.2306.14824

  20. [20]

    https://arxiv.org/abs/2408.01319

    work page arXiv