From Digital Waste to Planetary Insight

by

Author: Denis Avetisyan

A new approach, dubbed ‘Scrapyard AI’, repurposes discarded artificial intelligence models to monitor the environmental consequences of resource extraction.

This review explores the potential of leveraging obsolete AI and remote sensing data for sustainable mining practices and a more ethical future for planetary computing.

The relentless pursuit of ever-more powerful AI systems ironically generates a wealth of discarded, yet capable, models. This paper introduces ‘Scrapyard AI’, a frugal methodology for investigating large AI models by repurposing these obsolete resources for novel applications. We demonstrate this approach through Project Nudge-x, which leverages legacy AI and satellite imagery to map the environmental impact of mining operations-creating a shared record of planetary intervention for both human and artificial intelligence. Could this ‘scrapyard’ approach not only reduce the environmental cost of AI development, but also foster a more ethically-grounded relationship between technology and the planet?

The Relentless Cycle of AI Evolution and Resource Demands

The landscape of Large Language Models is defined by a relentless cycle of innovation and, consequently, obsolescence. Models like GPT-5, Claude 3.5, and Gemini represent not simply incremental improvements, but paradigm shifts that rapidly diminish the practical value of their predecessors. Each new generation boasts enhanced capabilities – greater fluency, improved reasoning, and expanded knowledge – effectively rendering older models functionally outdated for many applications. This isn’t a gradual decline; the pace of development means that models considered state-of-the-art just months ago are quickly surpassed, creating a continuous churn and prompting a constant demand for the latest advancements. The implications extend beyond mere technological upgrades; it establishes a dynamic where substantial computational resources invested in previous models are often sidelined in favor of pursuing – and quickly replacing – the newest iterations.

The swift succession of increasingly powerful Large Language Models isn’t merely a technological upgrade; it echoes established patterns of resource extraction, prompting critical questions about the sustainability of current AI development. Like the rapid depletion of natural resources to fuel economic growth, the creation of each new AI generation often necessitates abandoning prior models, even those still functionally capable. This ‘AI Obsolescence’ cycle demands substantial computational power, data storage, and ultimately, electronic waste, mirroring the environmental costs associated with traditional resource exploitation. The pursuit of ever-more-capable AI, while driving innovation, risks establishing a paradigm of disposability – one where value is placed primarily on novelty rather than on optimizing and repurposing existing infrastructure, ultimately raising concerns about long-term ecological and economic viability.

The drive to create ever-more-powerful artificial intelligence frequently eclipses the considerable utility remaining in existing models. While the unveiling of each new iteration – such as advancements beyond GPT-4 – captures attention and investment, previously capable systems are often relegated to obsolescence, despite their potential for continued service. This prioritization of novelty overlooks significant opportunities for adaptation and repurposing; models initially designed for one task could, with refinement, address new challenges or serve niche applications, offering a more sustainable and cost-effective approach than constant redevelopment. The consequence is a cycle where valuable computational resources and embodied knowledge are discarded in favor of the latest advancements, hindering a more circular and efficient paradigm within the field of artificial intelligence.

The swift advancement of artificial intelligence is establishing a pattern of rapid obsolescence, where newly developed models quickly overshadow their predecessors, creating a significant challenge for sustainability. This ‘throwaway’ culture in AI isn’t merely an economic concern regarding computational resources and energy consumption; it also represents a loss of embedded knowledge and refined capabilities within these discarded models. However, this disposability also presents an opportunity to rethink AI development, potentially shifting the focus from constant novelty towards refinement, repurposing, and efficient scaling of existing architectures. Investigating methods for model adaptation, knowledge distillation, and transfer learning could not only reduce the environmental impact of AI but also unlock a more sustainable and economically viable path forward, moving beyond a cycle of perpetual replacement.

Repurposing for Resilience: The Principles of Digital Sovereignty

The ‘Scrapyard AI’ concept advocates for the continued utilization of AI models after they have been superseded by newer iterations or are no longer actively maintained by their original developers. This differs from the conventional practice of discarding these models, which are often considered obsolete due to performance limitations or evolving requirements. Repurposing involves adapting these existing models for new, potentially less demanding tasks, or fine-tuning them with smaller datasets for specialized applications. This approach aims to extract continued value from previously trained AI assets, offering a sustainable alternative to the resource-intensive cycle of developing and deploying entirely new models for every application.

The adoption of ‘Scrapyard AI’ directly supports principles of Digital Sovereignty by enabling localized control over AI infrastructure. Traditional AI development often necessitates reliance on large, external providers for models, data, and computational resources, creating dependencies that can impact data privacy and operational autonomy. Repurposing existing, deprecated models shifts this paradigm, allowing organizations and nations to maintain and adapt AI systems independently. This localized control minimizes exposure to external vendor lock-in, enhances data security through on-premise or regionally hosted solutions, and facilitates customization to specific needs without relying on external development cycles. Ultimately, this approach strengthens technological independence and reduces vulnerabilities associated with concentrated control of AI technologies.

Extending the operational lifespan of existing AI models offers significant environmental benefits by reducing the demand for frequent retraining and associated hardware upgrades. AI model retraining is computationally intensive, requiring substantial energy consumption and contributing to carbon emissions. Furthermore, the development and deployment of new models necessitate the manufacturing of specialized hardware – including GPUs and TPUs – which carries its own environmental footprint in terms of resource extraction, energy use, and electronic waste. By maximizing the utility of currently deployed models through techniques such as fine-tuning and knowledge distillation, the rate of hardware obsolescence and the need for new model development can be slowed, thereby lessening the overall environmental impact of AI technologies.

Current AI development frequently relies on extensive data extraction – termed ‘Extractivism’ – to train new models, creating a cycle of resource consumption and potential dependency. The ‘Scrapyard AI’ approach directly challenges this by prioritizing the reuse of existing, deprecated AI models. This reduces the need for continuous large-scale data harvesting, lessening the environmental impact associated with data storage and processing. By extending the operational lifespan of current models through repurposing, the demand for new data acquisition is diminished, promoting a more sustainable and resource-conscious paradigm for AI development and deployment, and lessening reliance on external data sources.

Nudge-x: Demonstrating Sustainable Intelligence Through Planetary Observation

The Nudge-x project showcases the capabilities of Scrapyard AI through the application of repurposed Large Language Models (LLMs) to the analysis of data acquired from Sentinel-2 satellite imagery. This approach deviates from training new models from scratch; instead, existing LLMs are adapted for a specific task – in this case, the interpretation of visual data. Sentinel-2 provides multi-spectral imagery with a high revisit frequency, offering a substantial data source for monitoring Earth’s surface. By applying repurposed LLMs to this data, Nudge-x demonstrates a resource-efficient method for extracting actionable insights from planetary-scale observations, bypassing the substantial computational and energetic costs typically associated with training state-of-the-art models.

The Nudge-x project employs Retrieval-Augmented Generation (RAG) to analyze Sentinel-2 satellite imagery and quantify the impact of mining activities on planetary resources. RAG functions by first retrieving relevant data segments from a knowledge base-in this case, information pertaining to geological features, mining permits, and environmental regulations-based on features detected within the satellite images. This retrieved information is then combined with the image data and provided as context to the underlying Large Language Model (LLM), enabling it to generate more accurate and informed assessments of mining impacts, such as deforestation rates, land degradation, and water resource usage, than would be possible with the LLM operating solely on image data.

The Nudge-x project utilizes an Embedding Model to transform high-dimensional satellite imagery data into vector representations, capturing semantic meaning and facilitating efficient comparisons. These vectors are then stored and indexed within a Vector Database, enabling rapid retrieval of similar image segments based on their embedded characteristics. This approach bypasses the need for exhaustive searches through raw image data, significantly reducing processing time and computational resources required to identify and assess specific features or changes related to mining operations within the Sentinel-2 satellite imagery. The combination of embedding and vector database technologies is critical for scaling the analysis to planetary-scale datasets.

The Nudge-x project utilizes Llama 4 as its core Large Language Model (LLM), a strategic decision driven by its significantly lower energy consumption compared to more recent models like GPT-5 Pro. Independent evaluations demonstrate Llama 4 requires 4 to 5 times less energy to operate while still providing sufficient performance for analyzing satellite data. This efficiency underscores the viability of repurposing existing, pre-trained LLMs-rather than solely relying on the latest generation-as a practical and sustainable approach for deploying AI at scale, particularly in resource-intensive applications such as planetary-scale impact assessment.

Beyond Llama 4: Towards a Future of Multimodal Planetary Understanding

The demonstrable achievements of Llama 4 within the Nudge-x project underscore a pivotal advancement in environmental observation: the power of Multimodal Large Language Models (MLLMs). This success isn’t merely about processing data; it’s about interpreting complex environmental signals derived from various sources. Llama 4’s ability to synthesize information beyond simple text – incorporating elements like sensor readings and preliminary image analysis – reveals a capacity for nuanced understanding previously unattainable. The model effectively bridges the gap between raw data collection and actionable insights, offering a glimpse into a future where environmental monitoring isn’t just comprehensive, but truly intelligent. This highlights the potential for MLLMs to become indispensable tools in addressing pressing ecological challenges, offering a scalable and adaptable approach to planetary-scale observation.

The evolution of Large Language Models (LLMs) is demonstrably shifting towards multimodal capabilities, with models like DeepSeek-Chat leading the charge. Unlike their predecessors focused solely on textual data, these next-generation LLMs are engineered to process and integrate information from a variety of sources – including images, audio, and sensor data. This broadened input capacity allows DeepSeek-Chat, and similar models, to move beyond simple text generation and engage in more complex reasoning tasks. By simultaneously analyzing diverse data streams, the model can identify patterns and correlations previously obscured, potentially revolutionizing fields reliant on comprehensive data analysis, such as environmental monitoring and planetary science. This ability to synthesize information across modalities represents a significant leap forward in artificial intelligence, enabling a more holistic and nuanced understanding of complex systems.

The convergence of satellite imagery, textual data, and diverse information streams is revolutionizing the study of planetary-scale phenomena. Previously isolated datasets-remote sensing observations detailing land cover change, textual reports documenting local ecological conditions, and sensor readings measuring atmospheric composition-are now being synergistically analyzed. This integration allows for a more holistic understanding of complex environmental processes, revealing patterns and relationships that were previously obscured. Researchers can now model deforestation rates with greater accuracy by combining satellite-derived forest loss data with on-the-ground reports of illegal logging activities. Similarly, predicting the spread of wildfires benefits from the fusion of thermal imagery, weather forecasts, and historical fire incident data. This capability extends beyond environmental monitoring, offering potential for improved disaster response, agricultural optimization, and a deeper comprehension of Earth’s dynamic systems.

Planetary computing is undergoing a significant transformation through the strategic implementation of large language models. This approach moves beyond traditional data analysis by enabling the synthesis of information from varied sources – satellite imagery, sensor data, and textual reports – to create a holistic understanding of Earth’s systems. The resulting insights are not simply produced, but rigorously vetted; generated text undergoes evaluation across five key dimensions – relevance, coherence, accuracy, fluency, and conciseness – ensuring a minimum quality threshold is consistently met. This focus on verifiable outputs facilitates more sustainable and impactful environmental monitoring, offering a pathway toward data-driven decisions that address complex planetary challenges with increased confidence and precision.

The concept of Scrapyard AI, repurposing existing models instead of relentlessly pursuing novelty, echoes a fundamental principle of efficient systems. This approach aligns with Marvin Minsky’s observation: “The more we understand about intelligence, the more we realize how much of it is simply a matter of arranging things.” Scrapyard AI doesn’t seek to create intelligence anew, but to skillfully arrange existing components – discarded models and readily available remote sensing data – to address a critical planetary challenge. The system’s strength resides not in computational power, but in the elegant orchestration of readily available resources, creating a scalable solution that prioritizes sustainability and responsible AI futures. This focus on arrangement, rather than sheer processing capacity, exemplifies the core tenet of scalable systems: clear ideas, not server power.

Salvage and Synthesis

The proposition of ‘Scrapyard AI’ reveals a fundamental tension within the field: the relentless pursuit of novelty often overshadows the utility of what is already known. The current trajectory favors building larger, more complex models, discarding perfectly functional, if momentarily unfashionable, systems. This is not simply a technical issue, but a reflection of a larger cultural tendency to equate ‘new’ with ‘better’, even when evidence suggests otherwise. Future work must rigorously examine the performance characteristics of these discarded models – what precisely is lost when a model is deemed obsolete, and what latent capabilities remain?

The application to remote sensing of mining impacts offers a compelling case study, but also highlights a critical limitation. Interpretation of environmental damage, even with augmented generation, relies on pre-existing datasets and human-defined metrics of ‘impact’. A truly sustainable AI, however, demands a shift towards systems capable of independently assessing ecological health, not simply quantifying deviations from a human baseline. This necessitates a move beyond pattern recognition towards genuine environmental understanding – a task far exceeding current capabilities.

The promise of planetary computing, as demonstrated here, is not simply about scaling up existing techniques, but about fundamentally re-evaluating the relationship between computation and the natural world. It is a frugal approach, yes, but frugality must not be mistaken for sufficiency. The real challenge lies in crafting systems that are not only intelligent, but also – and perhaps more importantly – wise.

Original article: https://arxiv.org/pdf/2604.08803.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-13 23:32