Scrapyard AI
Discarded AI models piling up faster than anyone can use them may be a resource, not just waste — one project is already putting them to work.
The Thesis
Every time a bigger AI model ships, the previous generation gets quietly abandoned. This paper argues that 'model churn' — the rapid obsolescence of capable AI systems — creates a freely available stockpile of powerful tools for researchers and projects that can't afford frontier compute. The authors call this the 'AI scrapyard.' Their proof-of-concept, a project called Nudge-x, repurposes legacy models to document the environmental damage caused by mining operations worldwide. The catch is that this is a conceptual and artistic framing, not a technical benchmarking paper — the evidence for broad utility is illustrative rather than rigorous.
Catalyst
The pace of AI model releases has accelerated sharply since 2022, with major labs deprecating models within months of launch and open-weight releases (models whose parameters are publicly shared) making abandoned systems permanently accessible. Storage costs have dropped to the point where hosting large model weights is feasible for small institutions. This combination — fast churn plus cheap storage plus open weights — is what makes a 'scrapyard' strategy newly viable.
What's New
Prior discourse around AI resource constraints focused mostly on parameter-efficient fine-tuning (adapting large models with minimal extra compute) or distillation (compressing a big model's knowledge into a smaller one). Those approaches still assume access to current, maintained models. This paper reframes the question entirely: instead of adapting to resource limits, it asks what can be done with models that have already been left behind and are freely available, requiring no licensing negotiation or API budget.
The Counter
This paper is closer to a manifesto or art-project writeup than a research contribution. Nudge-x is presented as an example but there are no benchmarks, no comparisons to alternative approaches, and no data on whether legacy models actually perform adequately for the environmental-monitoring tasks described. The 'scrapyard' metaphor is evocative, but the claim that obsolete models are a 'potent opportunity' is asserted, not demonstrated. Legacy models also carry real costs: they may reproduce outdated safety mitigations, have no ongoing maintenance, and can be harder to run reliably than the authors suggest. The AI field already has a robust open-source ecosystem — Llama, Mistral, Falcon — so framing old models as uniquely neglected resources overstates the gap. Finally, the mining-site documentation use case, while compelling as social commentary, doesn't demonstrate that scrapyard AI is better or even comparable to existing satellite-monitoring or NGO-led approaches.
Longs
- HuggingFace (private) — primary host and distribution point for open-weight legacy models
- CLNE or similar ESG-data-adjacent plays — if AI-assisted environmental monitoring gains traction
- ESTC (Elastic) — infrastructure for searching and indexing large model repositories
- Cloudflare (NET) — low-cost inference hosting that makes scrapyard models accessible
Shorts
- API-only AI vendors — if users can get 'good enough' capability from free legacy open-weight models, paid inference APIs for older model tiers lose pricing power
- AI governance consultancies — the scrapyard framing complicates clean model lifecycle and accountability narratives they sell to enterprises
Enablers (Picks & Shovels)
- Hugging Face model hub — the de facto repository where abandoned open-weight models accumulate and remain accessible
- Ollama — open-source tool for running legacy large language models locally on consumer hardware
- Internet Archive — precedent and infrastructure for preserving deprecated software and model artifacts
- EleutherAI — nonprofit that has released and maintained older open-weight models that would otherwise disappear
Private Watchlist
- Replicate — hosts versioned AI models including deprecated ones, directly enabling scrapyard-style access
- Together AI — provides open-model inference infrastructure for non-frontier models
- Civitai — community platform for sharing and reusing open image-generation model variants
Resources
The Paper
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.