arXiv:2604.07530v2 Announce Type: replace-cross Abstract: Scaling laws describe how language model capabilities grow with compute and data, but say nothing about how long a model matters once released. We introduce time-to-peak and lifespan as measures of model obsolescence and use them to characterize the scientific adoption trajectories of 62 LLMs across more than 108k citing papers (2019-2025), separating active adoption from background citation to recover per-model trajectories that citation counts cannot resolve. We find that a model's longevity is shaped more by when it was released than
The rapid development and proliferation of LLMs over the past few years are now mature enough for researchers to analyze their adoption and obsolescence patterns empirically.
This research provides crucial insights into the rapid decay of relevance for specific AI models, highlighting the accelerating pace of innovation and potential challenges for long-term AI strategy and investment.
The understanding that LLMs have a rapidly shrinking lifespan, driven more by release timing than scaling laws, will shift focus from raw capability to adaptability and continuous innovation.
- · AI research institutions
- · Agile AI development companies
- · AI model infrastructure providers
- · Generative AI platforms
- · Companies with deeply embedded legacy LLMs
- · Investors in static LLM intellectual property
- · AI models without continuous updates
- · Monolithic LLM vendors
Increased emphasis on transfer learning, modular architectures, and model-agnostic applications to mitigate rapid obsolescence.
Accelerated cycles of investment and disinvestment in specific LLM architectures, leading to more dynamic market behaviour.
The concept of 'digital depreciation' for AI models becomes a standard accounting and strategic metric, influencing venture capital and acquisition decisions.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI