Open weights are quietly closing up - and that's a problem (martinalderson.com)

The landscape of open weights models in the large language model (LLM) space is changing, with notable companies tightening their licensing conditions and fewer models being released openly. Historically, open weights models like those from the Llama series allowed users to run AI on local hardware, providing benefits in privacy, flexibility for fine-tuning, and significant cost savings compared to proprietary models. However, recent trends towards restrictive licenses, as seen with Meta's Muse Spark and Alibaba's API-first releases, threaten the availability of these models. This shift could reduce competitive pressure on major players like OpenAI and Google, potentially enabling them to raise prices without concern for market alternatives. This situation has critical implications for the AI/ML community, as an oligopolistic market structure could mean diminished price competition, leading to increased costs for users. The erosion of a competitive open weights ecosystem risks a concentration of power and wealth among a few firms, limiting access and affordability of AI technologies. While advancements in hardware and model training techniques may mitigate some of these effects, the trend signals a need for vigilance. The ability to access open weights models has been a foundational assumption in the AI economy, and its decline could have far-reaching consequences for innovation and consumer value.