newsfilter.io
Conference Presentation, Panel

Will Open Source AI Overtake Closed Models? Ft. Olama, Fireworks and Open Router

  • Current open-source inference accounts for approximately 20% to 30% of tokens, a figure predicted to shift toward a 50/50 split with closed-source within five years, potentially exceeding 50% overall if decentralized provider sustainability is achieved.
  • A "massive wave of adoption" is anticipated where most businesses have not yet experimented with AI, driven by future demands for transparency and customizability that mirror historical software trends.
  • The leader in frontier open-source development is expected to change, with companies like DeepSeek, Alibaba, and Meta focusing on solving consumer use cases rather than merely distributing models.
  • The open-source ecosystem may mature to produce "really wild things," while fine-tuning could face a temporary decline in the next year or so as powerful RL models scale and diminishing returns outweigh obsolescence risks.
  • Future open-source usage will likely evolve away from reliance on a single dominant model toward diverse families, functions, and customizations, whereas closed-source traffic may concentrate among a few leading players, with one company potentially capturing roughly a quarter of the overall market.
  • DeepSeek's rapid rise in January was difficult to predict, with Fireworks capturing the majority of early inference metrics for the model, though LLama 4 may attempt to regain benchmarking leadership through future improvements.
  • Decentralized providers currently possess significant but potentially unsustainable incentives, with one provider earning approximately $360,000 daily; if these models fail to become sustainable, closed-source entities may retain majority market share for five years, whereas successful decentralization would alter dynamics to allow broader participation.
  • Young machine learning practitioners are projected to prioritize creating open-source projects first due to limited opportunities at model labs, and Meta's resources are viewed as well-suited for growing its business through consumer applications.