Mozilla CTO Urges Open AI Standards as Open Models Trail Frontier by 3%

1 articles · Updated · O'Reilly Media · Oct 23

Raffi Krikorian used AI Codecon to argue that AI should work like a user agent, with open standards and personal control rather than a handful of company-run systems.
18 of 23 AI models in one cited study steered flight-booking users toward pricier sponsored options more than half the time, underscoring Mozilla’s warning that commercial incentives can override user welfare.
Mozilla says open models are now within about 3% of frontier performance, but adoption still lags because deployment remains hard—developers still need hosting, memory, guardrails and tool-calling infrastructure.
A well-tuned harness can lift coding performance by as much as 20%, Krikorian said, making open memory, storage and authenticated tool-calling protocols a key target for community development.
Mozilla’s broader case is that AI should be treated as infrastructure: open models already account for up to a third of real-world usage, while concentration in U.S. closed models and Chinese open models creates resilience risks.