Just say no to OpenAI and Anthropic

You can now use powerful LLMs privately without paying OpenAI or Anthropic.

Micah Lee argues that users should avoid proprietary large language models (LLMs) like those from OpenAI and Anthropic, advocating instead for open-weight models. Lee criticizes OpenAI and Anthropic for their perceived 'Machine God' complex among leadership and their closed, privacy-invasive systems. He highlights that many open-weight models, such as GLM-5.3 from Z.ai and models from Mistral AI, are nearly as capable as top proprietary models, including for complex tasks like vulnerability research. A key advantage of open-weight models is the ability to run them privately on personal hardware or through providers like Tinfoil that use secure enclaves, thus avoiding data exposure to the model creators. Lee also notes that while Z.ai models are impressive, users should avoid their direct services due to past data handling issues. For users prioritizing privacy without self-hosting, Tinfoil and Maple AI offer solutions using secure enclaves, while services like OpenCode Go provide cheaper, albeit less private, access to various open-weight models.

The rise of capable open-weight LLMs, nearly matching proprietary models in performance, presents a significant challenge to dominant players like OpenAI and Anthropic. This shift democratizes access to advanced AI capabilities, allowing individuals and smaller organizations to deploy powerful tools without the privacy concerns or vendor lock-in associated with closed systems. The availability of private inference solutions further reduces reliance on centralized providers, signaling a potential fragmentation of the LLM market towards more user-controlled deployments.

This trend affects users by offering viable alternatives that prioritize privacy and reduce costs, particularly for cybersecurity professionals and researchers who face stringent guardrails on proprietary models. Companies that previously relied on OpenAI or Anthropic may re-evaluate their strategies, exploring open-weight models for greater control and transparency. Future developments to watch include the continued performance gap between open and closed models, the evolution of secure inference technologies, and the regulatory landscape surrounding AI data privacy.