The video discusses a recent privacy breach involving Grok interactions and highlights the importance of protecting user data when using AI services, explaining that mainstream platforms may retain or access conversations. It then presents three privacy-focused alternatives: running AI models locally, using O Llama’s Turbo mode which claims not to log data, and opting for privacy-centric services like Proton’s Lumo that do not record or use chats for training.
The video begins by addressing a recent privacy incident involving XAI, where over 370,000 Grok interactions were accidentally published online and made searchable. This raises significant concerns about the privacy of conversations with large language models (LLMs). The presenter emphasizes that if users want to keep their queries private, they need to consider alternatives to mainstream AI services that might store or expose their data. Privacy issues are common across many online platforms, often due to accidental leaks or malicious intent, and once data is out, it cannot be fully retracted.
The video then explains how popular AI services like ChatGPT handle user data. While ChatGPT offers a temporary chat mode that prevents conversations from appearing in chat history or being used to train models, it still retains chat data for up to 30 days for safety and review purposes. This means that even in privacy modes, there is a possibility that humans might access the conversations, which may not be acceptable for users with strict privacy needs.
To ensure complete privacy, the first recommended method is running AI models locally on one’s own computer using software like OMA (O Llama). Running models locally means no data is sent to external servers, eliminating the risk of exposure. Users can select models that fit their hardware capabilities and have full control over their chat history, including deleting conversations permanently from their device. This approach offers the highest level of privacy but requires sufficient computing resources.
For users who want to access larger models beyond their local hardware limits, the video introduces O Llama’s Turbo mode. This mode sends queries to O Llama’s servers but claims not to log or retain any user data, ensuring privacy while providing access to more powerful AI models. Although currently free during preview, Turbo mode is expected to become a paid service. This option balances privacy with performance, relying on the provider’s commitment not to store or misuse user data.
Finally, the video highlights privacy-focused AI services like Lumo by Proton, which explicitly do not record conversations or use them for training, ensuring that even the service provider cannot access chat histories. This service is designed with privacy as a core principle, appealing to users who want cloud-based AI without compromising confidentiality. The presenter invites viewers to share their thoughts on privacy concerns with AI and encourages them to subscribe for more content on this topic.