Video summary

How to use ANY AI privately - The most private LLM

Main summary

Key takeaways

Technology

Summary (tech + privacy focus)

The video argues that “running an LLM locally” is not enough for privacy in many cases—due to hardware limits, model quality limits, and the need for multiple models for different tasks. Instead, it proposes a “private AI” approach that combines:

  • De-identification for cloud AI, and
  • Local/offline AI options for higher-sensitivity use cases.

1) Why cloud AI is presented as a privacy risk

Major AI providers (e.g., OpenAI, Microsoft, Google) are described as having weaker safeguards, meaning user inputs may be:

  • Collected, stored, and retained for long periods
  • Linked to identifiers such as account details and IP address
  • Reviewed/annotated by human reviewers, increasing the chance confidential content is exposed to others
  • Shared with third parties without clear notice

The video also claims that research shows training data could potentially be extracted from language models via malicious attacks, so prompts sent to cloud models might be recovered by attackers. It further claims larger models may be more vulnerable.

2) A “choose your privacy technique” framework

Privacy concerns are split into two axes:

  • What data is collected (e.g., prompts/conversations)
  • Who it is collected from (identity/IP/account linkage)

The video suggests mitigation strategies vary by situation:

  • Sometimes de-identifying identity is enough even if chats are collected.
  • In other cases, you may need to prevent providers from learning the prompts (or at least prevent retention).
  • The most extreme option is preventing collection/identification entirely.

3) De-identifying cloud AI usage (example: ChatGPT)

For services like ChatGPT, the video focuses on de-identification, including:

  • Use a full-device VPN
    • Recommends ProtonVPN free to start (no account required)
    • Warns against “popular YouTuber sponsored VPNs”
  • Use a VPN + a fake email alias to create a pseudonymous account
    • Uses the concept of an email alias to separate identity from prompts
    • Recommends generating credentials via a password manager with unique passwords

Even with pseudonymity, it notes prompts may still be collected as a “permanent record” from the provider’s perspective, so this is not full privacy.

Key operational steps (as described)

  1. Run a full device VPN
  2. Create an account using a fake alias email
  3. Use a password manager to create a strong unique password
  4. Use the cloud AI service with reduced identity linkage

4) Cloud services presented as better for not retaining prompts

The video lists alternatives intended to reduce long-term retention or linking:

1. Venice AI

  • Described as “private” and “permissionless,” with or without an account
  • Claims to route prompts through a proxy
  • Claims it does not store/retain conversations
  • Suggests VPN + email alias reduces metadata leakage

2. Hugging Face: “Hugging Chat”

  • Uses multiple open-source models
  • Claims Hugging Face does not share conversations and offers deletion options

3. Brave Leo (Brave browser AI)

  • Works only inside the Brave browser
  • Claims records aren’t tied to IP/identifiers (though the video still recommends using a VPN “just to be safe”)
  • No account needed unless using advanced features

5) When de-identification still isn’t enough: local/offline AI

For highly sensitive tasks (e.g., private brainstorming, mental health, private document analysis, or creating an AI “character”), the video recommends:

  • Running LLMs locally with no internet dependency

It mentions tools such as:

Open WebUI + Ollama (AMA mentioned)

Example setup includes:

  • Install Ollama/AMA
  • Download models using command prompts
  • Install Docker
  • Deploy Open WebUI in a Docker container

The video claims Open WebUI can:

  • Host a local web interface
  • Customize models
  • Let users upload/feed documents for local analysis

Jan

  • Presented as a lesser known local tool
  • Claims it works with Hugging Face model downloads
  • States Jan can connect to cloud models, but recommends staying local for privacy

Model selection guidance

  • If hardware is limited, the video suggests using small models that can run “relatively smoothly.”

6) Maximum isolation approach: GrapheneOS

For users who may be adversarial targets, the video recommends stronger device-level compartmentalization:

  • GrapheneOS
  • Use a separate isolated user profile for untrusted apps
  • Run a full-device VPN inside that profile
  • Install a separate isolated anonymous Google account
    • No phone number
    • Not tied to real identity
  • Use gift cards for Google Play payment instead of PayPal/debit cards

It specifically mentions using this setup for a “non-trusted” app like ChatGPT when higher isolation is desired.

7) Meta / tutorial structure

The video advertises “secret methods” and claims to provide guidance on:

  • When to switch between local vs. cloud
  • What to do vs. avoid
  • Links to research papers and tool installation commands (implied via description)

Main speakers / sources

Primary speaker

  • The video creator (unnamed in the subtitles)

Cited/mentioned collaborators & sources

  • Naomi Brockwell (referenced for a local AI tutorial series on her channel)
  • ProtonVPN (tool provider)
  • ChatGPT / OpenAI (target cloud service)
  • Venice AI (privacy-focused provider)
  • Hugging Face (Hugging Chat)
  • Brave (Brave Leo)
  • Open WebUI / Ollama (AMA) (local tools)
  • Jan (local tool)
  • GrapheneOS (device OS)
  • Research papers (mentioned, but specific titles not provided in subtitles)

Original video