Skip to content

Frequently asked questions

You're looking at the docs-site FAQ. The repo-root FAQ lives at /FAQ.md and is kept in sync with this page on every release. If your question isn't here, check known issues, the troubleshooting guide, or ask via SUPPORT.md.

Terms in bold-ish code are links to the glossary.

General

What is PAI?

PAI (Private AI) is a bootable live operating system that bundles local AI inference, privacy tooling (Tor, hardened browsers), and crypto wallets onto a single USB stick. It runs entirely from removable media and forgets everything when you shut down — unless you opt into encrypted persistence.

See: overview, quickstart.

How is PAI different from Tails?

Tails focuses on anonymity via Tor and leaves no trace. PAI shares that amnesic foundation but adds local LLMs (via Ollama), GPU-aware model runners, and crypto wallets preconfigured for offline signing. Think of PAI as "Tails + a local AI workstation."

See: editions, architecture.

How is PAI different from Qubes OS?

Qubes is an installed OS that isolates workloads in Xen VMs on your primary disk. PAI is a live USB that boots on almost any AMD64 or ARM64 machine without touching the internal drive. Different threat models, different workflows.

How is PAI different from Kodachi?

Kodachi is a privacy-oriented live distro with a heavier desktop and a different trust model. PAI is narrower in scope (privacy + local AI + crypto), reproducibly built, and signed with minisign.

Is PAI free?

Yes. PAI is free and open source. See LICENSE.

Who makes PAI?

A small group of maintainers plus community contributors. See MAINTAINERS.md and AUTHORS.md.

Install & boot

Will installing PAI delete my data?

No. PAI writes to the USB stick you flash, not to your internal drive. Flashing the USB will erase everything already on that specific USB — so pick an empty one or back up first.

See: installation guide.

What size USB do I need?

Minimum 16 GB for a basic session; 32 GB or larger is recommended if you plan to enable persistence and download LLM models. Fast USB 3.0+ sticks dramatically improve boot and model load times.

Why won't PAI boot on my Mac?

Apple Silicon Macs (M1/M2/M3/M4) require the ARM64 image and may need reduced security mode. Intel Macs usually work but sometimes need the Option key held at power-on and Secure Boot disabled in Startup Security.

See: troubleshooting — macOS boot.

Does PAI work with Secure Boot?

Signed shim support is planned but not guaranteed on every firmware. Disable Secure Boot in your UEFI settings if the USB is not detected.

Privacy

Does PAI send anything to anyone?

No telemetry, no "phone home," no analytics. The only outbound traffic is what you initiate (opening a browser, syncing a wallet, or pulling an Ollama model).

See: architecture — network.

Does using Tor make me invisible?

No. Tor hides your network location, not your behavior. Logging into a personal account, using identifying writing style, or pairing Tor with a non-Tor identity all undermine anonymity.

See: threat model.

What does encrypted persistence actually encrypt?

A single LUKS container on the USB stores any files you explicitly opt to persist (model weights, wallet files, GPG keys, browser profiles). Everything outside that container is re-created at each boot and cannot be recovered.

See: usage — persistence.

Can PAI protect me from a hardware keylogger or compromised firmware?

No live OS can. PAI defends the software stack. If your hardware is compromised, no operating system can save you.

AI

Which models are included?

PAI ships with a small curated set of Ollama-compatible models sized for CPU-only use. Larger models can be pulled after boot if you have disk space and/or a GPU.

See: usage — AI.

Do I need a GPU?

No. CPU-only inference works for small models (quantized to 4-bit or lower). GPUs make larger models usable — NVIDIA (CUDA) and AMD (ROCm) are supported where drivers permit.

Can I use GPT-4 or Claude from inside PAI?

Yes, via their APIs, but you'll be sending prompts to a third party and that breaks the local-only threat model. Route through Tor if you must, and understand the trade-off.

How do I add a new model?

ollama pull <model-name> inside a PAI session. If you want the model to persist across reboots, enable persistence and move the model into the persistent volume.

See: usage — AI.

Crypto

Are the wallets safe on a live USB?

Safer than on a general-purpose OS for signing, because the environment is amnesic and the network can be fully Tor-routed. They are not a substitute for a hardware wallet for large holdings.

See: usage — crypto.

How do I back up my wallet?

Enable persistence and export seed phrases to paper. Never store seed phrases unencrypted on any networked device.

Troubleshooting

My screen is black after boot.

Usually a GPU driver or Wayland handoff issue.

See: known issues, troubleshooting — black screen.

Wi-Fi doesn't work.

Likely a missing non-free firmware blob for your chipset.

See: troubleshooting — wifi.

Ollama is really slow.

Expected on CPU-only with large models. Try a smaller quantization, close other apps, or add a GPU.

My persistence volume won't unlock.

Double-check the passphrase (case-sensitive, keyboard layout matters). If still failing, see troubleshooting — persistence.

In most jurisdictions, yes. A handful of countries restrict privacy tools like Tor or strong encryption — check your local law. PAI is a tool; how you use it is your responsibility. See ETHICS.md.

Can I use PAI at work?

Check your employer's acceptable-use policy before booting a live OS on company hardware. Many policies forbid it regardless of intent.

Can I use PAI commercially?

Yes, within the license. See LICENSE and GOVERNANCE.md.

Meta FAQ

Why does this FAQ exist in two places?

Two audiences, two entry points:

  • /FAQ.md at the repo root is for people browsing the source tree on GitHub, a mirror, or a cloned checkout. It needs to stand alone without a docs site to link into.
  • This page (docs/reference/faq.md) is for people reading the rendered docs site. It enriches each answer with glossary links and deep pointers into docs/.

The two are kept semantically identical. The docs-site copy adds links; it never adds, removes, or changes answers. PR reviewers should block any change that diverges the two, and the contributing-to-docs page documents this convention for authors.

Where do I propose a new FAQ entry?

Open a PR touching both /FAQ.md and /docs/reference/faq.md. If you're only comfortable editing one, that's fine — leave a note in the PR and a maintainer will mirror the change.

How current is this page?

The updated: field in the frontmatter is the source of truth; it is bumped on every release and on material content edits.