Is Local AI as Private as You Think? - Techlore Office Hours Live

Summary

In this live Q&A, the Techlore host answers viewer questions about privacy tools, local AI, browsers, mobile operating systems, and online services. A recurring theme is that privacy depends not only on choosing a tool, but also on how it is configured and how it fits a person’s threat model.

Local AI and privacy

Running a model locally can keep prompts from being sent to an AI provider, particularly when both the model and app work fully offline. This can reduce concerns about a provider collecting or using conversation data.

However, “local” does not automatically mean private. A misconfigured server may be exposed to the internet, and a client or connected service may still transmit data. The host says they found many internet-accessible Ollama servers during research. More advanced automation may also rely on online services or MCP connections, so using a local model does not necessarily keep an entire workflow offline. Small offline models can handle basic tasks but may not offer the automation capabilities people expect.

The host recommends Jan AI as an open-source, cross-platform starting point for local models. They describe it as relatively straightforward and say it offers models in different sizes.

Browser fingerprinting and compartmentalization

The host discusses using separate LibreWolf profiles for activities such as YouTube and Twitch, with different settings and extensions in each. This can help compartmentalize browsing, but it does not guarantee that sites cannot link profiles. Shared IP addresses, hardware signals, and other persistent characteristics may still enable tracking. Separate profiles are presented as a reasonable step, not a complete defense.

Brave is mentioned for its fingerprint-randomization approach. For stronger anonymity-oriented browsing, the host points to Tor Browser and Mullvad Browser.

For work, school, or other privacy-sensitive activities, the general recommendation is to keep contexts separate. Depending on the situation, this could mean using different browsers, accounts, virtual machines, or dedicated devices.

Websites, domains, and services

Owning a domain can make an email address portable between providers, though domains generally need to be renewed periodically. The host discusses Njalla for domain registration, noting its association with Monero payments and avoiding KYC. For privacy-conscious hosting, they recommend 1984 Hosting; for simpler static websites, they suggest GitHub Pages and Codeberg Pages. The host says Techlore uses Hetzner.

The host criticizes WordPress as heavy and cumbersome for sites requiring many features, often involving third-party plugins. Ghost is presented more favorably as simpler and easier to maintain.

Techlore’s updated site, Go Incognito V2, is also discussed. The host describes changes to its Ghost theme, interactive elements, account pages, and visual Easter eggs.

Privacy quiz and data handling

The updated privacy quiz encodes answers in URL fragments rather than storing results in an account. This allows users to retain or revisit results without the site maintaining a results database.

One drawback is that anyone with the full link may be able to see the encoded information. The host advises caution when sharing quiz links and suggests adding compatibility checks if the encoding method is used in software.

The quiz focuses on a person’s specific threats and privacy needs rather than a single universal measure of how much privacy they should pursue. The host says an updated version with a score feature is expected.

Other privacy tools and services

The host is uncertain about Privacy.com. They describe its value as limiting the card details shared with merchants, rather than providing complete anonymity, and express concern that the service may now request more information or impose more limits. MySudo is mentioned as another option, though the host says it is more expensive.

For self-hosted Pi-hole or AdGuard Home, the host sees no inherent reason to avoid self-hosting, provided the server is secured and the user is prepared to handle maintenance and reliability.

The host also emphasizes that hardware cannot be made completely open source: some components, such as cellular modem chips, remain proprietary.

Mac mini and local models

The host has been experimenting with a 48 GB Mac mini and a local model, describing the experience as interesting but still early. A viewer reports using KoboldCPP for text generation on a home server without agents or internet access. The host considers this a sensible starting point.

iOS and Android

For iOS, the host recommends reducing reliance on Apple services where practical, using privacy-oriented third-party apps, and enabling available protections such as Lockdown Mode and Advanced Data Protection. Examples discussed include Brave, Proton Mail, and DNS filtering with AdGuard.

The host points viewers to Techlore’s iOS privacy and security guide, noting that it is older but that much of it may still be relevant.

On Android, downloading apps through Obtainium while continuing to use Google Play Services may reduce the connection between app installs and a Google Play account. However, Play Protect may still scan apps on the device.

The host recommends gradually reducing reliance on Google rather than waiting to switch to a fully de-Googled phone all at once. This can make a later move to a custom ROM or another system easier.

Linux phones and Qubes-like systems

The host is concerned that Google’s growing control over Android could make privacy-respecting alternatives harder to build. Projects such as postmarketOS, Ubuntu Touch, and Sailfish OS exist, but lack the ecosystem and support available to desktop Linux.

Locked bootloaders, unavailable drivers, and limited hardware support are identified as obstacles to Linux phones. The host would like to see a genuinely viable mobile alternative.

When asked about alternatives to Qubes OS, the host does not recommend a specific replacement. Instead, they discuss the appeal of Qubes’ isolation model and the need for a more approachable system with useful security protections.

Online shopping

Whether to forgo discounts for privacy is presented as a personal threat-model decision. Loyalty programs can save money, but they also give stores more information about customers.

One suggested compromise is to use separate email aliases for loyalty accounts, limiting the reuse of a primary address across services.

Guides, tools, and reviews discussed

• Jan AI: Recommended as an open-source starting point for running local models.
• Techlore iOS Privacy and Security Guide: Recommended, with the caveat that it is somewhat dated.
• Go Incognito V2 privacy quiz: Discussed as a tool for assessing privacy needs and retaining results through a shareable URL.
• Browser options: LibreWolf profiles, Brave’s fingerprint protections, and Tor Browser or Mullvad Browser for stronger anonymity use cases.
• Hosting options: Njalla, 1984 Hosting, GitHub Pages, and Codeberg Pages are discussed as recommendations and opinions, not formal reviews.

Main speaker/source: The Techlore host, answering questions submitted by viewers and live-chat participants.