Moxie Marlinspike has a privacy-conscious alternative to ChatGPT | TechCrunch

"If you're at all concerned about privacy, the rise of AI personal assistants can feel alarming. It's difficult to use one without sharing personal information, which is retained by the model's parent company. With OpenAI already testing advertising, it's easy to imagine the same data collection that fuels Facebook and Google creeping into your chatbot conversations. A new project, launched in December by Signal co-founder Moxie Marlinspike, is showing what a privacy-conscious AI service might look like."

"First, Confer encrypts messages to and from the system using the WebAuthn passkey system. (Unfortunately, that standard works best on mobile devices or Macs running Sequoia, although you can also make it work on Windows or Linux with a password manager.) On the server side, all Confer's inference processing is done in a Trusted Execution Environment (TEE), with remote attestation systems in place to verify the system hasn't been compromised. Inside that, there's an array of open-weight foundation models handling whatever query comes in."

Confer is an AI assistant designed to prevent data collection and ad-targeting by ensuring hosts never access user conversations. Conversations are encrypted end-to-end using the WebAuthn passkey system; the standard performs best on mobile devices or Macs running Sequoia, with password-manager workarounds for Windows and Linux. Server-side inference runs inside Trusted Execution Environments (TEEs) with remote attestation to verify integrity. An array of open-weight foundation models handles queries inside the TEE. The architecture is more complex than standard inference setups, but it prevents use of user conversations for model training or targeted advertising and prioritizes intimate privacy protections.

Read at TechCrunch