Open Source AI and Confidential Compute with Anthony & Marks (Co-founders, OpenSecret)

Anthony & Marks were in town for LlamaCon so we got a chance to jam! We discussed a variety of topics including:

* The state of Llama, DeepSeek, and Open Source AI

* Dario’s take “On DeepSeek and Export Controls” https://www.darioamodei.com/post/on-deepseek-and-export-controls

* Data Quality is key - Garbage In, Garbage Out (speculation: Facebook data not nearly as valuable as Twitter or Reddit?)

* Trend of companies not using proprietary models to protect data privacy

* AI Alignment towards truth: how to ensure LLMs show real chain of thought

* The model host matters most - open source models get filtered in dramatically different ways depending on hosting and system prompts

* The value of open source models in preventing brainwashing/mind control

* Learnings from Mutiny that led to Open Secret

* Confidential Computing Explained

* Secure Enclaves moving from the iPhone to the cloud

* OpenSecret is like Apple’s private cloud compute but open source, verifiable, and anyone can use it

* How to craft better prompts (meta-prompting) to get the most out of AI

* How OpenSecret and Maple AI verify a chat is private. Enclaves as the next upgrade of HTTPS https://blog.opensecret.cloud/opensecret-technicals/

* Using AI to audit code

* Secure enclaves as an ingredient for autonomous digital life

* o1 isn’t a chat model (and that’s the point)

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit hivemindvc.substack.com