19 chunks · 11 episodes
Google's Private AI Compute is a great development. It uses confidential computing for frontier Gemini models - hardware-enforced privacy, not just policy promises. Not perfect, but radically better. This should be table stakes! OpenAI, Anthropic, every frontier lab: make this the new baseline. This
... intended, and that every peer could validate it had that property? That's what Private Cloud Enclaves give you.
Private Cloud Enclaves provide a server you don't have to administer that you can trust to be honest. If you run a server yourself, you have to do all the administration. ...
...A system that is unsafe or difficult to expand to more users has a low ceiling. Private cloud enclaves are like cutting the gordian knot. The agency and control of local; the power and simplicity of the cloud.
...ftware from the command line–a small fraction of the population. But if you use Private Cloud Enclaves, a user can verify that the laws of physics are being faithfully executed… even in VMs in the cloud that they aren't running themselves. This could ...
...an presume a small number of beefy devices that are available practically 24/7. Private cloud enclaves provide a different approach. The power of the cloud, the sovereignty of local.
...dient to establish a different set of laws of physics. Confidential compute and private cloud enclaves can be used to extend those laws of physics beyond physical devices the user controls, and into the cloud.
... based computation is someone else's remote code that has control on your turf. Private cloud enclaves extend the agency from your personal device into the cloud. The device is in control, not the cloud. Which end of the connection has agency? The end...
... or data, and would only do so if compelled to do so or by accident. That's why private cloud enclaves put more focus on the service provider than the cloud host.
Private cloud enclaves allow you to have the security and privacy of on-device... in the cloud.
...a, and Confidential Compute doesn't really do anything about that. The ideal is private cloud enclaves. Confidential from the cloud host. Private from the service provider. Verifiable remotely.
...eed a private, secure base to allow safe composition of your data. For example, private cloud enclaves and information flow control.
Private cloud enclaves are possible to construct today. They require an expertise to know what you're doing and that you've done it correctly. But even when you've done it...
.... It's hard to charge for software that is written once and doesn't change. But private cloud enclaves must be paid for continuously by the user. If the user didn't pay for their turf, then they wouldn't call the shots. This implies a business model, ...
...ust right you get a kind of alchemical change. You get something you might call Private Cloud Enclaves. Think of a Private Cloud Enclave as your turf, in the cloud. Completely unlike traditional cloud computing. For a service to be considered a Privat...
...xtension of the device? For example, proving via remote attestation that it's a Private Cloud Enclave. The key question about which side is canonical is "where do the user's keys canonically live"?
...d SRE Or a subpoena from law enforcement. But now with confidential compute and Private Cloud Enclaves, we get the benefits of the cloud, but also under your control. The best of both worlds!
...ly local on devices the users controlled to be their agent. "On device" But now Private Cloud Enclaves allow you to extend your agency and turf off your device. An embassy in the cloud.
Both Confidential and Private Compute are about "provably locking observers out", but for different observers. Verifiably Private Compute says "The author of this service has locked themselves out from peeking inside, and as a user you can verify that externally." Authors can use tools like remote a