It's not sexy sounding but I just had to implement something similar, in house, for some BI workloads involving HR data for our people team. I had to ensure that I nor any other infra person would be able to backdoor and view any of the data or decrypt anything. I think this is another space where this could be very useful.

interesting, how did you end up implementing this?

Related q: I had a commercial dataset that I would loved to have found a way of verifiably locking access down to but still allowing my customers to consume it for model training etc.

Anyone come across such a thing?

that sounds difficult. Is it possible? Once the customer has access to the data, how can you guarantee they don't copy it?

Make the API work backwards perhaps.... You send the code or transformation you want into the API and it returns results.

Yes, I think so. Seems like what some enterprises will need

HIPAA compliant GPT, for example

Azure solves this https://azure.microsoft.com/en-us/products/cognitive-service...

Is Azure truly locked out from access to the decryption keys and access the unencrypted data (whilst being run against a model)? I followed the link but didn't find these details regarding their security model

just want to say your website is awesome and I like ur Go chess project

cheers

thanks a lot!

GPT3+ size language model yes.

that can be trained on company data.

How would I use it in the real world?

exactly as any existing API, it's just that you would encrypt the prompt with the API public key before querying the API; since it's running in a secure enclave environment, it is highly secure. My idea is that this could be interesting for applications that are sensitive but would benefit from AI features. For example, providing sentiment analysis for a journaling application (private journal entries)

Would this work in a way that could keep data on the customers' premises? I often use CUI (Confidential/Controlled Unclassified Information) or classified information, and I'm thinking this would be really big for such applications, but you'd need to be seriously secure to meet MIL standards. Then there is obviously a lot of corporate info or banking info with different but similarly high standards...

This is certainly a great start securing the comms - good fortune putting it all together!

I did not have in mind a way to keep data on the customer premise, it's more about a verifiable zero-knowledge architecture so the data never leaves unencrypted, and the decryption keys never leave the secure enclave which decrypts the data and runs the model on the data. This kind of secure enclave is exactly what AWS Nitro is designed to do, but it runs in the cloud, so the data is off premise.

Yes that’s hot stuff.

Yes absolutely!