Inspiration

LLMs have opened up a new type of privacy risk. Any API call or chat with LLMs now leaks thought patterns and big model providers can develop a cognitive fingerprint of any user, this can be very concerning.

What it does

The product priweight.ai provides private inference, where all the prompts and chats cannot be read by anyone else. It also verfiably proves that the inference is coming from GPT OSS 120b

How we built it

The GPT OSS 120b is hosted inside a GPU TEE. The prompts are encrypted on the client side, sent to the TEE that decrypts it, runs the model and re-encrypts it and strams the token back.

No one can see the plain text ever, it can be crypographically proven that is the case.

Challenges we ran into

High gpu costs

Accomplishments that we're proud of

Building this with not much knoweldge in cryptography

What we learned

We need more GPUs to scale to many users

What's next for priweight.ai

Built With

Share this project:

Updates