If You Can't See Inside, How Do You Know It's THINKING? [Dr. Jeff Beck]

So, how do you know if the AI is actually thinking? That's kind of the question, right? We're going to dive into that today.

It's a really interesting space because we're talking about things that are happening inside of a black box. And if you can't see inside, how do you know it's thinking?

That's the question we're going to try to unpack today with Dr. Jeff Beck, who's the founder and CEO of Zama. He's building cryptography for AI, and he's got a really unique perspective on this.

Jeff, welcome to the show.

Thanks for having me.

So, let's just dive right into it. What is cryptography for AI?

So, the core idea is to use cryptography to protect data when it's being used, not just when it's stored or when it's in transit, which is what most people think about when they think about cryptography.

The idea is to be able to compute directly on encrypted data. And the most promising technology to do that is called FHE, Fully Homomorphic Encryption.

And so, what FHE allows you to do is to encrypt data, send it to a third party, have them compute on it, and send you back the encrypted result. You decrypt it, and you learn the answer, but the third party never learns anything about the data or the answer.

So, how does that apply to AI?

Well, AI is all about data. It's all about training models on data and then running Inference on data. And so, if you can protect the data during these two phases, you can unlock a lot of new use cases.

For example, you could train a model on medical data without ever seeing the data. Or you could run Inference on financial data without ever seeing the data.

So, what are some of the specific use cases that you're seeing people get excited about?

So, the first one is privacy-preserving machine learning. So, this is where you want to train a model on data from multiple sources, but you don't want to share the data with each other.

For example, you could have multiple hospitals that want to train a model to predict cancer, but they don't want to share their patient data with each other. With FHE, they can train a model on all of the data without ever sharing the data itself.

Another use case is privacy-preserving Inference. So, this is where you want to run a model on data that you don't want to share with the model provider.

For example, you could have a financial institution that wants to run a fraud detection model on their customer data, but they don't want to share the customer data with the model provider. With FHE, they can run the model on the data without ever sharing the data itself.

And then the third use case is privacy-preserving Agents. So, this is where you want to have an Agent that can access your data and perform actions on your behalf, but you don't want the Agent to be able to see your data.

For example, you could have an Agent that can manage your finances, but you don't want the Agent to be able to see your bank account balance. With FHE, you can have the Agent access your data and perform actions on your behalf without ever seeing the data itself.

So, how does FHE actually work? Can you give us a high-level overview?

So, the core idea is to encrypt the data in a way that allows you to perform computations on it without decrypting it. And the way that this is done is by adding noise to the data.

So, you encrypt the data by adding a lot of noise to it. And then you perform computations on the encrypted data. And the computations also add noise to the data.

And the key is that the noise doesn't interfere with the computation. So, you can perform computations on the encrypted data, and you'll get the correct answer, even though the data is encrypted and noisy.

And then, at the end, you decrypt the data, and you remove the noise. And you're left with the correct answer.

So, it's kind of like adding a bunch of static to a radio signal. You can still hear the music, but it's harder to hear. And then, at the end, you remove the static, and you're left with the clear music.

So, what are some of the challenges with FHE?

So, the biggest challenge is performance. FHE is very computationally expensive. It's much slower than performing computations on unencrypted data.

So, that's the main challenge that we're working on at Zama, is to make FHE faster and more efficient.

Another challenge is that FHE is still a relatively new technology. There aren't a lot of people who understand it, and there aren't a lot of tools and libraries available.

So, we're also working on making FHE more accessible to developers by building tools and libraries that make it easier to use.

So, what does the future look like for FHE?

I think that FHE is going to be a very important technology in the future. I think that it's going to enable a lot of new use cases that are not possible today.

I think that we're going to see FHE being used in a lot of different industries, such as healthcare, finance, and government.

I think that FHE is going to help us build a more private and secure world.

So, if someone wants to learn more about FHE, where should they go?

So, the best place to start is our website, zama.ai. We have a lot of resources available on our website, including blog posts, tutorials, and documentation.

We also have a community forum where you can ask questions and get help from other people who are working with FHE.

And then, if you're interested in contributing to the development of FHE, we have a number of open-source projects that you can contribute to.

Awesome. Well, Jeff, thanks so much for coming on the show.

Thanks for having me.

Key Takeaways:

• FHE allows computation on encrypted data, protecting it during use.
• It unlocks new use cases in privacy-preserving machine learning, Inference, and Agents.
• Performance is a key challenge, but advancements are making FHE more accessible.

Link: zama.ai