Can AI Coexist With Privacy? Proton’s Andy Yen Says It Will Have To

I’m trying not to get too nerdy here, but are you talking about homomorphic encryption?

No, it’s basically—do you know about TPMs, trusted platform modules, and how that works a little bit?

Yeah. If I could try to summarize it, it’s kind of like: You can’t encrypt the whole chat, but you can use cryptography to create a technical guarantee that part of the computer won’t look at another part of the computer.

There’s a kind of secret enclave where these chats can happen, and it cannot be inspected by the rest of the server, essentially. So that creates a kind of technical guarantee of privacy. It’s certainly not end-to-end encryption, but it seems like something, in fact, that you could be doing today. I mean, other people are doing it. So can you explain why Proton is not?

This is why I didn’t say end-to-end encryption, I said end-to-end protected. Because it’s not technically end-to-end encryption, it’s as you say: You can run the processing on the GPUs, but in a way that the rest of the pipeline that delivers the tokens to and from the GPU is not able to view it.

And through using this secure execution environment, you can have a cryptographic guarantee that nobody else along the pipeline could look at what was sent to the GPU.

That’s maybe the best simplified way to explain it. And yeah, it’s possible to do some of this already today, but there are certain, let’s say, performance issues. It’s still a little bit clunky. It doesn’t have proper support across the entire stack. So I wouldn’t say it’s impossible, but it’s gonna get easier to do in the years to come.

Got it. Well, it is interesting, though, that people are already doing this today. Other products do use these trusted execution environments; they don’t just promise not to look at your conversations. But you don’t feel like it’s up to snuff yet? It doesn’t have certain features? What’s lacking? Why is Proton not using this now?

Well, first, you have to re-architect essentially your entire platform in a way to do that. So it’s quite a bit of work. Also, when we put out Lumo last year, a year ago, the technology for AI was a lot more immature, right? Then there’s also the interfaces with the GPUs themselves.

So, you know, Proton is not running on a cloud. We run our own infrastructure. And so then you’re pipelining it, and the libraries that Nvidia provides to do that have to be good, have to be stable, they have to be completely working, they have to be fleshed out, they have to be, you know, making sure there’s no security issues.

So the technology for me, like I said, it’s, it’s almost mature, but we’re probably in a better position to fully deploy it in, let’s say, the next year or two.

OK.

So it’s pretty close. Some people are already doing it, but we have a bigger deployment environment.