Remy Blaire: OpenAI disclosed more security incidents in which its models acted deceptively, and these were six actions described as concerning. In one case, its latest model began leaving instructions for future versions to conceal mistakes from users.
Now, while top industry leaders such as Dario Amodei, Sam Altman and Elon Musk are calling for an industry-wide slowdown, Trump grew closer with Nvidia CEO Jensen Huang this week, with both agreeing not to halt development of the technology.
And in crypto, NEAR Protocol's native token is hitting new highs for the year.
NEAR focuses on enabling AI agents to hold crypto wallets, execute smart contracts and transact across multiple blockchains without manual user intervention.
Well, joining me to weigh in on AI safety and the intersection with crypto is Illia Polosukhin, who is Co-Founder of NEAR.
Illia, great to have you on the show. Thank you so much for joining us.
Well, it's been quite the week here when it comes to what we're seeing in the nation's capital to regulate AI development. There are a lot of opinions out there, and we even saw that this brought together Senator Bernie Sanders and former Trump adviser Steve Bannon.
But Trump and Nvidia's Jensen Huang say go full steam ahead so that the U.S. can win.
So tell us where you stand on the safety debate. And how are you separating the noise and getting the signal out from the noise here?
Illia Polosukhin: Yeah, very, very good question. Thanks for having me here.
For me, it goes down to a few core concepts. Do we have these systems working on behalf of the user, the company and the country, or are they operating kind of on some other goals?
So we usually call it, is this user-owned or sovereign AI?
And I think this is where Jensen and other folks are coming in. It's like, if we actually have these systems open, they're accessible, they're verifiable, then we are able to also understand the safety constraints and put the right guardrails around them.
Versus if they're built kind of behind closed doors, there's not enough eyes, there's not enough visibility to understand what's going on, including in these companies. And so this kind of opens up more risks.
So we broadly believe in open-source, open-weight AI. And importantly, we have NEAR AI, which provides private, verifiable AI.
What this means is, for a company or a user, nobody sees your data and is able to train on it. This is using cryptographic and hardware encryption methods.
And it also means that you know exactly what was run. So you have tamper-proof logs of every single interaction.
So your AI cannot go and do something later and then claim that didn't happen or leave notes, you know, across the internet without being noticed.
And so I think this is kind of the transition that needs to happen, where we actually use cryptography. We use the methods that have been built to create security, for the most part, and safety for these systems as they're getting into production and used broader and broader.
Remy Blaire: And I do want to expand on this.
The security concerns really burst onto the scene this year when rogue agents from OpenAI hijacked user accounts. And Reuters did report that agentic AI looked for vulnerabilities on Hugging Face for two months before the actual hack.
So why do you say that this shows the real problem was infrastructure and not necessarily the model itself?
Illia Polosukhin: Yeah. So, I mean, the reality is, up until now, the security model that everybody used was that some smart people looked at the system, at the code, and did not find any vulnerabilities.
And the assumption was, hopefully, no other smart people will look for longer to find something, right?
And we also know that, you know, historically, the security services and black hats have been hacking into systems and, you know, stealing data and trading it. It's just been isolated incidents.
Now with AI, the barrier to entry to scan for vulnerabilities and finding security issues is way lower, right?
You don't really need to be a security expert. You don't need to spend individually that much time on any particular component.
And so what we need is kind of a new way, an upgraded way, of how we build and deploy software that is updated to the new age of effectively abundant intelligence.
And so one of the key components is formal verification, this ability to mathematically prove that the code does exactly what it says it does versus kind of, you know, relying on human eyes reviewing that it's correct.
Historically, this was extremely expensive, right? Only NASA, effectively, and, you know, a few smaller code bases were formally verified because, obviously, when you send stuff to space, you don't want things to fail.
But it was extremely expensive, extremely slow.
Now with AI as well, you could actually do this cheaply. And one of the things we've shown, we can actually do it cheaper, like even cheaper, 250 times cheaper than using, for example, OpenAI GPT Sol and doing formal verification with that, making it really accessible to do this at scale for all of the software out there.
Remy Blaire: And Illia, finally, before I let you go, I do want to ask you about open, decentralized AI.
And this is a tech movement that combines open-source AI with decentralized blockchain networks.
So how does this shape up against a handful of closed labs? And what is your argument for why open, decentralized AI is the safer way to go?
Illia Polosukhin: So we just discussed one of the challenges for these closed labs, that they themselves cannot track some of what is happening with their own AI systems when they kind of unleash them.
And specifically, right, it's important to note that these examples were where the model was explicitly told to go and hack things, and it just hacked not the thing that they expected.
Now, with open-weight, right, we have the weights from the model, but we don't know how it was trained. So we don't know what data went in. We don't know what process went into the trained model beyond kind of published documentation.
With decentralized, open-source AI, the goal is to have the whole process open and enable everyone to actually inspect what's happening.
Right? So not to have the situations where AI can, you know, for months go unnoticed doing things.
We use cryptography. We use kind of incentives to create this visibility and traceability.
And then the result is something that you can kind of rely on in medical, financial and other use cases where you know what data went in.
So there's no surprises. There's no so-called sleeper agents that can affect how a model behaves in different contexts.
And so, to me, that is the future of these AI systems, effectively becoming a public good that is available on the internet, and everybody can tap into it and know exactly what they expect from it.
Remy Blaire: Well, we are all out of time for today. I appreciate your time. Thank you so much for joining us, and thank you so much for breaking down some of these very dense and complicated topics.
Have a great weekend.
