What does this AI safety tester really know?

This is all reliant on this stuff not getting into the hands of rogue states, bad actors, all that.

I mean, reassure me.

Any of the frontier models that you would have access to, the ones that we're actually more worried about because they have the highest level of intelligence, we're talking about what's coming out of OpenAI with ChatGPT, Anthropic, with Claude, Google, Gemini, These models are restricted.

So when the public is interacting with them, you cannot go to chat GPT and ask it how to make a bioweapon.

It is clearly going to not allow that to happen.

So there's all sorts of guardrails in what the public has access to.

I think it's actually a really interesting and unusual moment in history where The public is having access to what's happening in science and R&D.

When we think about the Manhattan Project, it was all behind closed doors.

No one really knew what was going on except the scientists involved.

Now we have all this frontier science happening to create AI models, and you can go on your phone and you can download an AI app.

It's quite an unusual situation, and so a lot of focus is put on making sure what people actually access is passing through lots of safety layers.

So the next generation models that come out, because these things are coming out at a rate of knots, they're going to be programmed to not...

So if somebody finds a prompt to cleverly try and extract from this model the best way to make a biological weapon, say, this model is programmed just not to deliver that information.

Yeah, that's the goal.

How can you tell that you've worked out all the

different devious prompts? So there's the monitoring and the guardrails that go around that.

And then I think what we're getting to is having enough friction involved that it's not easy, even if someone has that bad intent.

Let's say I wanted to create a biological weapon.

I don't have the resources, the money, the tools, the experience to be able to do that.

So the knowledge is not actually the limiting factor at that point.

But there are terror groups sitting in deserts in the Middle East or whoever who do.

You've got the resources and the willpower and the time and the technology.

Yeah.

And so that's, I think, why we're putting, there's this big shift, so there's this race dynamic, and the race has been characterized by let's make super intelligence, let's make the biggest, smartest thing possible, whereas what we're starting to see is actually when that interacts with society and when that interacts with business, Boardrooms don't want to be responsible for having an AI model in their company that goes and discriminates against customers or creates a legal or PR risk.

So the economic incentive is changing to say, we want to have these really intelligent systems, but we want to do it safely, responsibly.