So come back to China and what is coming out of China with respect to these concerns about the intelligence compounding and becoming super intelligent and getting away on us? What are we hearing out of China?

So a massive paper, really, really good paper, came out of China about two weeks ago called The Last AI Built by Humans Toward Genuine Recursive Self-Improvement.

And it surveys everything that we know.

And at the very top of the paper, they put a little, you know, TLDR too long, didn't read.

If you want to know what's inside of it, what they say is like human evolution, AI evolution will unfold through a vast and extraordinary history.

Everything AI has achieved so far is just a drop in the ocean.

And the paper goes on to explain, yes, recursive self-improvement could probably be a thing.

Why not? Let's go to that one.

How the hell do you wrap your head around this issue?

In some ways, the easy bit had been you have OpenAI or you have Anthropic or Google as a provider and you're paying, you were paying by subscription.

Everyone's now been moved to pay as you go, right? You basically pay by the token.

And in some ways that makes the accounting straightforward, but people are now getting bill shock.

The bill shock is not because provider costs have gone up.

They have kind of gone up, but really not.

It's because, in fact, organizations are finding more ways to use tokens.

Now, you shared with me an article this morning, and we were talking about it earlier, just around this kind of agentic, you know, what does the world look like with the agents operating? Can you just speak a little bit about what you think is going to happen over the next 12 months, here in Australia specifically?

Sorry, Mark.

But if we put this in a personal context, remember that Anthropic is the child of a divorce, right? And they were all at OpenAI together, and they were all working toward superintelligence, as they call it.

And Amadei decided that Altman wasn't as safe.

And so went off and did his own thing.

And now there's a competition between the two of them about who can do it.

And the way they think about it, and I've had people really come to me and very strongly say, Mark, this is a cult because of the way they think about it.

But the way they think about it is that it is a literal winner takes all.

meaning it's not connected to another device and it's not connected to the Internet, so it can only do damage on its own machine.

But in acknowledging how fast it is learning and how good it is getting, he is foreseeing a real sweep through jobs, and of course, then you end up in our current societies, if you have very high levels of people unemployed, you have the financial and the personal stresses that creates for the society.

Yes.

His, his data's a little out of date.

The, the economists have been looking very closely at this, Catherine, and the current best read is that it does seem to have some impact about people entering the workforce, and that's changed, although there's data that points to the fact that it's more work from home than AI that's affecting that.

But in terms of actual working professionals, no.

You know, Geoffrey Hinton, the father of modern machine learning, predicted five years or 10 years ago there would be no radiologists left in five years, and now there's more radiologists than ever because it turns out having good AI radiology tools increases the demand for radiologists.

So making blanket statements like it's gonna put everyone out of work, which seemed very convincing a few years ago, seem a lot less supportable now.

Were there spurts? What would've been some of the real, uh, sort of bursts of progress?

Carry on with the

agents.

And the thing is, the folks who have been tracking this, METR is the name of the organization, METR.org. They've been tracking what we would call the task horizon, how long an agent could get away with it before it falls over.

And again, 2023, about a minute.

But the thing is, is that there has now been, similar to Moore's law, there's been a regular doubling.

And that doubling cadence has about a seven-month cycle.

When you get to November of last year, in a three-week period, you get, first, Gemini 3.1 gets released by Google.

effectively a worker, right? That's a worker that costs $6,000 instead of $65,000 a year.

That's the current advice people are giving and getting.

The problem being, if one of them is smarter than the others, the others are all going to disagree, but they don't know.

So you're going to get the wrong answer.

Now, that may be OK in the situation.

But if you have one that's brighter than the others and brighter than you and most of the people you could bring to judge it, for that area, we can say, OK, that's relative superintelligence.

And we know that with mythos, mythos is really, really good at finding bugs and bashing bugs and actually creating exploits from chaining bugs together.

And we know that its abilities are basically better than a handful.

But I also don't want AI doing the directions because it might riff on them and tell me to drive off a cliff or something.

Now, can AI help with that? Can AI help people monitor their process?

Okay, I guess the answer is theoretically, right? Right.

But we haven't seen a lot of good tools on that yet, right? Right now, what we're seeing is mostly tools for tool makers.

So, you know, you talk about... people getting very productive.

Right now that's happening largely with people who are using programming tools.

And there's, of course, this new phrase that's going around called vibe coding, which is where a single person who may or may not know how to code sits down and directs enough prompts to a powerful enough audience language model, a frontier model like Gemini or Claude or one of the later OpenAI models, and basically gets it to build all of the code around a specification which is not particularly technical, but is more like, you know, I wanted to do this thing and I wanted to do that thing.

A really good example of this, which is current to the moment, you know, we know about Coldplaygate.

But on the other hand, how does that affect your clients when you talk to folks about this? Are they thinking rationally about this or have they been terrorized into AI?

Now, can AI help with that? Can AI help people monitor their process?

Okay, I guess the answer is theoretically, right? But we haven't seen a lot of good tools on that yet, right? Right now, what we're seeing is mostly tools for tool makers.

So, you know, you talk about... people getting very productive.

Right now that's happening largely with people who are using programming tools.

And there's, of course, this new phrase that's going around called vibe coding, which is where a single person who may or may not know how to code sits down and directs enough prompts to a powerful enough audience language model, a frontier model like Gemini or Claude or one of the later OpenAI models, and basically gets it to build all of the code around a specification which is not particularly technical, but is more like, I wanted to do this thing and I wanted to do that thing.

A really good example of this, which is current to the moment, we know about Coldplaygate.

Because that's the big thing where that executive and girlfriend slash CPO were trapped on the camera, the kiss cam at a Coldplay concert.

But on the other hand, how does that affect your clients when you talk to folks about this? Are they thinking rationally about this or have they been