Anthropic envoy says AI progress will ‘move at the speed of trust’

Por Ryan Heath — POLITICO – TOP Stories

CANBERRA — Former U.S. Ambassador to Australia Jeff Bleich travels between the U.S. and Australia as Anthropic’s international special envoy, a role that has placed him at the center of an increasingly urgent debate about how governments should respond to frontier AI.

In recent days, Anthropic researchers have publicly warned that rapidly advancing systems could escape human control and, in the most extreme scenario, kill all humans. One senior alignment researcher put the chance of human extinction within a decade at 10 percent, while another warned that the industry is moving ahead without a credible plan for managing the risks.

Anthropic has stressed that these are personal views, not company forecasts, while CEO Dario Amodei has called for a slower, more internationally coordinated approach to frontier AI. Against that backdrop, Bleich spoke to Ryan Heath at the Sydney Dialogues AI Masterclass, hosted by the Australian Strategic Policy Institute.

This transcript has been lightly edited for length and clarity.

As we move from worrying about robot overlords to AI extermination, it seems to me that either we’re either not proposing enough action, or we’re overstating the risks? How should we price the risks?

The risk grows as the capability gets bigger and bigger. At this point, the models are powerful and there are risks associated with them, but we understand how they are operating. You don’t want the model to get too far out ahead.

People should be appropriately cautious about this.

Do you or will you need help to keep pace with the risks around these models? Do we need to bring government into the room?

Yes, absolutely. That doesn’t mean regulate for the sake of regulating.

Do we need to develop an early warning system amongst model makers? Or does it require congressional action or a presidential executive order?

We’re heartened by what appears to be unanimity among the frontier AI labs that we should be doing this. We should be bringing in external evaluators. The devil will be in the details.

I think we want to work with other labs to make sure that we set standards and hold ourselves accountable. We’ll probably have to work with the government as well, in part to make sure that nothing we’re doing violates antitrust laws. And I think we also want to work with the government generally because they’re in the best position to hold this to account.

You can do that without slowing our ability to achieve super intelligence.

If you can’t get action among all model makers or from Washington, can the Australian government get involved in setting the pace on AI safety?

I don’t think it makes any sense that something this important would be decided either by just company executives or by leaders in only two countries. And so I took a lot of comfort from the fact that a nation like Australia, which is so forward-looking, which has so much capability internally, that has punched above its weight globally, is one to take on that mantle.

Any country which is doing training of AI is going to have to establish norms, rules, conditions within their country relating to that.

The old adage out of Silicon Valley was “government can’t keep up with us.” Now that the AI could or is getting ahead of the tech companies, do you have some extra sympathy for regulators?

I was a regulator, so I’ve always had sympathy for regulators.

This is a moment where we’ve been so much part of an attention economy, and we’ve given up on rational thought in too many cases when it comes to important things. You want people who are serious-minded and human-centered to be working on this in a collaborative fashion.

I look at the history of Silicon Valley and the success that happened there, and what was the story? It was government sending signals. It was working on mitigation. It was making it so that people believed in that technology and it flourished.

The Australian government is sending a bunch of signals to you right now that it would love some Anthropic investment. How certain are those plans? What sets Australia apart as an interesting place for your investment?

Australia has anticipated issues before and made major changes and adapted to avoid them. It is a well-recognized and well-respected country as a thought partner and place for governance. So I think the capacity to build an ecosystem in a safe and responsible way.

In terms of talent and capability, we’ve got a ton of Aussies at Anthropic in San Francisco, who I think would join my kids and move back here if they had the opportunity to do training in Australia.

There’s a tradition on all sides of politics here to be straight with people. I think it matters. It was the only major economy that didn’t go through a recession during the global financial crisis because of some hard thinking up front.

How critical is a copyright or intellectual property resolution for Anthropic to turn up and build compute here?

They are issues that have to be worked out democratically so that there is a true social license for frontier developers to operate [and] also some legal certainty, like what can we do, what can’t we do? We’ve had a very honest dialogue. I’ve shared information, and others from Anthropic have shared information about technically what is and isn’t possible.

It won’t be easy to get these things done, but that actually makes us feel better about going to Australia.
We didn’t come here because it’s easy. We came here because we knew you would make it hard.

You’d rather get it done right, even if it takes a little while longer.

Well, that’s the issue: We have a limited amount of time

We’re talking months?

Yeah. We’ve been very candid about these sorts of timelines. We have to continue training. We need to, and if you’re putting in more compute for training, we’re going to be putting in more compute for safety and safety evaluations.

If anything, you need to accelerate because you need even more compute for the research related to safety and staying ahead of the models themselves.

The most important thing is that there’s trust. The adoption of this technology, the capacity of it to serve humanity, that’s going to move at the speed of trust.

We’re not just in an intelligence race, we’re in an adoption race. And you don’t achieve deep adoption if you don’t have trust, if workers don’t feel like this is something that’s going to benefit them rather than steal their jobs. How can you make sure there is a bottom-up movement for people to adopt AI?

We need the democratic governments to come together and work through specifically what the rules of the road are going to be.

I think it’s widely acknowledged we have a productivity crisis. But we can easily see how AI becomes a major labor market disruption and problem. Can you work with the Australian government to make AI be a job creator, rather than just something that tears up industries and leaves a generation without jobs?

This opportunity is like the second industrial revolution, and this is a country that is particularly well positioned to be one of the leaders in it. The countries that stepped up and adapted to this new technology were far more productive. They were far more secure. And they ended up having their economies catapulted.

And a lot of their policies become not only admired around the world, but eventually emulated around the world. So the same is true with AI here in Australia. If Australia becomes a magnet for AI talent and capital, what is going to flow [is] an ecosystem which will produce far more jobs and opportunities.

I started in my legal career in the relatively early days of Silicon Valley. So I’ve seen precisely how that works in Silicon Valley. It started out as, you know, Hewlett Packard and Intel and one or two other companies that made Silicon Valley. But on the other hand, the question was, well, what’s going to come next after those? And the people who worked in those companies started their own companies. And people who worked in their companies started their own companies.

Before you knew it, it has been the economic engine of California and California’s the fourth largest GDP in the world.

The message is: Get a Silicon Valley in Australia?

I don’t want to over-promise and under-deliver, but as an honorary citizen, I would hope so.

And you could see that opportunity, that kind of wealth generation occurring. And when you’ve got that kind of wealth generation, then when you set ambitious goals like 100 percent renewable energy at some point, you could see how you’d be able to get there because you would have long-term massive capital commitments into Australia.

The old thinking about move fast and break things and just put out spin, we’ve all been broken and burnt and spun enough. I think the public is ready for straight talk. And there should be an obligation on the part of companies to do it.

What I think our experience to date has demonstrated is that doesn’t cost you in your ability to be more successful, because again, in this space where the technology is so powerful, progress will move at the speed of trust.

Fonte: POLITICO – TOP Stories