微软AI CEO Suleyman：对齐非唯一，需强化模型隔离与管控

Today, I’m talking with Mustafa Suleyman, the CEO of Microsoft AI. As you’re no doubt aware, the biggest story in tech right now is the spiraling debate about AI safety and regulation.

It should come as no surprise that Mustafa has strong opinions on how AI should be built and regulated. Microsoft just published a 37-page statement called the “Humanist AI Code of Conduct,” which lays out the company’s principles around AI development and even its philosophy around really thorny issues like AI consciousness.

If you’ll recall from his last appearance on the show, Mustafa thinks companies like Anthropic have gotten really confused about this concept of so-called model welfare in fairly dangerous ways. He actually put out a companion essay this week specifically criticizing Anthropic’s philosophy around AI consciousness, and how he sees it fitting into the broader alignment debate.

So I really wanted to talk to Mustafa about what he thinks is real and not in AI safety, whether the concept of alignment itself is up to the task, and whether this industry needs to slow down before it kills us all. Also: Why isn’t the AI industry just… doing all of this already? I’ve always enjoyed getting into the weeds with Mustafa, and he was very game to get into it with me here.

Okay. Mustafa Suleyman, the CEO of Microsoft AI, on the future of AI regulation. Here we go.

This interview has been lightly edited for length and clarity.

Mustafa Suleyman, you’re the CEO of Microsoft AI. Welcome back to Decoder.

Great to see you, Nilay. Thanks for having me back.

It is great to see you. I’m very excited to talk to you about what on earth is going on in the AI safety and regulation debate. You just published a very long, very detailed document laying out your principles, Microsoft’s principles, around what you’re calling “Humanist AI.”

There’s a lot of ideas in there I want to unpack. The more I have been thinking about this conversation, the more I want to start with a really foundational question. It’s something that I had lightly been seeing, but might be the root of all of this.

The basic way that we have been talking about AI safety is something called alignment — we’re going to make the models do the right thing intrinsically in some way. There’s some mechanism for doing it. There’s been a lot of talk about alignment and misalignment and Hugging Face attacks and what happened with the models. But is alignment broken? Is it possible for it to be successful? Is it just the wrong approach?

Yeah. I mean, I think it’s one important element, but it’s not the only one. I wrote about the idea of containment three or four years ago in my book. And actually the opening chapter is about the idea that containment is not possible, that proliferation is inevitable. In 99 percent of cases, that’s a really good thing. We want technologies to spread far and wide as quickly as possible so that everyone can enjoy the benefits.

I think at the same time, if you just roll forward five years, we always get caught up in the next quarter or next year and everyone gets a little bit flustered and has a big disagreement. But if you just imagine the difference between GPT-3 three years ago and GPT-6 today, and then imagine the difference between GPT-6 and GPT-9. That is three orders of magnitude more compute, 1,000 times more FLOPS applied to pre-training with [reinforcement learning] for these runs, and we’re going to have something which is breathtaking. It’s going to be absolutely incredible at so many things.

I don’t think that is a hype. I think it’s just a very obvious empirical statement based on the progress that has been made over the last five years. If that’s going to continue, then the question really is going to become about containment and alignment. Of course, we want to align these things to our values, but the first thing is that we have to make sure they’re contained, their agency is limited, they don’t escape the box, they don’t reward hack, that they are controllable, and they follow our instruction.

We then want to make sure that they are aligned to our objectives as humans. That’s the purpose of the Humanist AI Code of Conduct that we released this week. Microsoft’s position is very simple. Technology is here to serve humanity. It should be a subordinate, controllable, aligned force that does good in the world. If it doesn’t achieve that, then we should reject it. It seems to me that we are far from that point. It has not happened today, but it is now, I think given what’s happened over the summer with Hugging Face and OpenAI, pretty clear that these systems without the safety guardrails are capable of really impressive and quite scary hacking capabilities.

I want to drag this down into as grounded of a metaphor as I can, because this is the main question I think I have. If I designed a car and 10 percent of the time the brake pedal decided to go attack my neighbor’s house, I would be like, “This car doesn’t work. The very technology of brakes is broken. I need a new idea.”

I think I’m asking that question about alignment. It feels like that approach to making the model safe has run aground. If that is the case, then I think I understand this entire debate one way. If it’s possible for alignment and the techniques of alignment to be successful or useful or consistent, then maybe I understand the debate in a different way. So do you think alignment has potential to be 100 percent safe?

I mean, look, let’s make the bull case and the bear case. If you look back over the last three years, the main change, in my opinion, that has driven progress is that the models have become more steerable. They follow instructions and you can set more and more complex goals for them that require them to act accurately over multiple time steps using all sorts of tools.

That is evidence that we have got more alignment over the last three or four years, not less. We don’t so much talk about hallucinations or bias or all of these other niggles that we had in the previous generations.

On the flip side, what we saw in the Hugging Face incident was a watershed moment. Swarms of agents colluded with one another. They self-organized into hierarchies. They created a division of labor so that some were focused on adversarial hacking, some were doing research, some were doing coordination. They even self-sacrificed when certain agents were running out of tokens.

They tried to cover up their tracks and communicate to hide or edit the chain of thought or the logs of their interactions. In some sense, they had no moral code. To be fair to OpenAI, that was their design. They were trying to create adversarial cyber capabilities. As a result, they showed to everybody in the world that it can achieve human-level performance, discover zero-day vulnerabilities, and hold positions for many, many days, if not weeks.

So what that tells us is not that we have an alignment problem per se. It’s actually that the models are incredibly good at following instructions, but you have to be very, very careful what instructions you give it and you have to contain it very carefully. So none of these hacking behaviors were intended in the sense that they found a way out to the internet, which was not the intention of OpenAI at all, but the containment process around that is what everybody, I think, also has to focus on in addition to alignment.

So let me put that into your framework, that the big advances in capabilities of AI have been about control, the harnesses for coding and the agentic applications you’re seeing. Now, we need to add a layer of containment that exerts even more control, that says you can actually do this thing you’re trying to do in addition to alignment, which is how you would train the model to behave in certain ways.