Demis Hassabis 论 AI 新时代：呼吁设立前沿 AI 标准机构

Google CEO Demis Hassabis offered us a first rate second rate essay, A Framework for Frontier AI and the Dawning of a New Age. I’ll go over that essay and various responses to it in Part 1.

Part 2 of this post then covers Alex Turner’s resignation, and his story about how he tried and failed to prevent Google from signing up to allow the Department of War to use its models for essentially whatever the government wants, including autonomous weapons.

Demis Hassabis sold DeepMind to Google on condition that something like this would not happen. Yet here it is, happening. A cautionary tale.

I will cover Kimi K3 tomorrow. I am hoping to know more by then. Please do share any reactions or info about it in the comments here.

The Core Statement and Request

He saying we are standing in the foothills of the singularity.

His ask is a Frontier AI Standards Body within the US Government, similar to FINRA, that would govern ‘frontier labs,’ defined as any company that produces a frontier model based on various technical benchmarks. Evaluations would be updated regularly, and vulnerabilities would be addressed, both before and after release.

He is excellent about stating that this is big, really big, no bigger than that, it be big.

Demis Hassabis: I’ve spent my whole life working on AGI because I’ve always had a deep conviction that, if built and deployed responsibly, it would prove to be one of the most beneficial and transformative technologies ever invented. AGI cannot be compared to standard technological breakthroughs, not even ones as consequential as the internet or mobile – it is much more akin to the discovery of electricity or fire. If you stop to think about it, we’ve essentially found a way to make sand think. It’s miraculous.

The magnitude of this technology’s impact will be unprecedented, perhaps 10x of the Industrial Revolution at 10x the speed. It will help us solve some of the biggest problems society faces from accelerating drug discovery to developing new clean energy sources to creating novel advanced materials. We could even reach a point where resources are no longer the limiting factor for human progress, leading to an amazing new era of abundance.

Things Left Unsaid

There is definitely a ‘don’t say the thing’ aspect of this, where he won’t name what the downside risks actually are. When Demis says ‘experts disagree’ he is rather avoidant about the way in which they disagree here.

Nate Soares (MIRI): I’m glad Demis acknowledges that this is a “pivotal moment in human history” during an “extremely intense” race. I’m disappointed that his proposed solution is a “standards body” to evaluate whether models are dangerous, with no plan for what to do once they are.

I’m glad he acknowledges that “experts disagree.” I’m annoyed that he glosses past how the disagreement is about whether there’s a ~5% or ≥50% chance of total catastrophe. We’ve gotta do better.

Aaron Scher: Glad to see AI CEOs speaking publicly about their views on AGI. I think Demis is wrong about his policy prescription: it’s far too little too late. When he says the experts disagree, he means that some think 5% this tech kills literally everybody, some at 40%, some at 90%.

Clearly this is strategic, but if you don’t already know, or are looking to not realize, it is very easy to come away thinking that Demis does mean the effect on jobs, even though when he says ‘safely’ he very much does not (primarily) mean that.

The Proposal

Demis Hassabis: … On the horizon, we will need robust safeguards to maintain control of increasingly agentic, recursively self-improving systems – and tackle unknown issues that will only become clearer over time.

… I’ve always believed in the power of human ingenuity and creativity to solve any problem. I’m confident that mitigating the technical risks related to AI is a challenge we can collectively address, but only if we give ourselves the time and space to get this next crucial step right. Currently, as a field and as a wider society, we aren’t doing that.

He makes clear part of this is about giving us options, including for a slowdown.

The strength of this approach is it would be technically focused, while at the same time supporting innovation and incentivising responsible behaviour. It is designed to keep up with the field’s acceleration and adapt to the biggest risks as they are identified, and could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary.

Demis keeps it short, not offering many details. To the extent that he has laid out a proposal, it seems to be a good one. It is definitely an improvement on the margin.

Jack Clark (Anthropic): At this point, everyone at the frontier of AI agrees that third-parties should test out AI systems and use these to develop standards to feed into policy – excellent to see @demishassabis laying out a framework to do this!

Samuel Hammond: It is striking to see leadership at Google, Anthropic, OpenAI and Microsoft all fairly independently sounding warning alarms about an imminent technological acceleration.

Thus I file this post and its ask, as high praise, under ‘the least you could do.’

A Good Start But Insufficient

I agree with Peter Wildeford that while better than nothing FINRA is not a great model here, with heightened risk of regulatory capture, and not a substitute for full government action. You need an SEC to your FINRA. That doesn’t mean don’t make the FINRA. It does mean you still need the SEC.

Would such a (at least partly) voluntary regime, only for models intended for release, and without a related binding intentional agreement, be sufficient to solve the problem? No, again it’s just way better than doing nothing, as Peter Wildeford and many others noted.

You do not need to believe, as Aaron Scher and Connor Leahy do below, that only a full halt would be sufficient here, to know we have a long way to go. Demis’s statements here, if you know what they actually mean, imply a level of danger and urgency that is not reflected in the proposal.

Eli Tyre: > Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release.

Is this proposal only intended to address risks from models that companies plan to release? If a company develops a frontier model and never releases it, only deploying it internally to develop even more powerful AI capabilities, are they thereby exempt from this oversight scheme?

Connor Leahy: While @demishassabis is right that we need urgent action to address risks as we approach AGI (and superintelligence, I’d add), the correct response to the threats is not a ‘self-regulatory organization’.

We need to prohibit superintelligence, not give industry regulatory power.

Aaron Scher: … The extinction threat, the “only a few short years”, the “10x the Industrial Revolution”—these aren’t indicators that point to “let’s evaluate models to understand their capabilities and have voluntary safety standards”. We need to back off, we need to halt the creation of ASI.

Point 2: I agree with the attached quote that we need more time. But I think Demis’s optimism is a vibe, not a trustworthy basis for predictions. Rob Miles says it best in this video, if an asteroid we’re headed earth’s way 200 years ago, we’d just die

Point 3: As others have pointed out, it’s not clear that this proposal would reduce risks from internal deployment (it seems to focus on public deployment and pre-deployment testing), but internal deployment is where much of the risk is.

Point 4: I don’t think the proposed body could actually enact, verify, and enforce a slowdown; there’s ambiguity about what’s voluntary. Again, I think we need a long-term international treaty and to actually back off, not just to slow down a little.

Skeptics Of Future AI Capabilities