It’s The Compute, Stupid

the conversation on AI Safety in China needs some common sense

I have read a whole lot of piping hot takes on Chinese AI Safety over the last 48 hours; many of them brilliant, many of them idealistic, but not all of them sensible.

These last 48 hours also corresponded to my aunt’s 60th birthday party; a woman who has never used AI in her life. I explained my job to many people who have never used AI, and that experience was valuable- so much so that I decided to write it up, in case you’re interested in reading a sensible solution to our current situation, pitched in a non-technical way.

My view is right there in the headline: it’s compute. That’s it.

a sensible explainer & pitch

Compute is a fancy term for computer power- the chips and the other technology needed to run AI on the chips.

You use compute for a few things:

Inference: giving access to the Model that People and Businesses Pay For.
Training: making a Bigger, Better Model for People and Businesses to Pay For
Capability Research: improving the Bigger, Better Model People and Businesses Pay For
People Safety: Preventing Awful People from using the Bigger, Better Model
Alignment: Preventing the Model itself from becoming Awful People

If you’re American, there’s status and money and compute available to do all these things.

If you’re Chinese, there’s status but much less money and much less compute.

In China, some of these things are non-negotiable, like Bigger and Better, because that makes your stock price go Up, a direction we all like very much because it leads to more money and more compute and also makes your parents and friends proud.

If you need Bigger and Better but you don’t have the compute for Capability Research, you probably need to figure out ways to do research really efficiently, and you might even decide to help yourself to someone else’s Capability Research.

Once you’ve hit a certain level of Bigger and Better, you start worrying about Awful People. Once you’re really Bigger and Better, you worry about models becoming Awful People.

Now if you’re a China Idealist, you probably think “it’s good to give China all the compute”, but many Americans who love America very much would become angry, so that’s unlikely to happen.

If you’re a China Pessimist, you probably think “it’s bad to give China any of the compute”, but many Chinese people who love China would become angry, so that’s not great either.

So what if we gave China compute to help them ONLY with the Awful People Problem, and also compute to help with the Models Becoming Awful People Problem?

We could assemble some of the people who care about these problems, give them lots of compute in Singapore (a place where neither the Chinese nor the Americans think is Bad) and then we’ll use it for safety work. We’ll block work aimed at making Bigger and Better models so nobody gets tempted. We’ll invite the Chinese labs to use the cluster to safety-test their new models and do other safety work, and we can talk with them about Awful People and Awful Models if they ask us questions.

While we’re at it, we could give Americans compute to help them ONLY with the Awful People and Awful Models problems, too! Researchers like compute a lot, it’s probably their favorite thing. This compute could be controlled by the safety people themselves as a special treat for spending all their days thinking about Awful People and Awful Models.

We could name it something like Peaceful Compute.

https://pacificcompute.org/