Pete Huang

So if you suddenly had a million PhD level AI models that could just do more AI research, doesn't that pretty much guarantee that you will have ASI very shortly afterwards?

Okay, let's take a breather for a second.

We were just talking about GPT-2 and how it was no smarter than a preschooler just a few moments ago, and now we're talking about Skynet?

Like, geez, that's a bit of a jump, isn't it?

The specific arguments behind it kind of go like this.

If you want to say, let's take the same improvement from GPT-2 to GPT-4 and see what happens if we continue,

then you first have to know how much actual improvement there was in GPT-2 to GPT-4, right?

Leopold does this by measuring how much computing went into training these models.

Again, remember that GPT-2 to GPT-4 was simply a matter of more training with the same general model architecture.

Leopold breaks it down into three buckets.

Bucket number one is just how big the computers are that you use to train.

Bigger and more efficient computers are more powerful and therefore you put more training in these models.

Big machine is better, basically.

The second bucket is how you train those models, the algorithms that you run, the methods that you use.

For example, you could train a model in a different way such that the whole thing runs faster and is more efficient.

Then the third bucket is sort of a miscellaneous bucket of rather small changes that actually translate into very big wins.

For example, people discovered that asking Chat2pt to think step by step through a problem before actually solving it improved its accuracy.

Just a small little change, but that actually resulted in a big win in its reasoning capability.

Leopold's estimate is that from GPT-2 to GPT-4, the models got somewhere like six to eight orders of magnitude more training.

And each order of magnitude is 10 times.