Pete Huang

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.

So what he's saying is 1 million to 100 million times more.

Then he's looking forward and estimating from GPT-4 today onwards, we're gonna get another five to six orders of magnitude.

So that's another 100,000 times to 1 million times improvement.

And parts of that are somewhat believable, right?

Like Microsoft is already saying quite publicly that they helped build an absolutely massive computer to help OpenAI train GPT-5.

And people are already calling for even more computing resources, even larger,

computing clusters to train these models.

Ultimately, Leopold looks at all this and comes to the conclusion saying, look, the train isn't stopping here.

I mean, each of these three buckets still has lots of room to run.

But even for people who deal with the real world, like plumbers, a super intelligent AI could definitely on its own figure out how to make robots that can replace plumbers.

Okay, so the consequences are massive if you can get there.

So with all of this as the backdrop around AGI and ASI and the consequences, with such dramatic outcomes, people's instincts immediately tend to be like, okay, sure, that's definitely happening.

I mean, how many times have people tried to sell others on some big lofty dream like this, right?

It's such a far-reaching sci-fi world that it can be hard to picture as reality.

Which brings us to some of the debate at hand about this set of essays that Leopold Aschenbrenner has written.