Azeem Azhar

But if you try to get the LLMs to write anything, even a network of LLMs, what you get is incredibly mid-copy.

It doesn't have that interiority.

It doesn't have that idiosyncrasy.

It doesn't have the details that come out from being a human who's lived through something and has experienced things and knows when to break the rules and when not to break the rules.

And there's some empirical evidence around this.

The Arc AGI General Intelligence Benchmark, the Arc AGI-2, showed that pure LLMs score 0% on tasks requiring genuine fluid intelligence.

They're good

at being able to work their way through certain classes of crystallized problems.

But I think that means that right now they're not very good at doing the thing that makes the writing good.

I say right now because the LLMs, of course, aren't just LLMs now.

There's lots of other things going on in these systems to make them perform and make them be much, much more reliable.

So we don't know if that frontier doesn't get crossed.

What I have learned, of course, is that they can identify golden threads and they can analyze writing styles, but they can't write good quality paragraphs.

And one of the challenges, of course, if you are someone who is delivering what I'm doing now, which is like an expensive, authentic experience, I'm spending an hour of my time, you are all graciously giving me an hour of your time.

We have to work with LLMs.

We have to work with the best tools available.

Otherwise, we just can't cope with the demands of the work.

But it's that hard human work, the work with your pen.

Where's my pen gone?

Here it is.