Azeem Azhar

There's a really interesting idea in there, which starts with the fact that LLMs don't get addicted to dopamine.

And so you can put this material in there.

And one of my experiences of using the different chatbots is that they are incredibly reasonable.

I mean, you can get them to be unreasonable.

You can push them in different directions, but they are this aggregate of human knowledge and their training and the reinforcement learning to get them to behave in particular ways.

But it's quite hard to get them to be extreme by the time they reach us.

Now, there's a question as to whether that will always be the case and whether as the companies get more and more powerful, whether the owners won't start to tune them in one direction or another.

But there is also this really important point that you've made, which is we can start to identify more clearly where there are gaps and opportunities.

And one of the things that struck me when I do deep research queries is that Gemini or ChatGPT or Claude will go out and look at 500, 600 websites

And I would have asked a really simple query that needs deep research.

And I'm thinking, an academic, a human wouldn't have gone to 600 websites to figure out when the bicycle boom in Peoria was and what the impacts were.

You'd go to two or three good sources.

So there's also this question that,

these machines are rather greedy and they're a bit dumb, frankly, about how they look at the content that's out there.

So you described the Swiss cheese model, the sense of filling in the gaps, but one of the things to make markets work, including information markets, is to put in a little bit of scarcity, a little bit of the price mechanism that says you have to pay more for this because it's better and this is free because it's slop.

So how do you think about that?

There are some models.

I mean, financial information is a very good example.

You can't get real-time data from the CBOE or the New York Stock Exchange without paying a lot.

You can't get Platts oil data.