Speculation about ChatGPT/Bing/LLMs

speculating on a question from a reader; Marc Andreessen on a podcast

This post has two parts. First, a reader asks,

"what economic factors should I study to understand whether LLM trainers will keep brute forcing increasingly huge models, or start focusing on smaller, specialized models?" 

It is early to speculate, but I cannot resist. Let me start by describing two paths that large language models like ChatGPT could take.

First, there is the General Purpose path. It leads in the direction of Artificial General Intelligence, where the models keep acquiring more and more knowledge and surpass humans in their capabilities. Along the way, the use cases are substituting for humans as writers and analysts.

Second, there is the Simulator path. It leads in the direction of providing training and education. We create an effective coach for helping young people learn. And we create an expert in, say, trust and estate law relevant to residents of Maryland.

The Simulator path also leads to creative artificial interactions. I could have a conversation with my late father. Or we could watch a simulated debate between Keynes and Hayek about today’s economy. Or a simulated basketball game in which Wilt Chamberlain and LeBron James are on the court at the same time.

As of now, the General Purpose path seems to have the attention of many researchers and pundits. But I think it could turn out to be unsatisfying and a dead end. I am skeptical that large language models can be more intelligent than the data with which they work, which consists of samples of human writing. We are going to see that the behavior of the General Purpose AI’s based on large language models will be highly contingent on the people who train the models.

I think of model trainers as steering the way that language models respond to prompts. Think of the trainers as seeding the models with artificial data.

Once upon a time, economists figured out a way to hack a linear regression program to do Bayesian estimation, in which the results are a sort of average between what the data says and your prior belief. For example, let’s say that your “prior” is that for every $100 of additional income people get, they spend $80. In addition to putting actual monthly data on household income and spending into the computer, you seed it with some made-up data that is consistent with your prior. Insert a data point with monthly income of $5000 and spending of $4500 and another data point with monthly income of $5100 and spending of $4580, so that the marginal spending is $80 out of $100 in marginal income. The more made-up data that you give the computer, the higher the weight on your “prior.”

Similarly, large language models produce output that is a combination of the writing it sees in the data and the prior beliefs of the trainers. As users come to see the clear influence of prior beliefs, they will tend to view these models as puppets of their trainers rather than objective sources of information.

If I am correct that the Simulation path will prove more productive than the General Purpose path, then the balance will tilt toward specialized models. That will make it less likely that the resulting industry structure will consist of one or two giant corporations protected by a “moat” consisting of an enormous data set that is difficult to replicate. It will make it more likely that we will see profits from domain expertise and to skills in generating multimedia simulations.

Marc Andreessen and others

Self-recommending. I don’t think anything Marc says makes me want to take back anything I said above. Here are a couple of excerpts:

we have moved so quickly to sort of rule these hallucinations into the category of bug or flaw um and I'm on the exact opposite page like I think that I think that the hallucinations themselves are an incredible breakthrough. . .this sort of Magic Machine comes along that like is making stuff up all over the place and we just have this like reflexive knee-jerk reaction that there's something wrong with that I think there's something amazing about that which is we have apparently like as part of this we have apparently solved the problem of computer creativity

and

every kid is going to grow up now with a friend right that's a bot um and that bot is going to be with them their whole lives um and it's going to know every it's going to have memories gonna know all their private prior conversations it's going to know everything about them um it's going to be able to answer any question um it's going to be able to explain anything it's going to be able to teach you anything um it's going to have infinite patience um and you know for like as close as a machine can get to loving you like it's going to love you

Neal Stephenson’s Young Lady’s Illustrated Primer.