A brief diversion from where we were. One of the most fun uses of ChatGPT is having it create new works in the style of other authors or imaginary conversations among people. Here I decided to see how it would do recreating an episode of Tyler Cowen’s podcast Conversations with Tyler featuring Vaclav Smil — but I didn’t give it any of that context.
Let’s see how it does!
Not a bad start, though usually Tyler’s first question is more direct.
I’m only noticing this after many re-reads– but I’m surprised ChatGPT makes the mistake of saying “a interesting” instead of “an”… given that the tools is trained to select the most likely word to follow another word, it seems unlikely there are many cases of “interesting” following “a”. Has anyone else noticed agreement issues like this?
Anyway, moving on:
It gave a few more responses in the series, and as you’ll see in a second, I didn’t think this was particularly similar to Tyler’s style of question or Vaclav’s style of answers. But we’ll come back to that.
I wanted to test how much ChatGPT “remembers” about the conversation it’s in the middle of.
Ok, so it does “understand” in some sense who is speaking in which order. To help it out on the question side, I gave it a few examples of the way I think Tyler asks his questions.
(This was not a totally random question! While Vaclav is maybe most known for his work on energy technologies, in his 2012 book he explores the Japanese diet transition, including how it relates to climate: https://mitpress.mit.edu/9780262017824/japans-dietary-transition-and-its-impacts/ Unfortunately, ChatGPT’s version of this answer seems to not know much about the book.)
Ok, we didn’t like what ChatGPT gave us, can we make it better? How well does ChatGPT do in responding to feedback?
Here it just copied my prompts but made Vaclav’s responses more concise, which I told it wasn’t really what I wanted.
Oh well, let’s see what it can do on its own:
ChatGPT got the format much closer here, but the content is mostly there, but maybe not the way Vaclav would frame it.
Vaclav seems to agree that subsidies and energy efficiency are underrated and that carbon capture is overrated. (See, for example this NYT interview, notably published in 2022– after ChatGPT’s training cutoff, meaning that ChatGPT has not seen this interview.) I don’t have any definitive sources, but I think Vaclav might be more nuanced on solar (my guess is he’d say that on it’s own, solar might be appropriately rated, but as part of a system which requires a significant change in the power grid to deliver it, he might say it’s overrated). And on EVs I think Vaclav would take exactly the opposite position.
ChatGPT is trained on a wide set of data, and by prompting it to focus on the work of Cowen and Smil, I hoped it could bias its answers a little more towards what they would have said in the past. Still, it’s not hard to imagine what the next version of the tool could do.
And now, to close out the conversation like Tyler often does — a step too far for ChatGPT:
