The following chart of Go decision quality over time has been making the rounds. In short, the development of AlphaGo and its domination in the Go world has seemingly catalysed an increase in the decision quality for Go players worldwide:
Or maybe you’ve seen this video of a dog “stuck” behind an open sliding glass door which needs to be let into its own house.
Perhaps you’re into business and have read this HBR article on what the 4 minute mile taught us about limits:
I’m not sure how much this concept of feeling limited until someone else shows us something is possible generalises, or if it’s even been researched. How many invisible ceilings are there really in the world, and could we predict them ahead of time? Even without those questions answered, the examples above are enough of an intuition pump to consider what other invisible ceilings might be blown away by AGI and/or LLMs. Let’s try one.
In the improv comedy world, I’ve heard that LLMs are too logical and unfeeling to write jokes or improvise scenes that would leave audiences cry-laughing. Puns are easy for a machine, but “real” jokes are an art.
Sure, the few times I’ve tried to craft jokes with ChatGPT, they would’ve been rejected even by the US government’s dad joke repository. But that’s true for mine too, especially that one I tried to curate for my colleague at LEGO (ask me over a coffee 😬). And on the capable side, a quick news search yields plenty of articles about how funny ChatGPT can be. We've got a longer history of actual humans being funny too. So this just nets out at a wide distribution of joke quality for us and the machines. This isn't a great test for ceilings; let's try another one.
What about stand-up routines? How many of the stand-up acts out there were crafted by or with a LLM? And if some were, would we even know? I’m assuming that audiences crave authenticity. I was thinking of communicating this point using the Overton Window concept but wasn't sure how to translate it, so tried the machine of the hour:
but...
Oh well, let's go with it anyway. So I'm assuming that LLM-generated comedy is outside the Cultural Acceptance Range for most consumers, and would turn them away. This is based on other machine-generated entertainment not quite catching on (e.g. hologram concerts). So even if stand-up comedians are using LLMs, they'd have a pretty strong incentive to keep that quiet. What if LLMs helped stand-ups to breach a ceiling that had been effectively limiting their growth? Would breaching this ceiling lead to downstream impacts via increased consumer demand, or would the number of consumer-hours of comedy watched remain flat, but consumers would be happier and maybe would pay more? As I dive down this rabbit hole, I realize I know very little about comedy success metrics. Time to explore this one in a future, more in-depth post!
Now, what does this mean for improv? I'll be sharing more on this soon...
Questions: What ceilings do you consider to be limitations in your field or hobby? Where do you think you could surpass the best-in-class? Would you want to watch a stand-up comedian who did an entire act based on LLM outputs? What about a stand-up act that was partially developed using a LLM as a tool?
With thanks to: Rothbaum for many years of chess followed by an adventure in learning Go together
