Accidentally Funny AI

A few days ago, something I built got to the front page of Hacker News. I took a screenshot when I found out, and , was at #9 on the list:

The app is called YC Remix, and it takes a template sentence and randomly replaces two of the nouns in it. The hope is that this will stimulate more creative ideas for startups by ‘remixing’ YCombinator funded companies and their chosen targets or business models. (Like ‘Uber for X’, but allowing Uber to vary as well, so ‘Y for X’.) What happened on Hacker News was very interesting: most of the readers thought it was satire. The reason is that many of the phrases YC Remix produces are, well, ridiculous. For example, one reader posted this output from YC Remix:

The reader made this comment:

Everyone loves plastic surgery, but how do you organize groups to go and get procedures done?

This got a good laugh from me, and I don’t doubt that is a common reaction. While this might not seem significant, I believe it can teach us some things about Artificial Intelligence and innovation.

The Uncanny Valley

I own a book called Universal Principles of Design. From this book, I learned of a phenomenon called The Uncanny Valley:

Anthropomorphic forms are appealing when they are dissimilar or identical to humans, but unappealing when they are very similar to humans.

You probably can remember a face that looked almost human, but not quite there. These faces often feel creepy.

I have a hypothesis that there is an analogous phenomenon in Natural Language Output, which is when machines produce text or speech meant for a human audience. I called it the Funny Zone: natural language output that is very close to human speech, but not quite there, tends to be humorous. To get visual:

I have more experience with this than just my HN post. I worked previously on a similar project, one with Natural Language Output. In that case, the sentences always took the form ‘Try blending <seed term> with <other noun>.’ We imagined putting in a seed term like ‘Oreo,’ and getting something like ‘Try blending an Oreo with mint.’

But the output was much wackier than we anticipated. Often, those who read the machine’s ideas reacted with uproarious laughter. For example, one of the actual ideas for a new Oreo: ‘Try blending an Oreo with a Greek.’ The guy who read that first just fell apart. More on this in a bit. The Uncanny Valley is interesting, taken alone. But combined with The Funny Zone, it makes me wonder if there is some general rule of human perception that leads to both of them. For some reason, when stimuli get close to, but still distinguishable from, human-level realism, we have interesting emotional reactions to them.

Framing

I posted YC Remix on Hacker News two times. I know this is technically bending the rules, but I couldn’t resist an experiment in framing. The first time I posted it, I gave it a bad title. I simply used the name of the app, YC Remix. That was dumb. I really don’t know what I was thinking. I only got 2 points on HN that time. After a few months, I decided to give it one more try. I used the title ‘Idea generator that uses YC as Data.’ That much more descriptive line made a massive difference.

The point is that framing is important. Frames change the way we interpret content. Frames interact with our perception, and what we perceive is our reality.

This leads me back to Oreos. If you think about it, there is actually a not-so-nonsensical idea hiding in the statement ‘Try blending an Oreo with a Greek.’ It may be possible to create a tasty Greek-yogurt-inspired filling for Oreos. Now, it might not work out, but it’s not immediately laugh inducing. I would be quite curious to taste such a cookie. It might taste good. The lesson here is the the idea generator actually did a substantively good job stimulating creativity, but the framing really got in the way. In other words, the content of the output is not so bad. In essence, it is saying ‘Try using things associated with Greek culture to inspire a new kind of Oreo.’ But the output is framed in a way that leads to a different interpretation, one that is ridiculous.

How to Exit The Funny Zone?

Since I want to build idea generators, I have something of a catch-22 on my hands. An idea generator should produce output that is slightly different from what a human would naturally think, but not so far away as to be uninterpretable nonsense. That lands it exactly in The Funny Zone. So, how can I solve my problem? I think part of it lies in framing.

To get any idea off the ground, you need a critical mass of people behind it. The disconnect between perception and reality has interesting consequences here: sometimes good ideas die because nobody sees their value, and other times products succeed because enough people buy them, even though they seem objectively ridiculous. The market can be fickle, and unpredictable. A few years back, I read an in The Atlantic about the bias against creative ideas. It turns out things that are too radical are hard for people to accept, despite statements to the contrary. At the end of the article, the author gives this advice for easing acceptance of radical new ideas:

In Hollywood, the “high-concept pitch” offers a useful example. Film producers […] have to evaluate hundreds of ideas a year, but can only accept a tiny percentage. To grab their attention, writers often frame original ideas as a fresh combination of existing ideas. “It’s Groundhog Day meets War of the Worlds!” Or “It’s Transformers on the ocean!” In Silicon Valley, where venture capitalists also shift through a surfeit of proposals, the culture of the high-concept pitch is vibrant (Airbnb was once eBay for homes; Uber, Lyft, and Zipcar were all once considered Airbnb for cars; now, people want Uber for everything).

Creative people often bristle at the suggestion that they have to stoop to marketing their ideas. It’s more pleasant to think that one’s brilliance is self-evident and doesn’t require the gloss of sales or the theater of marketing. But whether you’re an academic, screenwriter, or entrepreneur, the difference between a brilliant new idea with bad marketing and a mediocre idea with excellent marketing can be the difference between success and bankruptcy.

Interestingly, it was this exact article that led me to frame YC Remix output in the way I did. Yet, it didn’t work, as evidenced by the reaction on Hacker News. The reason it didn’t work is not because the advice in The Atlantic article is bad, it didn’t work because machine generated ideas were being squeezed into an all-purpose frame.

So, it’s a tricky problem, and I’m afraid I don’t have the complete answer. It seems that getting the right content is only half the battle: even if machines can produce good content, they still need to frame it correctly. Rigid templates, like the ones described above, are the best frames I’ve found yet. But the search is on for much better ones. Perhaps meta-framing is a possible way to sidestep the issue in the meantime. That is, there may be a way to frame the idea generator itself such that people can find the kernels of value behind the humor. If that is possible, it may be the best of both worlds: you know enough about the output to take the good ideas seriously, and you get a good laugh to boot.