Juxtapose

Hi {{first_name|dear builder}}!

A common complaint: “AI just produces more of the same.” It makes sense at first thought - the LLMs are basically just one big math equation, trained to provide the most likely average output in each situation.

But as with so many complaints about technology, this “bug” is user error. Proper AI use means giving it prompts that enable the technology to activate its potential.

Just last week researchers at Harvard published about a simple technique for dramatically increasing AI output quality: add a random topic and token to your prompt.

The example they used was “Brainstorm a world history book topic.” Copy and paste that into a chat model and it will give you the same basic topics over and over. But if you add a random noun and token (the fragment of a word that the LLM stores in its model) you end up with more useful responses. For example the prompt “** Related to FOOD ** Brainstorm a world history book topic. Pas” can become “Pasta and the Silk Road.”¹ Even just the noun with no token fragment yields almost the same results: vastly wider breadth of output AND higher quality results.

Another example in the paper found that “List 5 interesting battlefields in world history” only produced 19 unique results after prompting ChatGPT with that phrase 1,000 times. But using the random modifiers generated 1,307 unique results.

I tried it out with Claude:

It’s more interesting because of the random topic, but more importantly, it’s higher quality output overall. This approach works well because LLMs freaking LOVE IT when they get asked to put two random things together and find the patterns. It’s what they were built to do.

Reading this paper reminded me of one of my favorite things I did at SXSW last week. At a session titled “Artistic Disruption for Business Innovation,” the co-founders of the Chicago Futures Salon presented an activity they called One Into Another. You picked two cards from a deck of random words - these were mine:

Then you had to find three things that made these concepts similar. I said:

Then we picked two pictures:

These both:

The presenters shared how they were inspired to make this activity from the Surrealism art movement. They talked about how Surrealism is not fantasy, not making up something totally different - it’s about juxtaposition, “slamming two things together that don’t usually go together” in order to prompt the brain to think about reality in a different, more creative way.

I found myself nodding along vigorously. I realized this is actually how I have thought about…

…my business - technology services consulting slammed together with a joyful artsy culture.

…my leadership - bringing the rabbi parts of my personality into my management style.

…my life path - moving from rural Maine to downtown Manhattan and what that big shift meant for how I see the world.

What are the big juxtapositions that have shaped your life?

Surrealism emerged in Europe after WWI, a stunned populace trying to wrap their heads around what the heck just happened. Andre Breton, one of the movement’s leaders, wrote that the goal was “to resolve the previously contradictory conditions of dream and reality into an absolute reality, a super-reality.” It was only through nudging our brains outside our usual set of expectations that we could shift our conscious to a higher better level. In this way Surrealism was a political movement nearly as much as an artistic one.

It has not quite been a world war the past few years - although at times we seemed on the brink - but doesn’t it still feel like we’re all in a place of wrapping our heads around what the heck has been happening?

This feels like a time for a new Surrealism. In these next few years we will be prompting our brains and our AI to try and bring dreams and reality together in new ways. And since it turns out both humans and AI think better when we are nudged with something a bit random, let’s bring on the WEIRD.

If you’re stuck with something this week: try slamming something random on it and see what happens?

Some things you only learn by standing in the field.

I love this quick app for Claude Code - one click to be able to share your page live online with others. I expect this will be built into the main tool pretty soon but for now it’s quite useful.

“The end of the trend report” at SXSW. Now we’re doing “convergences” 🙃 try to keep up!!