How to judge the speed of improvements in AI
This post was first published on my newsletter, Delightful
A fun thing that happens when you get excited is you think the future will be more exciting.
When you get sad the same the same thing happens: You think the future will also be sad.
Your prediction of your future emotional state is affected by your current emotional state.
Which is why giddy couples who just got married think, five years from now, they’ll be just as in flagrante delicto, and hungover people see nothing but a grey horizon of couches and overpriced Uber Eats.
This is called projection bias. We look at tomorrow and render it with the affect of today.
The future is a hologram of now.
Another fun thing that happens is that we overestimate the impact of technology in the near term and underestimate its effects in the months and years after that.
This is called Amara’s Law.
When the CEO of Ford said in 2015 that self-driving cars would be a thing in five years, but they weren't; and when Bezos said in 2014 that drone delivery was coming soon, but it didn't; and when observers said in 2003 the Human Genome Project would quickly end cancer, and then biotech slumped for like a decade and a half: That was Amara’s Law in action.
We were all very excited, and tomorrow looked exciting.
Alas.
To borrow a phrase, no optimism extends with any certainty beyond the first encounter with complexity.1
All that said, it’s difficult, in this moment, to think Amara’s Law applies to artificial intelligence.
That is, it’s difficult to suggest we’re underestimating the impact of AI in the near term, and overestimating its effects in the months and years after that.
I mean, finger count the news: there’s that new thing that, given a three-second clip, can simulate your entire voice, and that other new thing that generates complex songs from simple text prompts, and that other other new thing that can make it look like you’re always making eye contact. Not to mention GPT passed the bar, that was cute.
Kinda feels like we’re not underestimating anything! The internet feels like a peculiar sideview mirror. Objects are closer than they appear.
But it’s worth pausing to note what, exactly, is closer.
Because the speed of the things advancing in the sideview mirror depends on what exact things you’re talking about.
When people talk about how AI is changing creativity, one fun thing to ask is what do you mean, exactly, by “creativity”?
Do you mean auto-generating images to capture someone’s attention during a scroll, or in a purchase funnel, or for a twitter bot, or an online magazine?
Where the image or the idea needs only be “good enough"?
Or do you mean crafting a commercial idea that satisfies a client, and the client’s lawyer, and can be distributed on a broadcast network to try and affect, if only for a moment, the culture and the discourse? Like say a Super Bowl commercial.2
Because one of those looks similar to the business model of the underpants gnomes:
And the other one looks something like this:
And that’s not everything!
I’m just tired of typing the list!
You’ve still got whole big things like receiving the RFP, then translating the RFP, the pitch process, the close process, getting coffee, backsliding on cigarettes, working until 3am for weeks straight, ruining your marriage, all that.
One, because I’ve never been married.
Two, because software that uses machine learning to aid design choices or create complete designs already exists.
Netflix has used computer vision to determine which frame of a video is most suitable as for key art. They score each frame based on a character’s expression in that frame to narrow down choices for its creative team. Alibaba’s Luban auto-generates billions of image thumbnails and learns which are most effective.
And in the advertising world, there are already services like Pencil and Bannerbear, which create complete designs and visuals with only a few inputs, and tools like Seenapse, which help you develop ideas.
Obviously there are differences among these examples.
The former examples are automation at the industrial asset scale.
The latter examples are automation at the design commodity scale.
But they’re similar in that they’re both discrete parts of a creative workflow. Not the whole workflow. Part of the workflow. The parts of the workflow that deal with standardized units. Not the parts that are wildly bespoke. Which means the products fit a precise need (we need more thumbnails, we need more banners ads) within a precisely defined workflow.
Yes, more tools are coming, like Fantasia brooms, to sweep away increasingly complex tasks.
But some of them are going to take a minute, Mickey.
Somebody told me once, I can’t remember who3, that anything that can be described can be automated.
We are now in the process of describing everything, and much more quickly than ever before.4
But it’s possible, just possible, we’re currently overestimating how quickly creative AI tools with make a difference, and underestimating where and how (and btw already, see examples previous) our more prosaic AI tools will make an impact.5
And one way to think about that—to think about what AI can and will automate next—is to consider two reductive characteristics.
Where the factors affecting complexity are stuff like e.g.
Not to mention all the traditional characteristics of complex systems, like randomness, local rules, nonlinearity, hierarchy, and emergence.
This action/system complexity gives rise to a b-school obvious 2x2, where the simplest task in the simplest part of the system (lower left quad) would be rapidly generating low-stakes ideas that are "good enough” for a transactional audience. The activities involved are things that are easily explained.
In the top right quad, you’d have diligently workshopping an idea based on conflicting feedback and then passionately selling the revisions to a fearful corporation so you can ship the result to an equally quirky production studio. The activities involved are not as easily explained, and accordingly the difficulty level is FML.
I’d make that 2x2 myself but my Apple Pencil is broken (also FML), so please welcome to the stage, my friend and collaborator, Midjourney:
I like letters. So does Tita. She's my dog. She opens all the mail.
