Editor’s note: this post was written in September 2022 and obviously a lot has changed since then. I may write about the current state of AI soon, in the meantime please enjoy this blast from the recent past. —DD
This week I saw someone on Reddit posted about using OpenAI’s GPT-3 natural language AI model to generate outlines, and even first drafts, of schoolwork. So far I’ve been a bit of luddite when it comes to AI. Images from tools like DALL-E and Midjourney looked cool but weird, I didn’t know how people were even playing with these models, and sentences like “GPT-3 can write a novel” didn’t sound threatening so much as unrealistic. Um, no? Computers can’t do that?
But something about this schoolwork-via-robot scheme finally piqued my interest, so I signed up for an OpenAI account and started messing around.
Neil Gaiman wrote a great line in Neverwhere where he describes characters as wearing “the kind of suits that might have been made by a tailor two hundred years ago who had had a modern suit described to him but had never actually seen one.” For me that describes much of these AI models’ output: technically accurate in some respects, but wrong in ways that mark them as uncanny.
This is especially true of AI image generators like Midjourney. Midjourney V4 seems to prefer pictures of people to landscapes — if you ask it for a scene with people in it, people will dominate the composition — and it’s terrible at animals. Like all models Midjourney also has a “knowledge cutoff,” limiting its understanding of pop culture and current events, and because its sense of semantics is text-based it simply doesn’t understand memes or visual humor.
For example, prompts referring to “David S. Pumpkins” get images that are sort of on the right track—high foreheads, pale skin, curly black hair, as if it understands that this person shares attributes with Tom Hanks, but can’t quite make the leap that he is Tom Hanks unless I say so. And it can usually understand that “David S. Pumpkins” is a man with pumpkins on his suit, but often takes that to absurd places (and sometimes goes pumpkin-y in the wrong places, like the face.)
Coming back to DaVinci and GPT-3, I had asked it to write posts about video game consoles, about explaining football to my third grader, about using football as an analogy to explain the Electoral College to my third grader.
Eventually, I wondered if DaVinci could convincingly write a blog post about what it means to be a good product manager in 2022. I gave it some prompts, cleaned up the output, then re-ran everything again and again to flesh out the details. The result was a story I posted to my Medium profile this morning:
In the world of technology, product management has always been about features. The race to add the most features and get them to market quickly has been the name of the game. But as we move into the future, this mindset is no longer going to cut it.
To be successful, product managers need to start thinking about the bigger picture. They need to be focused on creating products that solve real problems for people. They need to become customer-obsessed.
A lot of business-related content on Medium and other platforms is regurgitated wisdom, similar to what DaVinci and other models generate, it’s just that the regurgitation is done entirely by humans. These posts yield claps and follows, which in turn leads them to be recommended to new readers, who write posts in a similar vein, and the cycle continues. Here, the regurgitation is being done mostly by a computer — the algorithm eating itself.
The result looks a lot like stuff I see all the time on Medium and LinkedIn, so I expect that it may get more engagement than if I’d written what I really think about data, roadmaps, and being customer-obsessed. (Honestly, this whole digression into AI—which I included in the original post, so friends don’t think I’ve been replaced by a pod person—will probably hurt my chances of becoming a product management thought leader. Le sigh.)
Having tried to get a complete, good text out of DaVinci, I’m skeptical that models like this can create anything interesting or inspiring. But they’re surprisingly good at summarization and structure. For example, if a writer (like me) is struggling to write and work from an outline, prompting an AI with Outline for a blog post comparing the following webcams is not a bad place to start.
But how much generated text is too much, and when does it matter?
That Reddit post about AI-written schoolwork raises an obvious ethical question and a less-obvious practical one. Ethically, is it plagiarism to pass off the work of an AI model (which is drawing from seed data written by thousands or millions of other people) as your own?
And, practically, if it’s OK (even expected) to draw ideas from prior art in a paper or essay, and you’re editing the AI-generated text into its final form… is the only difference that you didn’t craft every sentence the whole time?
Even now, having gotten into Midjourney and OpenAI’s Playground tool, there’s a big difference between “writing” or “painting” and whatever these tools do, as evidenced by my fake PM thought leadership post above. The models will only ever be as good as their seed data; Midjourney seems to really like glowing halos, and landscapes that seem realistic at a distance but make no sense when you zoom in.
My fake thought leader post was partly a lark, partly an experiment to see what happens when you give algorithms a taste of their own medicine. I’m still not sure what I think about all of this—generating content with AIs is a fun, sometimes surprising mirror on the collective output of the internet.
But I do know that we need to be careful about the stories we tell about AI, and the way we use AI to tell stories. AI is a powerful tool, but it’s not a magic bullet. We need to be thoughtful about the way we use it, and make sure that we’re using it to create the world we want to live in.
(Sorry, I suck at endings so I had DaVinci write that last bit.)