Re-posted from my newsletter. (Yes, I’m back doing newsletters. Subscribe for the latest — I’ll be writing more there.)
GPT-3 has been heralded as the next frontier in AI writing. It’s capable of “writing” op-eds, songs, Elon Musk poems in the style of Dr. Seuss, and mimic Kafka’s “The Metamorphosis” or any other literary authors. I put “writing” in quotes, because the quality of output is still largely dependent on human prompts and human editing.
OpenAI’s GPT-3 Playground, the off-the-shelf demo that introduced this to the masses, is pretty darn good. It’s perfectly capable of generating a five-paragraph essay responding to a typical middle or high-school writing prompt. It’s also pretty damn fun to let your imagination run wild. Everything highlighted in green below is generated by GPT-3; no editing was involved.
GPT-3 is just a slice of the broader wave of “generative AI” that can produce paragraphs, poetry, stories, art, video, code and apps, on par with or exceeding human capacity and imagination. Sequoia Capital summarizes current developments nicely and has a handy market map. There’s generative AI for interior design and life coaches. Chances are you’ve already come across something created with GPT-3 or DALL-E but didn’t realize it.
As far as where investors’ attention and checks are focused, GPT-3 is the new Web3. Except this time, the use cases are much more tangible and easy use. Jasper.ai has taken the most recent spotlight courtesy of a 2022-defying $125 million Series A round. It’s one of many tools tailored towards helping marketers generate blog posts, social media posts, sales emails, ads and other content marketing. This is by far the most crowded GPT-3 category, with everyone describing their value with some variation of “Write copy that converts 10X faster!”
I suppose it shouldn’t be terribly surprising that one of the first popular uses for a new technology trained on all the web’s content is to generate stuff optimized for clicks. Personally it feels like a low-hanging fruit and, as a former journalist, is probably the least exciting use case (it feels like writing clickbait headlines). On the flip side, there are other tools like Lex and Sudowrite tailored more for literature, longform and other pursuits in the creative and liberal arts.
What I find more fascinating is how this will change not just how we teach and assess writing, but also force us to rethink how we evaluate skills and capabilities.
GPT-3 and Education
GPT-3 made headlines shortly after its initial release in June 2020, and it is finally emerging as a topic of debate in education outlets. But it’s a couple steps behind. An Education Week article, published this August, claimed that “GPT-3 is not used in any regular K-12 classrooms” right now.
Newsflash! While it may well be the case that it is not a school-approved technology, students are definitely catching on. One doesn’t need to search deep on Reddit or TikTok to find them sharing tips on using GPT-3 to write assignments or essays. Last month, a post on the OpenAI Subreddit went viral, about a student boasting about getting straight A’s and making money by doing his classmate’s homework. The tech-savvy ones are even training their personalized GPT-3 AI to produce things in their voice.
As someone who took shortcuts in school himself, I find this delightfully subversive. Let’s acknowledge that it takes some ingenuity and creativity to do this! More fundamentally, it raises big questions about how schools and students approach writing.
Will this change how we teach writing?
In previous generations of writing AI, humans were the writer, and AI the editor. But GPT-3 flips the script: AI produces copy, and humans are now the editors. GPT-3-generated copy is very clean; there are hardly any diction, syntax and other grammar issues, unless the writer wants to tweak them.
I don’t think this diminishes the value of teaching fundamental writing mechanics, just as calculators don’t obviate the value of teaching computation. But GPT-3 can ensure that mechanics are not a barrier to creative coherent output. Writing companies are also incorporating GPT-3 to turbo-charge the personalization of writing assignments, assessments and feedback. Quill, for instance, is building custom AI models on top of GPT-3 to provide customized feedback for students on individual writing prompts.
Can you be a good editor without being a good writer?
It’s a chicken-or-egg question that a couple paragraphs won’t do justice. My instincts say no. For now I’ll say this: One of the most basic responsibilities of any editor is to check facts, which is the one area where GPT-3’s limitations are noticeably apparent: inaccuracies are rife the more niche or esoteric the prompt.
This will place even greater value on critical thinking and media and digital literacy; after all people need to understand how the web works to effectively (and responsibly) use tools trained on the web. Also, getting good GPT-3 output is largely dependent on how one writes the prompt, which is not unlike how the quality of Google search results depends on what keywords we use. That’s an invaluable skill in and of itself.
Will this make writing more fun?
Many people (including writers *ahem*) don’t like writing. It’s like the “I’m not a math person” thing.
As a kid, I loved writing because I was encouraged to let my imagination run amok, without rhyme or reason. As we get older, writing tends to become more structured in ways designed to support a logical and coherent train of thought. Organization is important, of course. But structure can come at the expense of creativity. And depending on how it is enforced, it can suck the joy out of writing. I personally hate the five-paragraph essay.
If GPT-3 can provide structure and let kids focus on ideas, perhaps they will feel more free and empowered to propose bold things outside the box. In this spirit, one novel application comes from Koalluh, which is leveraging GPT-3 and DALL-E to encourage kids to create their own illustrated stories.
As an adult, I find it absurdly delightful to feed whatever random thoughts are circling in my head and see what it puts out (like the giraffe VC poem above). Sometimes it’s trash, sometimes it’s amazing. Just like Forrest Gump’s box of chocolates.
How will the education system respond?
GPT-3 will be a major flashpoint and understandably so. How should teachers feel about written assignments largely done by AI? How will admissions officers evaluate college application essays? Will every kid have equal access to these tools?
Based on precedence set by plagiarism checkers and proctoring software, a natural inclination would be to call this cheating and develop tools to specifically prohibit its use. Build a sort of Turnitin for GPT-3. (We can imagine how kids will feel about that.)
That would be disappointingly reactionary and short-sighted. Banning things, as we should well know by now, often has the opposite effect. It will also further drive a wedge between students, teachers and the education system, and widen the gap between what kids learn and what they’ll need to know later. It’s a little hypocritical and counter to the mission of preparing students for life and work.
Cheating is a social construct established in relation to rules. Rules change over time, and so should our models for defining terms like “academic integrity.” The optimist in me hopes GPT-3 will push us all to take a hard look at what we assign, how we assess, and why we’re doing it. What is the value of asking questions that can be answered by AI?
These questions are just the tip of the iceberg. There are many more and each deserve their own posts. If you’re thinking about and/or building in this space, please reach out. I’d love to chat!
Et Cetera
- This post, written by one of the creators of Lex, captures well the possibilities and limitations of GPT-3. “Writing is thinking. That’s it. And if we don’t forget that, I believe the future of writing is going to be very exciting indeed.” (h/t to Jomayra)
- I got a new role at Reach! More on what “Head of Platform” meansand also meet some of my colleagues.
- We also welcomed the newest colleague to the Reach team: Steve Kupfer. He’s our venture partner and, perhaps, my new golf partner.