To A Man With A Hammer
Don't overlook the importance of the chatbot interface
Have you thought about how much of the perception of ChatGPT is a consequence of its chatbot interface?
There’s no reason that the GPT large language model that powers it under the hood needed this particular interface. But if we’re not paying close attention it’s easy to miss the specific attributes implied by the chatbot interface. We need to pay attention not only to the nuts and bolts of how a technology works but also how we as humans interact with it. In the case of chat-based generative AI, it’s important not only to understand the basic concepts of how a large language model or generative pretrained transformer works, but also think carefully about how the way we interact with it will shape our understanding of it. Analyzing the nature of our interaction with AI tools is another important part of building our AI literacy. While AI is discussed widely enough for a basic level of literacy to be valuable for everyone, I see a particular need for educators to be well-informed about this before we jump into classes this fall.
The Wizard of ChatGPT
ChatGPT reminds me of The Wizard from The Wizard of Oz. Just like The Wizard—who after all is just a normal guy from Omaha, Nebraska—the large language model that powers AI tools like ChatGPT, BingAI, or Google’s Bard is really just putting on a show. I’m certainly not going to throw too much shade at what the AI tools can do; I’ve used them enough to see that they can be very useful and helpful in the right context. But seeing them as anything other than a dressed-up algorithm is misleading and ultimately unhelpful.
Regardless of the underlying technology, ChatGPT demonstrates how important it is to clothe a technology in an effective user interface. Just like the iPhone was effectively a better interface to technology that already existed before its invention in the Palm Pilot, the combination of the GPT large language model and chatbot interface of ChatGPT finally helped these types of large language models achieve widespread use.
The interface matters a lot. If the UI for ChatGPT wasn’t as accessible as having a chat with your friend over iMessage, I’d bet that we’d be seeing far less hype about it. The tool is valuable, but make sure that you are careful to understand what parts of the chatbot interface are helpful and what parts are giving you false impressions.
The medium of the chatbot
The analogy to The Wizard of Oz also reminds me of two of my favorite modern philosophers, Marshall McLuhan and Neil Postman. Neil Postman’s book Amusing Ourselves to Death was my introduction to Marshal McLuhan’s now iconic phrase: “The medium is the message.” Those five words have loomed large in my imagination ever since, giving me a lens that I can’t help but look through when I think about a specific technology and its broader influence. Postman writes about the shift from a text-based to a television-based medium:
As typography once dictated the style of conducting politics, religion, business, education, law and other important social matters, television now takes command. In courtrooms, classrooms, operating rooms, board rooms, churches and even airplanes, Americans no longer talk to each other, they entertain each other. They do not exchange ideas; they exchange images. They do not argue with propositions; they argue with good looks, celebrities and commercials. For the message of television as metaphor is not only that all the world is a stage but that the stage is located in Las Vegas, Nevada.
Peering through a set of McLuhan lenses is a worthwhile exercise. Think even about three different ways to have a conversation with one of your friends: text message, voice call, video call, or in-person chat over coffee. Even if you try to communicate the same message through all three mediums, the message will be unavoidably shaped by the method of communication you choose.
Approaching ChatGPT through this lens helps to draw the connection between it and The Wizard even more directly. The chatbot interface of ChatGPT is a specific communication medium. The chatbot interface presents ChatGPT as an illusion of a conversation. You start by prompting it with a question or request just like you would ask your friend or teacher. Then, for a short while, the cursor blinks on the screen presenting you with another illusion of a familiar activity: pausing in thought before responding. However, in this case, the computer is really just performing some computations behind the scenes. After the processing is done, ChatGPT starts to spit out the result of its calculation. But it doesn’t give it to you all at once like a search engine query or a function call in a section of code. It parcels it out one word at a time.
While I don’t have any inside knowledge about the specific design decisions in the ChatGPT interface, I would hazard a guess that these specific interactions are specifically chosen to hijack the wiring of our brain circuits to give us the impression that we are speaking to another human.
In many ways, the chatbot interface is one of the most challenging obstacles to overcome in helping people to understand large language models and their strengths and weaknesses. Yes, the booming influence of AI tools like ChatGPT is about the new GPT architecture that powers it. But it’s also strongly influenced by the medium of the chatbot.
The medium of the chatbot gives us the illusion of a conversation with an intelligent being. The slowly evolving text on the screen, rolling out word by word, only serves to strengthen this charade, giving us the impression that there is spontaneous thought occurring when in reality there is only iterative computation to calculate the next most probable word in the sequence.
Use your hammer wisely
When you’re holding a hammer, everything looks like a nail. It’s a trap that’s easy to fall into as the user of a new shiny technology. You hold a new type of hammer in your hands and proceed to search the world looking for nails. Even a screw seems nail-like enough that you’ll warrant a try.
So it is with generative AI text generators. See a blank page? Turn your magic machine loose on it. A recommendation letter or performance review that could be meaningful to another human but can just seem like a chore in the moment? Magic machine again.
I’m not arguing that AI text generators aren’t useful. I think they can be. But not for every application we can think of. If we just blindly adopt them and begin to use them without discretion we will be just as foolish as if we were happily walking around mindlessly trying to nail in screws.
Got a thought to add to the conversation or think I’m missing something? Leave a comment with a question or thought!
The Book Nook
After listening to a podcast interview with Josh Linkner last week, I’ve really been enjoying his book Big Little Breakthroughs. His thought process is very resonant with a prototyping mindset, exploring how iterative small experiments and prototypes help to unlock big breakthroughs.
I’m only a few chapters in, but Josh has the book divided into two parts. In the first part, he talks about some of the fundamental principles. Then in part two he walks through what he calls “The Eight Obsessions of Everday Innovators.” I’m sure that I’ll be reflecting on some more takeaways in future essays.
The Professor Is In
A few weeks ago I shared how I was prototyping some new mentoring ideas for my research students this summer using individual development plans. After they completed their plans I met with each of them one-on-one for a half hour to discuss the plans and talk about anything else that was on their mind.
This first round of chats went great and I really enjoyed getting to know the students better in an intentional conversation. This week will be two weeks since that initial conversation and I’m looking forward to following up on how things are going and what they are learning.
Leisure Line
In case you were under the false impression that my pizza-making operation was without failures…I’m about to disabuse you of that notion! Of course, after I wrote last week arguing that we all need to embrace failure to get better, I had an adventure with one of my pies this weekend. While that first one was pretty rough, the last two photos show two of the more successful pies. And boy were they good. Chicken and roasted red pepper is a winner. And the freshly shredded mozz is still a game changer.
Still Life
More food pictures. On Saturday morning I made a Dutch Baby from this New York Times Cooking recipe. Super easy to make (basically blend everything together and pour it in the pan) and pretty yummy. I’m sure a little powdered sugar on top would have made it even better.
This is why I focused my class around using tools like SoduWrite that integrate into the writing process. The future of AI writing isn't chatbots. They are useful, but the real writing happens in tandem with AI and that doesn't really happen in a chatbot.
This is also why I take a structured approach. I never talk to ChatGPT like a person, and I think it handles structured content better ... well, so do humans, but we usually ignore that.
100% valid. People think "Clippy" when they should be thinking "Hal."