PBS launched a popular children’s TV show “Ghostwriter” (Chinese translation “Little Ghost and Word Ghost”) from 1992 to 1995. In the play, a group of teenagers, with the help of typewriter ghosts, unraveled the mystery of Brooklyn, New York. Some neighborhood mysteries. This haunted typewriter is haunted by the ghosts of escaped and murdered slaves during the Civil War. There’s a typewriter that happens to be the show’s namesake, also called Ghostwriter, but has nothing to do with anything supernatural, but is the work of Arvind Sanjeev, himself an interaction designer, artist, and Lumen.world CTO. There are ghosts attached to the typewriter in the show, and Sanjeev’s typewriter is based on OpenAI GPT-3 technology.
Sanjeev first demonstrated GPT-3 technology in an artistic form in 2021. “This is a poetic reminder that allows us to calm down and think about the new creative relationship that is gradually forming between humans and machines.” Spent a few weekends with The Ghostwriter typewriter built at night can talk to users through typing, allowing both parties to freely communicate and create together through paper.
Sanjeev told Engadget via email: “I wanted Ghostwriter to evoke a warm feeling that people feel comfortable interacting with a typewriter, which is why I chose the typewriter as a mental model. The typewriter has a taste of nostalgia, the old days. Technology is mostly tangible and is a tool of everyday life.”
“People can trust and feel comfortable using a typewriter because they understand that the sole purpose of the typewriter is to create on paper.” He added: “This is completely different from today’s technology, those black boxes trying to use the ‘attention economy’ to promote unethical business model.”
The original Ghostwriter was an electronic Brother AX-325 antique typewriter with a keyboard matrix for easy coding. Sanjeev chose the GPT-3 model for several reasons. He is familiar with the model as an adjunct assistant professor at the Institute of Interaction Design (CIID) in Copenhagen. Another reason is that he was impressed by the model’s “creative content generation” function. “Because the API is easy to access, I integrated it into Ghostwriter,” he said.
Sanjeev removed most of the device inside the typewriter and replaced it with an Arduino controller and a Raspberry Pi. Arduino can read the user’s input on the Ghostwriter keyboard, and then import the input into OpenAI’s GPT-3 API through the Raspberry on the control board. Then, the AI will use generative magic to generate responses, and Ghostwriter will be responsible for printing these responses on the page for humans to read.
“Ghostwriter slowly typed out responses on paper in a tangible way, word by word, so people couldn’t help chewing and savoring the meaning between the lines in the AI’s response,” Sanjeev said. “Digital interaction on a text editor is so fast, it’s often unintentional to overlook these details.”
One of the biggest challenges of the whole project was actually teaching the system to hit the correct keys in response. Sanjeev first had to decode the matrix of existing electronic keyboards, which is the device that converts the physical pressing of keys into corresponding digital signals. He wrote: “I press each button, read the signal scan line triggered by it, map it to the corresponding button, and finally create a driver that executes on the Arduino based on this.” Users can even use two Tactile knobs to adjust Ghostwriter’s “creativity” and “response speed” parameters, thereby affecting the final answer provided by the AI.
Sadly, there will be only one Ghostwriter typewriter in the world for a while. But Sanjeev is stepping up efforts to open source the project so that other developers around the world can create similar devices of their own. “I’m hoping to free up some time soon to clean up the code and package it all together,” he said.
“The generative AI fad is by no means a fad,” Sanjeev said, though it’s not a “magic bullet” when it comes to creating content. “It’s clear that we’ve crossed a tipping point for AI creativity that past AI pioneers thought was impossible,” he continued.
These tools can not only help us outline our ideas, but even inspire entirely new concepts, but in the end they should only be tools for our creativity, not substitutes. “AI is a magic pen that allows creators to do something brilliant,” Sanjeev said. “Whether it’s built with AI or not, it’s a masterpiece with humanity and life.”
The final application of GPT depends on the medium of presentation. It is an active tool for creating digital content, but more often it is a “concept database for inspiration” for creators. “The key to unlocking the potential of ChatGPT in the maker base is building useful physical interfaces,” he said. “Artists or creatives who use AI will play the role of curators in the true sense, carefully selecting from AI works and entering the next stage of the design process.”
He also expects similar synergies for knowledge workers. In the past few months when ChatGPT became popular, the automatic text generation system has been the focus of close attention from the media and the industry. The technology has shown itself to be good at not only writing Linux code, composing haikus, passing Wharton Business School admissions exams, and explaining finance on CNET. Knowledge workers such as lawyers, business analysts and journalists are naturally concerned that such automated systems will one day take their jobs, as BuzzFeed recently brought AI into its newsroom.
However, Sanjeev believes that the role of AI should be less prominent, and gradually develop from general creative use to specialization in specific areas of knowledge. “Just as cloud computing is ubiquitous today and supports most applications, AI will also retreat into the background of our lives after the hot period, becoming a ubiquitous element,” Sanjeev believes.
The revolution unleashed by AI should ease the burden of this type of work, automating many of the tedious parts of the job. “Because of AI’s ability to collaboratively organize vast amounts of niche data in specific domains such as software engineering, law, and business, it is being used to train hyperspecialized AI in individual domains,” Sanjeev said.
OpenAI itself also provides custom training kits in its system, making it easier for customers to accelerate the establishment of their own personalized AI doctors and robot lawyers. However, in the event of an accident—for example, an AI doctor misdiagnoses, or an AI lawyer makes a mistake that is enough to disqualify a lawyer—who should bear the responsibility in the end is still a difficult issue before everyone’s eyes.
original:The Ghostwriter typewriter brings generative AI to the printed page, by Andrew Tarantola