← IndexLarge Language WriterAbout
June 2024 • design-led research, AI • Made for : Design Investigations

Large Language Writer

→ How will we write in the future? Designing honest interfaces for Large Language Models.

a person sitting in a large beautiful building writing on a device
© Fritz Enzo Kargl

Interfaces

The interfaces between humans and machines play a crucial role in shaping perceptions of technology. Current interfaces deployed by major tech firms often lean towards mystification and concealment, presenting technology as if it operates like magic. This top-down approach can breed scepticism and distrust, categorising these interfaces as "Dishonest Interfaces." This project explores what it means to design "honest" interfaces.

Picture of Andres Sergovia playing the spanish guitar Picture of a control interface designed by Ettore Sottsass

Dishonest Interface

I've noticed a concerning trend in design, especially in interface design, that emerges in "fully developed" products like baking ovens, stoves, and car interiors. Take the baking oven: its practical features are optimized, but companies still need to drive sales. Instead of relying on planned obsolescence, which could tarnish their reputations, they create demand by adding features promoted through advertising. This explains the rise of WiFi in home appliances.[1] Traditional controls are replaced by terms like "4D-Heat" or "Fish Mode." While corporate greed is a simple explanation, the nature of design itself may offer deeper insights.

Contemporary technologies of networked computational things and artificial intelligence, as well as the data capitalism they have made possible, differ from the logic of industrial production. Not only that, they fundamentally challenge the conceptual space designers have created to cope with complexity. For instance, with runtime assembly of networked services, constant atomic updates, and agile development processes, the boundary between produc- tion and consumption is almost fully dismantled.

Elisa Giaccardi, Johan Redström, Technology and More-Than-Human Design, DesignIssues: Volume 36, Number 4 Autumn 2020

Image of a modern car interior showing only a screen and a steering wheel Image of a modern baking oven interface showing pictures on its screen

Societal Impact

In our technocratic era dominated by Big Tech, the prevalence of "Dishonest Interfaces" in modern consumer products has fueled growing skepticism towards research and technology, often spiraling into conspiracy theories. For instance, a simple warning light in today's digital cars, which once indicated an easy fix, now typically demands an expensive workshop visit, creating a sense of opacity and loss of user control. This dynamic breeds frustration and mistrust, making users feel governed by their technology rather than empowered by it, thereby perpetuating skepticism towards progress.

Graph showing low interest in science in Austria
Decreased interest in science in Austria.

Works like magic

We observe these tendencies across various design domains. The phrase "it just works like magic" might appear to be a guiding principle, and while those deploying it may genuinely believe in its effectiveness, it merely exemplifies the problem highlighted by Elisa Giaccardi and Johan Redström, where intentional obscuration becomes the preferred guiding principle, rather than addressing the underlying issues.

"Clarke's Third Law: Any sufficiently advanced technology is indistinguishable from magic."

Arthur C. Clarke, "Hazards of Prophecy: The Failure of Imagination“, Profiles of the Future (revised edition, 1973)

Screenshot of apple introducing the iPhone. Backdrop reads "works like magic".
iPhone introduction 2007

Deus ex machina

The "Deus ex machina," a figure in Greek theatre lowered by a visible crane to resolve impossible dilemmas, offers a compelling analogy. The crane's visibility allowed the audience to trust the magic without feeling deceived. In design, we should "expose the crane" by making the mechanisms behind technology visible, fostering trust and transparency, and enabling users to engage without feeling manipulated.

Sketch showing the mechanism of deus ex machina

Honest Design

With all these insights, ideas, and challenges, the question remains: What does it take to design honestly? It seems that honest design, achieved through seamful[2] interactions, could be a cornerstone for fostering a more positive and confident perspective on technological development.

Good design is honest.

Dieter Rams, 10 principles of design

design needs to be anticipatory, able to craft desirable relations between people and emerging technologies, and thus proactive in the associated processes of research and development.

Elisa Giaccardi, Johan Redström, Technology and More-Than-Human Design, DesignIssues: Volume 36, Number 4 Autumn 2020

I set out to design a prototype of an "honest interface," along with hardware components that adhere to the same philosophy. The objective was to develop methodologies for creating such an interface, using the specific example of interfacing with Large Language Models. Without further ado, let's delve into the five principles of honest design.

Sketches showing the design-process of a AI writing device

01 Place a Spotlight

AI is currently a buzzword, and as its tools become more widespread, significant questions about their impact arise. While generative AI in image creation remains somewhat niche, systems like ChatGPT raise broader concerns. As students use AI for homework and others rely on it for drafting documents, the value of writing as a creative practice and a human method of documentation is increasingly being questioned.

Screenshot of ChatGPTs interface Screenshot of Llamas interface

The future of the Interface

While Big Tech favors chat-based interfaces, we should remain curious about alternative methods. Beyond the outdated command-and-control approach lies the Centaur Approach, theorized by Garri Kasparow, where humans and machines collaborate, each contributing their strengths. In 2021, I explored this by having children and generative AI co-design toys, demonstrating how the Centaur Approach can integrate humans into the generative process.

Mosaic depicting a centaur
© Mary Harrsch / Flickr

02 Break down the tech

When considering "honest" tech, we encounter a paradox: while interfaces like ChatGPT obscure their workings, the most transparent interaction might be Ishan Anand's "Spreadsheets is all you need", a fully functional GPT-2 in Excel. Though transparent, it lacks usability. As a designer, I wondered if there's a middle ground, a sweet spot between transparency and usability. To explore this, I examined three essential ideas…

Timelapse of Spreadsheets are all you need
Spreadsheets are all you need

Tokens and Relations

In a Generative Pretrained Transformer (GPT), tokens are the smallest text units, like words or subwords. The model learns relationships between tokens by analyzing large text datasets. The transformer's attention mechanism weighs the importance of each token in context, capturing patterns, syntax, and semantics. This allows the model to generate coherent and contextually relevant sequences by predicting the next token based on learned relationships.

Dataset

The dataset used to train an AI model shapes the quality, scope, and biases of its generated text. A diverse and extensive dataset enables more accurate and creative outputs, but it also embeds any inherent biases. After training, the model cannot learn from new data, meaning these limitations are fixed. However, a model can be fine-tuned on additional datasets to adapt or update its knowledge, allowing for some post-training adjustments.

Probability

In AI text generation, probability drives word prediction, with the model selecting the most likely next word based on context. This often results in text that feels "in the middle," as it favors safe, generic choices to maintain coherence. Interestingly, even improbable ideas like time travel or aliens can emerge when the model calculates them as the most probable among unlikely options, giving the illusion of creativity. However, the overall focus on high-probability words often makes AI text sound neutral and predictable.


For deep-divers I recommend LLM Visualization by Brendan Bycroft


03 Open up loops

In the conventional approach, the process begins with a prompt and ends with a result, which can be refined through iteration and prompt engineering. However, this method largely removes the human element from the generative part of the process. This project introduces an alternative approach that advocates for a collaborative writing mode inspired by Generative Pre-trained Transformers (GPTs). By emphasizing mutual understanding and fostering "honest" interaction with AI, this approach actively involves humans in the generative process, placing them directly at its core. During the writing process, which occurs within a dynamic feedback loop, three key mechanisms are integrated: "Dataset," "Emphasis," and "Probability."

Sketch showing circles and the words "Emphasis", "Dataset" and "Probability"

The Main Loop

Simply put, LLW introduces a new interactive feedback loop. The process starts with a prompt, followed by the AI extracting a theme and generating an initial sentence. The loop then continuously offers continuation sentences based on the user’s settings. We will delve into the interaction modes (Emphasis, Dataset, Probability) shortly. Once a sentence is selected, it becomes the new "last sentence," and the loop repeats until the user concludes the writing process. This approach immerses the writer directly in the generative process.

Large Language Writer Screenshot: Main Loop

Emphasis → steer the feedback-loop

By incorporating AI's core principles into the user interface, LLW fosters an intrinsic understanding, leading to a trustworthy and honest relationship between users and their tools. The Emphasis strategy, based on token relationships, exemplifies this approach. In Emphasis mode, users can select words or strings by holding Shift, then assign a weight (1 to 3) before exiting the mode. The generated continuation sentences will reflect these user defined weights.

Close up of a pcb with a rotary encoder
© Fritz Enzo Kargl
Large Language Writer Screenshot: Emphasis

Dataset → Inform the loop

We have to recognize the fundamental differences between human and more-than human intelligences and embrace them. The non-obfuscation of this reality is an important step in designing honest for AI. Human traits like inspiring moments or things, a cute present they recently got or their current surroundings can be fed into the dataset to inspire the generative process in unconventional ways. LLW is equipped with a camera-module that allow writers to capture whatever they imagine to flow into the writing.

A person is using the LLWs camera A person is sitting in a room writing on the LLW

Using the LLWs camera

When the user activates the dataset mode, they are prompted to take a picture using the LLW camera. There are no creative limits on the choice of subject. For instance, if something reminds the user of a particular moment, if they are in an inspiring environment while writing, or if they see something exciting, all these things can be captured with the camera and added to the LLW's "dataset." Even a word or a sketch can be photographed.

A person is lying on the floor, holding the Large Language Writer's camera up in the air.
© Fritz Enzo Kargl

Image recognition

In the background, the image is analyzed using AI image recognition, and a list of selectable keywords is generated. The user is always shown the complete image description beforehand to ensure that the derivation of the tags is logically integrated into the workflow.

Large Language Writer Screenshot: Dataset, Step1
Large Language Writer Screenshot: Dataset, Step

Probability → Break the loop

An essential problem identified by implementing feedback-loop based interactions into user-interfaces is the nature of the loop itself. To brake it is one of the main challenges. Up until know we leveraged the idea of „probable continuation“ in order to break up loops and let them continue somewhere else. Instead of predicting the most probable next word or sentence as AI does naturally, on the LLW keyboard students can use the probability button to ask the AI to give the least probable continuation of a story. This helps student to understand the nature of the technology and collaborate with it to playfully expand their personal perspective.

Large Language Writer Screenshot: Probability

04 Design Honest Hardware

AI is a fast changing landscape with new developments everyday. To honestly represent this ever-changing technology and give it a materiality we have to acknowledge that a design poured into the clear boundaries of a cast unibody, no longer proves sustainable. We therefore propose exploration into modular hardware like voxel-based modular design that is able to gracefully react to its obsolescence.

The Large Language Writer sits on a table
© Fritz Enzo Kargl

Modular throughout

The first prototype of the Large Language Writer consists of a 3D-printed, modular, voxel-based body. This system uses three color codings: red for volume, yellow for connection, and monochrome for function. Using this system, three main assemblies were created: 1. Display Module: This module houses a 2K e-Ink display encased in custom-cut, powder-coated aluminum sheet-metal parts, along with the computer of the LLW. 2. Keyboard Module: This module includes a custom-made circuit board with a hardware design derived from the UI/UX principles declared above. 3. Camera Module: A separate module for the camera.

A person is using the Large Languge Writer A person is using the Large Language Writer

Keayboard layout

We decided to design and manufacture a circuit board that adheres to the sizes of the underlying grid. This circuit board includes three function keys, a Shift key and three Mode keys: "Emphasis," "Dataset," and "Probability," which toggles the modes described above. There is also a "Write" key, comparable to a Return key. The cursor is controlled using a rotary encoder. A Pro-Micro, which sits on the underside, runs QMK. The keyboard can be connected to any computer via USB-C.

Close up of the Keyboard of the Large Language Writer
© Fritz Enzo Kargl
Close up of the Large Language Writer's screen
© Fritz Enzo Kargl

05 Involve the Real World

After developing the prototype and running the first semi-stable version of the software, five individuals from diverse backgrounds—each with a strong connection to writing, whether out of necessity or creativity—were invited to participate. They received a brief tutorial on operating the machine and were allowed to choose their own writing prompts. The participants wrote for 45 minutes to an hour and were subsequently interviewed. This initial testing phase yielded valuable insights, informing potential directions for the continuation of the project.

Philipp

Philipp, a 16-year-old upper secondary school student, contributed to the test by writing a letter to the editor. While the content of his letter wasn't focused on AI, it provides insight into how students in his age group are engaging with AI in educational contexts.

You are reading the newspaper report 'Nur keine Spompanadeln' by Michael Omasta from the weekly newspaper Falter dated June 22, 2016, and respond with a letter to the editor.

Initial Prompt

Flora

Flora, who studied law and is currently working in legislation, offers an important perspective to the tests. Questions of responsibility, accuracy, and contextual awareness, particularly in relation to AI, can be explored. Flora’s insights help to underscore the need for careful consideration of these factors when developing laws and guidelines for emerging technologies. Her contribution adds insights regarding legal and ethical implications.

A lawyer advises her client based on AI-generated legal research. Describe possible issues in her work using a case example.

Initial Prompt

Helmut

Helmut, a 56-year-old author, contributes a seasoned perspective to the test. His experience as a writer brings a unique viewpoint on how language, storytelling, and perhaps even AI intersect. While the content of his contribution is not focused on AI, his background as an author enriches these tests by offering insights into how creative professionals engage with evolving technologies.

Newspaper article about low-threshold free cultural offerings in public spaces in Vienna.

Initial Prompt

Flora

Flora, an art history student, brings a crucial focus on factual awareness to the testing. Her academic background underscores the importance of precision and context in interpreting information. This perspective is vital in discussions about AI, where accuracy and contextual integrity are essential. Flora’s contribution drew parallels between her field and the challenges AI faces in maintaining these standards.

My thoughts on Otto Wagner's design for the Peace Palace in The Hague, 1905-1906.

Initial Prompt

Yucheng

Yucheng, a young lower secondary school student, belongs to a generation where AI is increasingly part of daily life. His experiences highlight how early exposure to AI influences learning and interaction with technology. Yucheng's participation offers a glimpse into the growing role of AI in education and its impact on younger students.

Inner Monologue: Thoughts of Janine Phew! That was close, he almost caught me...

Initial Prompt

A person sitting on the carpet using the Large Language Writer
© Fritz Enzo Kargl

Credits

Supervisor: Univ. Prof. Anab Jain & Team from Design Investigations Studio
Project Lead: Leo Mühlfeld
Design: Leo Mühlfeld, Lucy Li
Hardware: Leo Mühlfeld
Software Development: Alan Schiegl
PCB Layout: Elias Mack
Pictures: Fritz Enzo Kargl
Operator: Mia Tešić

Special thanks to: Ursula Gschlacht & Team from the University Library, Max Kure, Florian Sapp, Stefan Schönauer and Viktor Windisch.

Image of guitar: Poran111, Flickr
Image of Olivetti ELEA 9003: Olivetti, Wikimedia
Image of Car-Interior: Leo Nguyen, Wikimedia
Image of Baking Oven: Gorenje
iPhone presentation video-still: John Schroter, Youtube
ChatGPT screenshot: ChatGPT by OpenAI
Llama screenshot: Llama by Meta
Centaur mosaic: Mary Harrsch, Flickr
Spreadsheets are all you need screencap: Youtube


  1. Admittedly, a practical application could be to activate these devices during periods of surplus in the power grid, which would require networking them. ↩︎

  2. Upol Ehsan, Q. Vera Liao, Samir Passi, Mark O. Riedl, Hal Daume III published research on Seamful XAI: Seamful XAI: Operationalizing Seamful Design in Explainable AI. ↩︎