Caption: Brittney Michele Dunkins in Philadelphia. Recently, I’ve been working on a few ideas about period tech, data storage, AI hardware devices, and the limits of financial services when managing a global life. I’ve been editing audio and video, writing and interviewing and updating my portfolio. There’s a design file I haven’t opened this week.

Alt text: A woman stands in front of a bicycle, a truck and a bus near Rittenhouse Square.

Reading an MIT Technology Review article about large language models (LLM) from a coffee shop, today. About a year ago, I was sharing ChatGPT with a team of content designers I managed in a weekly meeting. We all challenged the model to express tonal nuance in funny ways: “Write a passive aggressive roommate agreement for my puppy.” The exact language is stored somewhere, I’m sure.

Caption: Working on an article about large langauge models at a coffeeshop. Alt text: A woman typing on a laptop in the late morning.

News and academic writing about AI, and generative AI, have been published quickly in the last year. And while the tech may seem completely new, SXSW quotes about “tech super cycles” and my short-lived experiment migrating to new products like Microsoft Edge, are a repeat of conversations about how to explain machine learning.

I’m not an expert, but writing an explanation of new or changing tech for myself is often a way to design and write explanations people can understand. Below are a few notes, questions and terms. Send your feedback, explanations or questions; I’d love to know how designers are thinking about and understanding AI.

A table describing the cumulative proficiency of humans in parallel to large language models, AI.

Notes on training large language models

Large language models are trained with data to perform simple and complex functions. The article characterizes engineers as teachers that provide a function to be completed by connecting data to a model.

Training a model, with reference to OpenAI’s research, describes a model’s learning as a similar process to training any worker. AI in this article is characterized as learning to complete a task or job through training, rather than completing a complex function after an extensive education. For example: learning to write a recipe or learning and then creating a unique recipe with parameters, such as the omission of certain ingredients or a balance of flavors (sweet, salt, sour, bitter and umami).

A parallel to training large language models can be made to the commonly understood hierarchy of education, and, for most people, the required knowledge, skills, analysis—and at the highest level— the ability to produce a unique understanding, idea, object and action.

My questions

Is an engineer a parameter for model training?
What patterns can a large language model find?

Writing, design and ideas by Brittney Michele Dunkins unless cited.

Concept mapping is a content design method to analyze language before developing terms and definitions. I first used this method as a content designer at Meta Platforms, Inc., Facebook UK, Ltd., located in London where I was trained to be a terminology reviewer and later, a term lead. I’ll likely write more about concept mapping in a later post.

In an earlier post, I identified common terms that require standard definitions. The map clarifies the relationship between two identified terms: large language model and data.

A large language model (LLM) is data. Data is a word that the AI industry uses in the context of teaching or training an LLM. The concept map below includes the three types of data of an LLM, publicly available, licensed, and a word or sentence provided by a human or a graphics processing unit (GPU). See source material below.

The types can be understood as distinct limits. I’ll continue exploring the third type of data, or the language used to describe a token, prompt, neural network architecture, a statistical language model, and a device, including a computing device or a hardware device.

A note that the colors chosen for this post, HEX #F4F5E, #10B320, #6EB5FC, and #000000, are a reference to the colors used in the source material. Concept maps designed using Figma.

Concept map: Large language model data