Date: January 12, 2024
Author: Bob Royce
Reading Time: 6 min 17 sec

Looking back on last year, I have to say that it was one of the most intellectually stimulating of my career. My top CliftonStrengths are ideation, input, learner, and strategic, so I enjoyed absorbing the flood material on artificial intelligence and the emerging power of large language models. It was particularly interesting to me as it brings together several key threads woven through my life—education, design, and technology—and especially lately—information architecture.

As I’ve started to conduct workshops introducing AI, I found myself referring again and again to certain materials. Below is a collection of some of those that I think might be useful to others.

A Historical Perspective

Learning how important technologies came into being helps me better understand the foundational principles I use to put new things into context. So when I heard that Satya Nadella recommended “The Dream Machine” by M. Mitchell Waldrop, I started listening to it, and it was so good that I ended up buying a hard copy. As the publisher notes, it “tells the story of the birth of the computing revolution through the life of a man who shifted our understanding of what computers were and could be.” The birth of AI is described within this context, and the narrative was both enjoyable and instructive, so I feel like I better understand some of the key intellectual strands that have brought us to where we are today.

How Does Generative AI Work?

While AI advanced on many fronts in 2023, it was clearly large language models, the core technology behind software like ChatGPT, that got the most attention. There are a gazillion resources out there about this new invention, but the three below rise to the top for me:

When dealing with things that are so large and multidimensional that it’s hard to get your head around them, it is sometimes helpful to see how something works in miniature. That is exactly what David Corney has done with his wonderful effort, “A small language model made from matchboxes,” where he builds a machine out of paper and matchboxes that generates text in the spirit of Green Eggs and Ham by Dr Suess.

The Financial Times has an excellent visual explanation of the core software innovation behind this latest AI growth spurt in “Generative AI exists because of the transformer,” which may also help you get more out of Stephen Wolfram’s quite detailed examination, “What is ChatGPT Doing ... and Why Does it Work?” It is long and at times detailed, but it is worth wading through even if you’re not going to follow some of the technical aspects of his explanation.

If you’d prefer a video, Andrej Kaparthy recently posted a very good general audience introduction to LLMs. If you only have time for one video, I’d watch Maggie Appleton explain how to tame the Shoggoth and get practical work from LLMs in, “Squish meets Structure”.

Applying Large Language Models

In terms of applying large language models in everyday work, I’ve really enjoyed the writings of Ethan Mollick, Associate Professor at the Wharton School of the University of Pennsylvania. In “Centaurs and Cyborgs on the Jagged Frontier” he unpacks a well-designed study done with Boston Consulting Group that benchmarked the benefits and risks of integrating generative AI into consulting. In “Now is the time for Grimories” he makes a provocative assertion: “It isn’t data that will unlock AI, it is human expertise.” He then shows how to prompt LLMs to interactively engage with people to help them accomplish a task.

Although he is an academic, he also has a company that employs AI to help them do their work. He describes some of the benefits they gain in “Reshaping the tree: rebuilding organizations for AI” which ends with something of a warning to companies on the fence regarding AI:

“You don’t have time. If the sort of efficiency gains we are seeing from early AI experiments continue, organizations that wait to experiment will fall behind very quickly.”

Cutting through the Hype

One of the things I’ve been asking myself is, “How fast is this all going to unfold?” I surfed the dot-com wave in the ’90s all the way through the tsunami that wiped out the boom in 2001, and you can see many similarities to that era today. But I know that I’m an optimist—so I’m even more susceptible to the tendency people have to overestimate progress in the near term while underestimating its impact in the future. Generative AI looks VERY promising as a tool for accelerating knowledge work. But the devil is in the details when it comes to technology, and if you don’t know where the limitations are, you’re going to flounder and run aground. So, to ground my expectations on what is practical today, I am very thankful for the MLOps Community, a treasure trove of talks by people working to apply generative AI today. Want to know the good, the bad, and the ugly of applied machine learning? This is the place to go.

Just for Fun…

Think you know what GPT-4 can and can’t do? Test your wits with this GPT-4 Capability Forecasting Challenge from Nicholas Carlini at Google. He presents a variety of interesting prompts that pose different intellectual challenges, and you have to guess how likely you think it is that GPT-4 gets it right. I had fun using this as an audience participation activity in a workshop.

It’s Not Only Software that is Accelerating

While software gets most of the attention, a revolution is also happening in the computer hardware world. This is an important part of the story, making it all the more analogous to what happened in the dot-com boom. It wasn’t just Internet software driving the explosion. It was also the advent of the personal computer and the ability to put powerful processors in the hands of everyday people. “The Dream Machine,” mentioned above, tells that story nicely, but to understand where things are today, there may be no better storyteller than Nvidia CEO, Jensen Huang. At the keynote at SIGGRAPH this year, Huang tells the story of how they bet the company by focusing on designing and manufacturing GPUs, graphical processing units optimized for rendering the special effects and animation at the heart of today’s gaming and entertainment world. It turns out that the same kind of architecture is well suited to support the types of calculations required by generative AI. The video is over an hour long, and the whole thing is interesting. However, in the first 20-30 minutes, you’ll get a good overview of the computational side of the new economic flywheel that is starting up centered on advances in neural nets combined with advances in computing architectures.

Where are Things Headed?

We see and hear a lot from Sam Altman but less from the chief scientist of OpenAI, Ilya Sutskever, who helped oust Altman before reversing course. Speculation was his concern was safety related and you can understand why people thought that after watching this lively 12-minute documentary about him filmed between 2016 and 2019. His ongoing role at OpenAI seems in question, but given his contributions to date, he’ll likely continue to play an important role in the future of AI.

As Mustafa Suleyman, co-founder of DeepMind and founder of Pi, tells it, “Generative AI is just a phase. What’s next is interactive AI.” If that’s so, then you can expect to see more tools like Autogen from Microsoft and see more examples of this experiment to build an entire workforce of AI agents to build software as an example of what artificial general intelligence might look like.

Suleyman is not optimistic about the future and has written a book, the Coming Wave, as an urgent warning to prepare. No time for another book? Me neither! Here’s an animated overview of the key points he makes.

For a more positive perspective and a fairly deep dive into how this all works, watch this talk, From Machine Learning to Autonomous Intelligence, by Yann LeCun, Chief AI Scientist at Meta AI. LeCun is a pioneer in deep learning and is quite convinced that LLMs will soon run their course. They may persist for a while, but only as one component of a multifaceted machine composed of multiple models serving different functions and, in particular, know how to plan a course of action based on knowledge of the real world, something well out of reach of LLMs.

What do you think?

We hope you find this collection of resources helpful and would love to hear what you think. How is AI affecting your world? Are you ready to leverage its power? Send us a message on Social Media or connect with us on our Slack Workspace. We’d love to hear your thoughts.

Need Help with a Project?

Exploring how you can implement AI in your organization? Whether you need an IA to help sort through a complex project or are just getting started with a new product or strategy and want to build on a solid foundation, TUG’s experienced Architects, Designers, and Strategists are available to connect with you. Feel free to reach out to us on our Slack Workspace, or if you need to have an extended conversation to see how TUG can help, drop us a note so we can schedule some time with you.

// connect with tug

About the Author

Bob is TUG's co-founder and president. Leveraging digital technology to amplify human intelligence has been a thread throughout his career. From desktop publishing and digital design in the '80s to digital libraries and expert systems in the '90s to the emergence of generative AI today, Bob has worked to ensure technology is deployed in a way that is good for people. Learn more about Bob here.