What’s Next After ChatGPT?
So much ink has been dedicated to raving about ChatGPT that I have temporarily shelved my idea to write a post explaining what it is and what it does (and doesn’t). I still think such an article is useful, because most of the material I see misses the real point, which is how to incorporate language models like ChatGPT into business operations to make them better. Stay tuned on that front.
“My advice to you all is to look closely and critically at the wake created by step change developments in technology like ChatGPT.”
For now, I wanted to talk about another subject in that domain that gets very little attention—the “what comes next” regarding this breakthrough technology. Frankly that is where the most interesting and useful developments occur.
Machine-assisted human flight in the early 1900s happened with a fairly unremarkable device doing a single remarkable thing. The decade afterward witnessed the development of military and commercial air travel—much more exciting. The Apple Newton was a rare product failure for the company, but eventually gave us the smartphone 15 years later. So my advice to you all is to look closely and critically at the wake created by step change developments in technology like ChatGPT. Side note: ChatGPT was years in the making, enabled by the technology platform of GPT, started in 2017. Not the “overnight success” that some have indicated.
We are 2.5 months into the release of this most recent version of ChatGPT, so it is far too early to assess the knock-on effects. However it is not too early to speculate on what that might be, so that we can ready ourselves to capitalize on these developments.
“It’s longer what you know that counts, it is what you can curate into a machine.”
Before we get there let’s anchor ourselves in what ChatGPT actually is and is not. First and foremost it is a language model, which means that it accepts input and direction from a user to create understandable prose about a given subject. Oddly it doesn’t know what it is telling you.
For example, you can ask it to write a beautiful love letter to your significant other, include a few details about that person, and it will generate text that will practically bring you to tears. So much so that it tricks you into thinking that it understands the concepts of love and relationships. It does not. It was simply trained to understand what a love letter is (by looking at billions of them) and its structure and has just enough intelligence to string words together that make for a sensible story. That is directly opposite of what a knowledge system is, which has been trained specifically to respond to prompts where the prompt is digested and understood and then mapped to its knowledge network. That doesn’t make language models v. knowledge systems a competition, but I wanted you all to know that the two might seem the same in look, feel, and function but are not.
Stephen Wolfram wrote two wonderful posts here and here comparing Wolfram Alpha (probably the best public Knowledge System out there) to ChatGPT and I encourage you to check it out.
I see several things on the horizon. Check with me in a few years and see if I am right:
1. Convergence of knowledge systems and language models
This is a marriage whose time has come as Stephen Wolfram has beautifully described. Not only should we be able to create excellent prose, but we must combine that with a genuine understanding of facts and concepts that precipitated it. When these two AI developments merge…watch out. Another step change.
2. Domain-specific AI mutating off the public model
As a human I not only want to know about things in the public domain like scientific facts and historical events, but also about private domains like my company. I would love to ask my AI the following: “What are my earnings likely to be next year?” “Should I merge with competitor X and if so why?” “What are some potential risks to my strategy?”
OpenAI created ChatGPT by training it on a broad cross section of internet content. What if those same tools and methods could be unleashed on every piece of corporate data to develop a universal economic understanding of the company including deep analysis of its strategy, operations, and finances? At Business Laboratory we’ve actually done Proof-of-Concept (PoC) work around this idea, and I can report to you that such systems are within reach today.
3. Preemptive AI
ChatGPT begins with a prompt. Therefore the user must be skilled enough to form a prompt that will give them the answer they intended. That puts a great deal of burden on the user to master the tricky art and science of prompting ChatGPT.
What if instead the system formulated its own prompts and submitted those at just the right moment, perhaps anticipating what the user should know? We call this preemptive AI. Expect to see developments in this arena in the coming years.
Summary
Any new groundbreaking technology creates winners and losers (steam engine manufacturers = winners, Luddites = losers). The skill that wins the coming decade is the curation of human knowledge into machines, which in turn requires a non-trivial extraction process followed by a data upload process. Medicine, law, engineering, business strategy, public systems—all of these are fields that have substantial amounts of human-based and text-based knowledge but precious little computable knowledge. It’s no longer what you know that counts, it is what you can curate into a machine. Get ready. It is going to be a wild ride.
