Recent advances in AI have taken the world by storm. OpenAI’s release of ChatGPT introduced the world to Large Language Models, which do a great job of mimicking interaction with people. Ask it a question, and it seems to “get” you, and generally gives a pretty reasonable reply. Give it a task, such as creating a table in Excel, or writing some code in python, or even editing a document, and it seems to understand and do it for you, almost effortlessly.
This is such a huge advance over the state of the art just a year ago, that it’s difficult to imagine our lives before.
But it’s missing something fundamental, something so simple we take it for granted. It’s missing the ability to make decisions.
Large language models may appear to make decisions because they choose what to say. But in reality, they are just picking the most likely series of words. They’re a bit like a reflex reaction. No actual thought is going into the choice.
This shows up in some important, but subtle ways. For example, while the answer you get may read very well, it can also be utterly wrong. It doesn’t know what’s right or wrong. All it does is predict a sequence of words that seem likely in response to your prompt. Those words only make sense if it deems a sequence which makes sense to be more likely. But it has no way of knowing what “making sense” even means. To the model, it is simply a sequence of words. As several attorneys have discovered, preparing a brief with references to case law can lead to disastrous results when the model provides a seemingly logical brief—complete with references which turn out to be utter fiction. It turns out that judges expect actual research, and are not amused by fictitious references.
Similarly, these models work in a prompt-reply, prompt-reply form. Here’s a sequence you might see:
What color is the sky?
The sky appears blue during the day due to the scattering of sunlight by the Earth’s atmosphere. This scattering, known as Rayleigh scattering, causes shorter blue wavelengths to be scattered in all directions, making the sky look blue. At sunrise and sunset, the sky can take on shades of red, orange, pink, and purple due to the longer path the sunlight takes through the atmosphere, scattering shorter wavelengths and allowing longer wavelengths to dominate.
Here’s a sequence you will not see:
Hi! How are you today?
I’m doing well, thank you! How about you? Is there anything specific you’d like to discuss or work on today?
Did I lose you? You haven’t replied in several minutes.
Hello?
Hello? Do we have a bad connection? Shall I call you back?
The model simply will not generate a conversation. It requires a prompt, and will respond to that prompt. It will not initiate conversation on its own.
Decisions
A decision is a choice among alternatives, based upon some criteria and information, with a goal of achieving some value.
Human beings make decisions all the time. Our ability to consciously choose our actions is one of our key differentiators from other animals. But some animals also make decisions—or certainly seem to do so.
We make decisions in order to achieve a goal, or fulfill a value. We often lack complete certainty about the situation, but still must make a determination of the best path forward—the best decision.
We need three things to make a decision:
- Something we care about (values),
- Two or more actions which are available to us (alternatives), and
- Something we believe about the world, as relates to the decision and how our action might affect our ability to achieve our values (information).
Decision professionals call this the Decision Basis.
Since the only value a Large Language Model is attempting to fulfill is the sequence of words that maximize a scoring system, that is the only decision it makes. It lacks any higher-level values, and any information about how best to achieve those values.
In essence, a Large Language Model is a super fancy user interface, without the underlying system, or “brain.” It can respond reflexively to a prompt, but will never develop new medical treatments, or advance our knowledge of physics, or proactively reach out to someone in need. To begin to develop systems that can do more, we need to put the ghost into the machine.
Enter the Decision Engine
We need a decision engine behind the Large Language Model interface. That engine can drive higher-level behaviors, to achieve higher-level goals.
The decision engine must be able to accept data, input through whatever input systems the machine has, absorb it, and form a map of the broader information about the world—not just a set of vectors describing the likelihood of sequences of words.
The engine must be capable of identifying and generating alternative courses of action.
And of course, the decision engine must have some core, measurable values it is trying to achieve. Some of these may be innate, provided when the decision engine is created, while others may be learned over time, or adaptable to circumstances.
Finally, the decision engine needs to be able to estimate the results of the various courses of action against the value measures.
With this decision engine, the model can begin to create and take actions, using the Large Language Model as a user interface, and the Decision Engine as the higher-level guidance on what to do. The sequence of words it chooses must then be driven by its likelihood to achieve the higher level values.
By using a Decision Engine as the “thinking” part of the AI, and a Large Language Model as the interface, we can create an AI which is orders of magnitude smarter and more useful than the current generation. Then we can begin to finally realize the true potential of AI, with the resulting advances in fields like medicine, physics, climate change, energy, and more.
The possibilities are endless, and exciting.
