On #robot judges: part 2.

In part 1,¹ I suggested that a Large Language Model (#LLM) is essentially a very powerful predictive text generator.

Now, if you've played around with an LLM, you would know that that #AI models are able to generate output that often appears useful or partially useful.

But before we use the output, we might² want to know:
a) how the output was generated;
b) whether the output is accurate; and/or
c) what were the sources that the LLM referred to.

I think this boils down to a single broad question:

Is the output from an LLM the result of "reasoning"?

---

I think the answer depends on what we mean by "reasoning".

Let's consider, again, predictive keyboards.³

When a predictive keyboard suggests the next word, it doesn't just throw up a random word. Rather, it suggests the next word based on certain internal rules, likely along these lines:

a) when you use the predictive keyboard for the first time on a new smartphone with factory settings, the keyboard suggests the next word based on its preset rules. Perhaps the app developer had, in earlier versions of the app, collected data as to the most commonly typed word after "Hey", and the app suggests that word; and

b) over time, as you keep using your predictive keyboard on your smartphone, the app will keep track of what are the words that you most commonly type after "Hey". The app will eventually start to suggest these words to you.⁴

So there are reasons for which the predictive keyboard suggests certain words. A particular word is suggested based on the app's internal rules, and the internal rules are the reason that word is suggested.

But.

Are reasons reasoning?

---

Another illustration.

Say you forget to pay your credit card bill. When you check your credit card statement, you see that a late payment fee has been imposed. The reason for the fee being imposed is that in the bank's internal system, there are rules that can be expressed by way of the following pseudo-code⁵:

IF "date" = "15th day of this month"
AND "unpaid balance as of end of last month" > 0
THEN impose late payment fee

Now, there's obviously a reason why a late payment fee has been imposed. But would you say that in the course of imposing the late payment fee, the bank's systems have engaged in reasoning?

Again, are reasons reasoning?

---

I suspect that when we consider whether LLMs "reason", the meaning that we ascribe to the word "reason" goes beyond merely following preset rules.

So while there may be "reasons" why an LLM has produced a certain output in response to a certain prompt...

...I think many of us would be uncomfortable with the conclusion that the LLMs generate output as a result of "reasoning".

But. Is this just a matter of semantics?

Perhaps. Perhaps not.

In part 3, we'll explore the reasons why LLMs generate the output that they do.