On #robot judges: part 7 - an intermission,¹ and a confession.

In part 6,² I suggested that we can never be sure whether hallucinations have been eliminated from a Large Language Model (#LLM).

In this part, let's take a step back and go back to where we all began.

---

I started this series intending to address a column in The Business Times, in which the author suggests that there are "good reasons" to "embrace legally binding and enforceable AI-generated reasoned determinations".³

This statement, and the entire article, struck me as a little... odd.

After all, one can say that there are "good reasons" for almost anything. But surely that's not enough for a nuanced discussion? Shouldn't we also consider the downsides, as well as any good reasons against, if we are interested in serious discourse?

But rather than merely criticising on the sidelines, I decided to put myself to the test, and to see if I could properly articulate my concerns in writing.

In my initial draft for my 1st post, I went straight into addressing the author's propositions.

But then I realised that in order to address the propositions, I would first need to challenge my own understanding and assumptions as to how LLMs work.

And I figured that the way to keep myself intellectually honest would be to set out, in full detail, my thought processes and logic as to how LLMs work, and to open myself up to brickbats if I get it wrong.

So, in that sense, the previous 6 parts have been a primer, if you will, as to how (I understand) LLMs to work.

And I must confess: it was only in the writing process that I came to realise the mistaken beliefs, and gaps in my knowledge, which existed before I embarked on this series.

---

In the next few part(s),⁴ I'll finally get down to addressing the author's propositions, and why I disagree (or agree!) with them.

For those of you who have been journeying with me so far, thank you for hanging in there. It's going to get spicier from here out.