Moving the goal posts
I’m not a deep AI expert, more of an interested observer and consumer. But I do remember fairly clearly that, for many years, the received wisdom was that you’d never be able to do natural language parsing without semantic understanding. Sentences like “Time flies like an arrow, fruit flies like a banana” or “The police arrested the protesters because they feared/advocated violence” are common examples of how you have to understand the meaning of the sentence to be able to correctly identify how the words are being used and how the sentence is structured.
Now, though, we have LLMs that do a great job of parsing sentences like that (try it…I just tried the first sentence above and it correctly called out each interpretation of the second word “flies”). But now the received wisdom (at least in some quarters it seems) is that “just because you can parse natural language doesn't mean you understand it”! It can’t be both ways…can it?
To some degree, this is folks moving the finish line because they don’t like the implications of it. This happens so much that it has a name as a phenomenon: “No True Scotsman”. In this case you can either say “it’s not TRULY parsing natural language perfectly” (debatable, but you can find examples), or “that’s not TRUE understanding”.
That last interests me. I often feel that intractable arguments like this are really, at root, arguments about definitions, and people often enter into them without understanding this. This results in an indirect argument about definitions, using some example as an imperfect proxy. Frustrating!
Think of the “Ship of Theseus” paradox. The idea is that if you rebuild a ship one piece at a time, so that it’s completely replaced eventually, is it the “same” ship? You can debate this endlessly, but what you are really debating is which definition of “same” you have chosen. One might say that “my definition of same is identical in all parts” - the absolutist definition - and in that case, when the first part is replaced, it’s clearly not the same ship. Or you might argue for a more continuous definition - that the ship is 9999/10000th the same after the first replacement (if it had 10K parts), and so on, until it’s a wholly different ship with no trace of the original after the last part is replaced. Or you might even argue that it’s the always same if it’s being treated as the same ship, or even something else.
If you agree to any of these definitions at the start, the entire argument becomes…boring. It’s just a matter of terms now. The real argument is when people don’t agree on the definition.
But definitions are only useful if they are being used for something. Is the ship being insured? Used again? Is ownership being settled? Then we need a firm definition of “same”. Are we just standing around debating (and writing blog posts)? Not so much.
The same is true with the idea of “understanding” in the first example. That’s a very broad term. It’s not likely to ever be rigidly defined in all contexts. We should be precise about what definition we are using when we make statements like the ones at the top. And it’s possible, like the examples from the Ship of Theseus, that there are some nuanced or continuous definitions of the word “understanding” that could be more useful.
The point of all of this is - it’s not that helpful to argue over different versions of words with ambiguous meanings. Far better to focus on practical use, or, if definitions are really needed for a practical reason, have the debate about them directly, relative to the context in which they need to be used, not as indirect, general arguments that are virtually impossible to resolve. If you find yourself making a “No True Scotsman” statement, it might be the case that you’re moving the goalposts and should step back and think about definitions more clearly.