Are LLMs 'useful'?
Image attribution: MB&F, CC BY-SA 4.0, via Wikimedia Commons
Here's a thought experiment. Say I give you a clock and I tell you it's stopped. Is that useful? Can you use it to tell time?
You'd say no, right? Even though, there are two times over the course of the day when the clock is indeed right. A stopped clock doesn't tell time even if it is 'right' twice a day. Which is to say, being right isn't the same thing as being useful.
Here's a thought experiment. Say I give you a clock that tells time correctly 50% of the time. The other 50% of the time, it's 5:04. Is that useful? Can you use it to tell time?
Surely yes, right? You just look at it over and over again until it's not showing 5:04, and there's the time. If it actually is 5:04, well, that's an edge case, but you can handle that by just waiting a minute. It's not as good as a properly functioning clock, but it is a decent approximation of one.
Here's a thought experiment. Say I give you a clock that tells time correctly 50% of the time. The other 50% of the time, it shows you any other time at random. Is that useful? Can you use it to tell time?
Well, yes, if you think about it, right? You can take a sample of, say, 10 or 20 readings and you know the matching ones are right and the random ones are wrong.
Here's a thought experiment. Say I give you a clock that tells time correctly 80% of the time. The other 20% of the time, it's wrong, but not random; it just unpredictably runs fast or slow over the course of the day, such that it may or may not actually be lined up with the actual standard time, even though readings from it look coherent in isolation. Is that useful? Can you use it to tell time?
Well, not quite, but almost, surely? You can tell time 'with 80% certainty.' That's probably good enough for most applications. You could figure out what the margin of error is and just live with it. Most of us don't really plot out our movements throughout the day down to the minute, right? Our precision in arriving places is more like down to the quarter hour or so.
Here's a thought experiment. Say I give you a clock being operated by a minute, malicious demon. It knows that if it just showed you the wrong time all the time, you'd figure out that it's a broken clock and chuck it. It tells time correctly, oh, about 80% of the time - but you can't know the exact odds, and really they will change based on circumstance. It almost always gives you a time that is plausible. It mostly won't try to insist that it's 5:07 when the sun is directly overhead, for example. But it will, sometimes, be wrong; sometimes significantly. You can't know when or how often.
Is that useful. Can you use it to tell time.
I would argue that no, it isn't; it's not giving you any actual information about the time, not even the probabilistic information that the 80% clock was giving you. If you found yourself in a situation where telling time is actually difficult – say you're scuba diving, or at a very high latitude – the clock could be leading you astray by hours.
But: it is very easy to act like the clock is useful, isn't it? I mean, yeah, sometimes you're late to things. But usually it's nothing too important. Does it, ultimately, really matter what time it is? The clock tells you a time, and that, in a sense, solves the problem you had where you didn't know what time it was.
Is that enough?