Comment by TeMPOraL
2 years ago
Of course we can. Reliability is a spectrum, not a binary state. You can push it up however high you like, and stop somewhere between "we don't care about error rate this low" and "error rate is so low it's unlikely to show in practice".
It's not like this is a new concept. There are plenty of algorithms we've been using for decades that are only statistically correct. A perfect example of this is efficient primality testing, which is probabilistic in nature[0], but you can easily make the probability of error as small as "unlikely to happen before heat death of the universe".
--
[0] - https://en.wikipedia.org/wiki/Primality_test#Probabilistic_t...
There are two problems with this comparison. First, probabilistic prime generation has a mathematically proven lower bound that improves with iteration. There is no comparably robust tuning parameter with an LLM. You can use a different model, you can use a bigger variant of the same model, etc., but these all have empirically determined and contextually sensitive reliability levels that are not otherwise tunable. Second, the prime generation function will always give you an integer, and never an apple, or a bicycle, or a phantasm. LLMs regurgitate and hallucinate, which means that a simple error rate is not the only metric that matters. One must also consider how egregiously wrong and even nonsensical the errors can be.
I think the better statement is that, if, say, you're running the Miller-Rabin test 10 times, you can be confident that an error in one test is uncorrelated with an error in the next test, so it's easy to dial up the accuracy as close to 1 as desired. Whereas with an LLM, correlated errors seem much more likely; if it failed three times parsing the same piece of data, I would have no confidence that the 4th-10th times would have the same accuracy rate as on a fresh piece of data. LLMs seem much more like the Fermat primality test, except that their "Carmichael numbers" are a lot more common.
I compare LLMs to a door with a slot where you put a piece of paper with a request on it and you get something back related to that request. If it's the same every time, great. But it might be different or completely wrong. You don't know what goes on behind the door and measuring the error rate tells you little predictive.
The general point is not that the feature currently exists to dial down the LLM parse error rate, it’s that the abstract argument “we can’t use LLMs because they aren't perfect” isn’t a realistic argument in the first place. You’re probably reading this on hardware that _probably_ shows you the correct text most all of the time but isn’t guaranteed to.
There's no such thing as a perfectly-watertight roof, therefore there's no qualitative difference between fixing the roof and buying a bigger bucket.
Precisely this. People dismiss utility of LLMs because they don't give 100% reliability, without considering the basic facts that:
- LLMs != ChatGPT interface, they don't need to be run in isolation, nor do they need to do everything end-to-end.
- There are no 100% reliable systems - neither technological nor social. Voltages fluctuate, radiation flips bit, humans confabulate just as much if not worse than LLMs, etc.
- We create reliability from unreliable systems.
LLMs aren't some magic unreliability pixie dust that makes everything they touch beyond repair. They're just another system with bounded reliability, and can be worked into larger systems just like anything else, and total reliability can be improved through this.
EDIT: In fact, my example with probabilistic primality tests is bad because those tests are too nice - they let us compute tight bounds on the error rate in advance. LLMs are not like that. But then, a lot of systems we rely in our daily lives also have this property - their reliability is established empirically, i.e. we improve them until they work reliably enough, and then we hope they'll keep on working, and deal with random failures when they occur. So that's nothing new, either.
7 replies →