This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
This has been an annoying aspect of LLM AI hype. There are plenty of indicators of something going on but many of the test results are not of that set. If you train them on the question sets and answer keys for repeatably mechanically gradable exams like the SAT, GRE or bar exams then it should be expected that they will perform well on them.
What would be really nice is if whenever the AI produced content it had to also tell you the minimal edit distance between that content and some content (or a direct combination of contents) in its training set. That way you could have a good measure of how much original content it was actually producing vs. how much it was just paraphrasing its training set. Or at least it would be useful to have extensive data on the average edit distance between a response and some item in the corpus.
Your proposed method doesn't work - even if you just turned a query into an embedding, picked the closest text in its dataset, and then ran that text through google translate and back a few times to obfuscate word choice, order, and other things like that, it'd change enough that literal edit distance would still be very high.
An analogy to image models, here's a claimed example of of taking inspiration from a particular photo in a training set. It's really not that close.
Not really, no. Edit distance is relative to which operations count as primitive "edits" and the "cost" of each use of that operation. There are specific forms of edit distance of which what you are saying is true, but you could also have an edit distance where "run it through Google Translate" is a primitive edit operation. Obviously, you would have to pick the operations to fit the specific model, e.g. what external resources it has access to.
Okay, I assumed you meant character-level edit distance, because that's what the article you linked was exclusively about. But without that, 'edit distance' isn't really a useful term, as we don't know what a 'primitive operation' is in the context of a LLM because we do not know that much if anything about what they actually do internally.
Based on what I've seen, 'most similar example in training set' doesn't capture the extent to which LLMs memorize things. Even if they are memorizing a lot, it's memorization in a very complicated way - otherwise the 'write a story about X in the style of Y', where X and Y hadn't ever been done before, just wouldn't work.
how is this memorizing? or this? Like, there's some extent to which it's memorizing things more than humans do, certainly. But it's a very vague sense, and positing a metric like edit distance that fully captures that sense just restates the problem, because we don't know what that is
That isn’t true, the formal definition doesn’t restrict what operations there are (if you disagree then you should quote it and tell me where), it just requires that they be operations on strings. Whatever else using Google Translate is, it’s an operation on strings.
I’m not suggesting that we base primitive operations on what the LLM does internally. Your initial example wasn’t about what the LLM does internally, it was about “what if the LLM ran things through Google Translate a few times to trick you.” I’m saying you could supplement a more basic measure with additional operations to capture when the model has access to external programs like Google Translate.
I never said that the AI was just “memorizing” things, that’s an obvious strawman. All I said was that edit distance would help you tell how much the AI was paraphrasing from its training set, which doesn’t presuppose that it’s not paraphrasing very little. You seem to have misinterpreted me as saying that an edit distance would show the AI was paraphrasing a lot, when all I said was an edit distance would be a helpful metric of how much it was. I didn’t make any specific commitments about how much that would be across a wide range of cases. Nor did I suggest that an edit distance would fully capture things, just that it would give a good sense of how close things were.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link