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 -
LLMs can reproduce 96% of the text of Harry Potter verbatim. Even if they do not store all their training data with perfect fidelity, their underlying operations are such that it doesn't matter. It's data compression with variable loss depending on what they were trained on. When 1:1 outputs from their memories of training data can't exist, they reach for similar patterns and smooth over the disjunctions using sophistry. They must be commended for semantic fluency.
What is this supposed to prove? There are people who have memorized the Torah or the Quran. It's still not the case that they are merely doing some kind of database lookup when you ask them about a verse, and that implies a fidelity that simply doesn't exist. And if you concede that there isn't perfect fidelity, one wonders what the purpose of discussing "database lookups" in the context of LLM inference is other than rhetoric.
Dismissing as mere sophistry novel LLM-discovered software exploits and math theorems is absurd.
It goes towards proving the basis for what we observe: that LLMs are very good at recalling large and disparate amounts of knowledge but are poor for functionally utilizing said knowledge, especially in matters complex, unusual, or otherwise not 1:1 with stuff from their training material. Whether this proves or disproves they are sentient or intelligent or whatever is a matter of semantics, but what it does do is give us a clue as to why we observe certain disparities in their capabilities, and can help inform our expectations about what further capabilities might emerge.
Humans lean on theory, trained pattern spotting, and various heuristics or memorized devices (i.e. king opposition) when playing chess. Memory plays a role to, but outside of maybe Magnus Carlson it is dwarfed by the capacities of LLMs. This is a level of intelligence that can also be employed for creating architecture or symphonies. LLMs lean a lot harder on brute memory recall (although I won't discount entirely their capacity for higher-tier reasoning) through hyper-intensive statistical calculations, and these make it very good for things like discoursing on a broad variety of facts or semantically juggling abstractions, but they do not, apparently, allow LLMs to create complex architecture, symphonies, or do anything else involving the complex interlocking of smaller elements.
The small elements are found in its memory and can be expurgated intact individually, but the LLMs do not possess the intelligence to complexly fit them together. The LLMs do not operate at a level of intelligence that would allow that. They are hyper-intensive exploiters of lower order processes but not high tier ones. That's what's suggested by the fact they can recall 96% of a novel. That they lean on highly scaled relatively brutish methods to repeat stuff verbatim, or close enough.
Like I said:
"LLMs haven't written a beautiful symphony or designed a beautiful building" is simply moving the goal posts. There's no reason that those are the true test of putting things together and theorems and exploits don't count.
More options
Context Copy link
More options
Context Copy link
I take the ‘opposite’ view that LLMs are becoming extraordinary intelligences, but I also think the distinction between memory, recall, training set, database etc is unnecessarily importing computer science distinctions into what is a relatively robust colloquial understanding of these models.
If you watch three thousand chess games and then play a chess game and see a move and think “I’ve seen this before, I’m going to do x” and you’re right but you can’t perfectly recall that it was actually a YouTube video of a 2003 Chess regional championship quarter final between… then are you recalling or remembering or did you learn?
This is just not a relevant distinction when it comes to the human concept of memory. I’ll keep pushing this because “actually, an LLM doesn’t have memory of the training set” isn’t really true. It often does have recall of the training set, just like often you really might be able to remember the book you first saw an unusual turn of phrase in or the chess game where you first saw a particular move. And in any case, memory encompasses both that and a relational, situational, partial and often metadata-free recall but it still counts.
The counterargument here isn’t “no LLMs don’t do this”, it’s “so do you”.
This doesn't exist, at least on this forum on down. There's at least one person I talked to who really thought that LLMs were looking through the training data at inference time. It turns out that people using sloppy language ""colloquially"" ("joke's on you, I was only pretending to misunderstand LLMs") can cause people to believe the literal meaning if they don't know any better.
Agreed.
This isn't what I said. I said it doesn't have access to the training set, in the same way that if you take an exam without "access" to the textbook you're not allowed to bring it in and leaf through it when answering the problems. It doesn't preclude you from reading the textbook a thousand times and memorizing it verbatim though.
Again, if we change “it accesses the training set” to “it recalls / accesses / understands [delete as appropriate] the conceptual relationships represented by the training set” what really changes?
What changes is that one is completely wrong and the other is defensible. One leads you to think that an LLM recognizing Shakespeare (or a lesser author) from a sample of writing is an unremarkable feat, an information retrieval problem that's been solved for forty years. The other causes you to realize that what's going on here is much deeper.
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