this post was submitted on 09 Mar 2024

99 points (89.6% liked)

Technology

59534 readers

3195 users here now

This is a most excellent place for technology news and articles.

Our Rules

Follow the lemmy.world rules.
Only tech related content.
Be excellent to each another!
Mod approved content bots can post up to 10 articles per day.
Threads asking for personal tech support may be deleted.
Politics threads may be removed.
No memes allowed as posts, OK to post as comments.
Only approved bots from the list below, to ask if your bot can be added please contact us.
Check for duplicates before posting, duplicates may be removed

Approved Bots

founded 1 year ago

MODERATORS

How Quickly Do Large Language Models Learn Unexpected Skills? (nautil.us)

submitted 8 months ago by dominiquec@lemmy.world to c/technology@lemmy.world

19 comments fedilink hide all child comments

So-called "emergent" behavior in LLMs may not be the breakthrough that researchers think.

top 19 comments

sorted by: hot top controversial new old

[–] UraniumBlazer@lemm.ee 52 points 8 months ago (3 children)

TLDR: Let's say you want to teach an LLM a new skill. You give them training data pertaining to that skill. Currently, researchers believe that this skill development shows up suddenly in a breakthrough fashion. They think so because they measure this skill using some methods. The skill levels remain very low until they unpredictably jump up like crazy. This is the "breakthrough".

BUT, the paper that this article references points at flaws in the methods of measuring skills. This paper suggests that breakthrough behavior doesn't really exist and skill development is actually quite predictable.

Also, uhhh I'm not AI (I see that TLDR bot lurking everywhere, which is what made me specify this).

[–] Venator@lemmy.nz 16 points 8 months ago (1 children)

Also, uhhh I'm not AI

An AI would say that... 😂

[–] gravitas_deficiency@sh.itjust.works 4 points 8 months ago

Clearly, the AI is learning deception

[–] inspxtr@lemmy.world 8 points 8 months ago (2 children)

re: your last point, AFAIK, the TLDR bot is also not AI or LLM; it uses more classical NLP methods for summarization.

[–] far_university1990@feddit.de 3 points 8 months ago

https://github.com/RikudouSage/LemmyAutoTldrBot readme say summarization is in summarizer.py which use sumy, specifically LSA which documented here

[–] dirtySourdough@lemmy.world 1 points 8 months ago

Natural language processing falls under AI though, and so do large language models (see chapters 23 and 24 of Russell and Norvig, 2021 http://aima.cs.berkeley.edu/).

[–] dust_accelerator@discuss.tchncs.de 7 points 8 months ago (1 children)

Also, uhhh I'm not AI

That's exactly what an AI would say that got an emergent skill to lie

🤥

[–] XTL@sopuli.xyz 4 points 8 months ago

Or a model that picked up on a pattern of sources saying that.

[–] Norgur@fedia.io 27 points 8 months ago (3 children)

What always irks me about those "emergent behavior" articles: no one ever really defines what those amazing"skills" are supposed to be.

[–] General_Effort@lemmy.world 12 points 8 months ago

The term "emergent behavior" is used in a very narrow and unusual sense here. According to the common definition, pretty much everything that LLMs and similar AIs do is emergent. We can't figure out what a neural net does by studying its parts, just like we can't figure out what an animal does by studying its cells.

We know that bigger models perform better in tests. When we train bigger and bigger models of the same type, we can predict how good they will be, depending on their size. But some skills seem to appear suddenly.

Think about someone starting to exercise. Maybe they can't do a pull-up at first, but they try every day. Until one day they can. They were improving the whole time in the various exercises they did, but it could not be seen in this particular thing. The sudden, unpredictable emergence of this ability is, in a sense, an illusion.

For a literal answer, I will quote:

[Emergent abilities appear in an] arithmetic benchmark that tests 3-digit addition and subtraction, as well as 2-digit multiplication. GPT-3 and LaMDA (Thoppilan et al., 2022) have close-to-zero performance for several orders of magnitude of training compute, before performance jumps to sharply above random at [13B parameters] for GPT-3, [68B parameters] for LaMDA. Similar emergent behavior also occurs at around the same model scale for other tasks, such as transliterating from the International Phonetic Alphabet recovering a word from its scrambled letters, and Persian question-answering. Even more emergent abilities from BIG-Bench are given in Appendix E.

[–] NounsAndWords@lemmy.world 11 points 8 months ago (1 children)

Emergent behavior is pretty much anything an old model couldn't do that a new model can. Simple reasoning, creating coherent sentences, "theory of mind", basic math, translation, I think are a few examples.

They aren't "amazing" in the sense that a human can't do them, but they are in the sense that a computer is doing it.

[–] admin@lemmy.my-box.dev 5 points 8 months ago

They aren't "amazing" in the sense that a human can't do them, but they are in the sense that a computer is doing it.

... without specifically being trained for it, to be precise.

[–] anlumo@lemmy.world 11 points 8 months ago (2 children)

One of those things I remember reading was the ability of ChatGPT to translate texts. It was trained with texts in multiple languages, but never translation specifically. Still, it’s quite good at it.

[–] variaatio@sopuli.xyz 3 points 8 months ago (1 children)

That is just its core function doing its thing transforming inputs to outputs based on learned pattern matching.

It may not have been trained on translation explicitly, but it very much has been trained on these are matching stuff via its training material. Since you know what its training set most likely contained..... dictionaries. Which is as good as asking it to learn translation. Another stuff most likely in training data: language course books, with matching translated sentences in them. Again well you didnt explicitly tell it to learn to translate, but in practice the training data selection did it for you.

[–] anlumo@lemmy.world 6 points 8 months ago (1 children)

The data is there, but simpler models just couldn’t do it, even when trained with that data.

Bilingual human children also often can’t translate between their two (or more) native languages until they get older.

[–] bus_factor@lemmy.world 3 points 8 months ago

That's interesting. My trilingual kids definitely translate individual words, but I guess the real bar here is to translate sentences such that the structure is correct for the languages?

[–] bionicjoey@lemmy.ca 2 points 8 months ago

A lot of the training set was probably Wiktionary and Wikipedia which includes translations, grammar, syntax, semantics, cognates, etc.

[–] kromem@lemmy.world 6 points 8 months ago

I'm not sure why they are describing it as "a new paper" - this came out in May of 2023 (and as such notably only used GPT-3 and not GPT-4, which was where some of the biggest leaps to date have been documented).

For those interested in the debate on this, the rebuttal by Jason Wei (from the original emergent abilities paper and also the guy behind CoT prompting paper) is interesting: https://www.jasonwei.net/blog/common-arguments-regarding-emergent-abilities

In particular, I find his argument at the end compelling:

Another popular example of emergence which also underscores qualitative changes in the model is chain-of-thought prompting, for which performance is worse than answering directly for small models, but much better than answering directly for large models. Intuitively, this is because small models can’t produce extended chains of reasoning and end up confusing themselves, while larger models can reason in a more-reliable fashion.

If you follow the evolution of prompting in research lately, there's definitely a pattern of reliance on increased inherent capabilities.

Whether that's using analogy to solve similar problems (https://openreview.net/forum?id=AgDICX1h50) or self-determining the optimal strategy for a given problem (https://arxiv.org/abs/2402.03620), there's double digit performance gains in state of the art models by having them perform actions that less sophisticated models simply cannot achieve.

The compounding effects of competence alone mean that progress here isn't going to be a linear trajectory.

[–] spujb@lemmy.cafe -1 points 8 months ago

what material benefit does having a cutesy representation of phrenology, a pseudoscience used to justify systematic racism, bring to this article or discussion?