this post was submitted on 20 Nov 2025
48 points (98.0% liked)
Fuck AI
4728 readers
1065 users here now
"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
AI, in this case, refers to LLMs, GPT technology, and anything listed as "AI" meant to increase market valuations.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Are you able to elaborate on this at all?
I don't personally work in the AGI space, but there have been some massive improvements within the last 2 years even. The public only has access to the most constrained, well-understood models, and even those are pretty good. I (and LeCunn, Hinton, other big names) don't think transformers + massive compute are the solution to AGI, but even that combination now leads to emergent unexplainable capabilities. I work in the field and even I think it's magic sometimes.
edit: lol I should've expected to be downvoted for talking about AI, in the Fuck AI community
Ok, but that's only a sign that you don't understand it enough. Which would be catastrophal in case of AGI.
Absolutely. That's called "learning". We observe some kind of behavior, investigate, make discoveries, then incorporate into the broader knowledge base. As for AGI, it's going to be just like humans, where we can probe behaviors on the surface and investigate 2nd order effects (through things like linear probes), but we won't be able to understand 100% of exactly how it makes a decision.
Really curious to hear about these. Do you have any examples?
Basically LLMs can do things it wasn't explicitly trained to do once a certain relative scale is reached. This is for LLMs but other model families show (considerably less) potential too. Keep in mind this is from THREE YEARS AGO: https://arxiv.org/pdf/2206.07682
and it's only accelerated since
What are some examples?
too much to write here, look at Table 1 in the paper posted above, and you can explore from there
I don't find that terribly compelling... It looks like there's also a large body of research disputing that paper and others like it.
Here is just one such paper that presents a pretty convincing argument that these behaviors are not 'emergent' at all and only seem that way when measured using bad statistics: https://arxiv.org/pdf/2304.15004
As with anything, especially in a field moving this fast, yes of course it's not black and white. Here's an article I just found that goes into more detail if you're curious. The first paper I shared was the one I read a while ago but there are dozens of them. Also I don't work in NLP, more in computer vision and physics-informed neural networks (PINNs), so I don't know all the most recent developments of LLMs (though I use ViTs in my work all the time).
I really don't wanna sound like a dick, but did you actually read that article? It basically just concludes that there's no consensus on whether or not LLMs are exhibiting emergent behaviors — only that they're very difficult to predict. Funny enough, it even spends half the article discussing the exact paper I shared above.
One thing it doesn't discuss but that I also think needs to be brought up is that even if a model shows emergent behavior at one level of scale, that's no guarantee that further emergence effects will continue to 'unlock' at higher scales. So yeah, it's definitely worth doing more research on... but the idea that LLMs might have emergent behaviors and that they might get even more emergent at scale should be enough to justify some expensive research grants, but not a trillion dollar industry.
That's why I shared it. It covers the general viewpoints and presents evidence for each. It's not a well established area so researchers can claim various things based on their specific metrics and definitions. Still blows my mind overall how far this has come in such a short time.
I don't think anyone has claimed that explicitly (I haven't).
Bigger picture is that LLMs are still not perfectly understood and because of that there's a lot of big claims from salespeople that want to sell their product. Also as of last year (which is the last time I did a dive on AGI), there were many that didn't believe that simply scaling LLMs is the way to AGI (btw you should look up the prevailing proposals for how to reach AGI). That doesn't mean there isn't a ton of use for LLMs + application-specific engineering in the meantime, because that's already shown to be true. This is where companies are making bank, not in the fundamental research level.