On the slow death of scaling
papers.ssrn.com
papers.ssrn.com
I’m trying to turn that into something testable with a simple constraint: “one hobbyist GPU, one day.” If meaningful progress is still possible under tight constraints, it supports the idea that we should invest more in efficiency/architecture/data work, not just bigger runs.
My favorite line >> Somewhat humorously, the acceptance that there are emergent properties which appear out of nowhere is another way of saying our scaling laws don’t actually equip us to know what is coming.
Regarding this paragraph >> 3.3 New algorithmic techniques compensate for compute. Progress over the last few years has been as much due to algorithmic improvements as it has been due to compute. This includes extending pre-training with instruction finetuning to teach models instruction following ..., model distillation using synthetic data from larger more performant "teachers" to train highly capable, smaller "students" ..., chain-of-thought reasoning ..., increased context-length ..., retrieval augmented generation ... and preference training to align models with human feedback ...
I would consider algorithmic improvements to be the following 1. architecture like ROPE, MLA 2. efficiency using custom kernels
The errors in the paper 1. Transformers for language modeling (Vaswani et al., 2023). => this shd be 2017
Disclosure: my proposed experiments: https://ohgodmodels.xyz/
In the area of AI, secrecy and inability to reproduce/verify can become a huge systemic and social problem, the possible damage is literally unbounded.
That's why I like open source AI, including training data and process, it solves the above problem as well as the problem of duplication of effort which leads to a huge waste of resources, waste that is economically significant on national and global scales.
Is this actually accepted? Ever since [0], I thought people recognized that they don't appear out of nowhere.
In my experience with agent assisted coding, how well it works seems very closely tied to the quantity and quality of training material. It also has some identifiable qualities like verifiability that make it a particularly good target for an LLM. I would not call that surprising or emergent.
I don't think that paper is widely accepted. Have you seen the authors of that paper, or anyone else, use it to successfully predict (rather than postdict) anything?
If they can't predict new emergence, then 'explaining' old emergence by post hoc prediction with bizarre newly-invented metrics would seem to be irrelevant and just epicycles. You can always bend a line as you wish in curve-fitting by adding some parameters.
I think there is a critical flaw in the paper though. Not critical from a technical standpoint, but from a reviewer standpoint. They don't bridge the gap to the final step of transitioning to a hard loss. But you can easily experiment with this by yourself even on smaller models and datasets and it is pretty effective. I think the logic is quite straight forward though and this isn't actually necessary to prove their point, which is why I think they didn't do it. But most ML people are hyperfixated on benchmarks and empirical evidence. Hell, that's why we kinda killed small scale research. It isn't technically wrong to ask for more scale and more datasets but these types of questions are unbounded so can be used too much as a crutch.
FWIW I also think the original Emergent Abilities paper has a critical flaw. Look at their definition (emphasis my own)
Specifically, we define emergent abilities of large language models as abilities that are not present in smaller-scale models but are present in large-scale models; ***thus they cannot be predicted by simply extrapolating the performance improvements on smaller-scale models.***
But I also don't know why we've accepted this definition of emergent abilities. It isn't useful.
Without their explicit distinction of extrapolating we'd call nearly every model emergent. Here's my proof: for any given model there is almost surely a smaller model that performs worse. Dumb, but that's the problem with the definition. But using their distinction we run into the problem of concluding things are emergent simply because we didn't realize we were doing things a certain way.
And using the more classic definition of emergence[0,1], distinguishing between strong and weak, we should recognize that all neural nets are by definition weakly emergent. (Emphasis from [0])
high-level phenomenon is *strongly emergent* with respect to a low-level domain when the high-level phenomenon arises from the low-level domain, but truths concerning that phenomenon are not *deducible* even in principle from truths in the low-level domain.
high-level phenomenon is *weakly emergent* with respect to a low-level domain when the high-level phenomenon arises from the low-level domain, but truths concerning that phenomenon are *unexpected* given the principles governing the low-level domain.
But strong emergence requires a very difficult proof. We must prove that we aren't just so dumb that we do not know how to deduce the results but that we cannot deduce the results. It means there must be a process that results in an unrecoverable information loss. I think everyone should be quite suspicious of any claims of strong emergence when it comes to AI. I mean... we have the weights... so the results are de fact deducible...
So I don't know why we talk about emergence the way we do in ML. I frequently hear people say things are emergent phenomena because they didn't realize they were teaching the model certain capabilities but that doesn't mean someone else wouldn't be able to (and boy are there many "emergent phenomena" that ML people "can't" predict but a mathematician would).
I especially agree with your point that scaling laws really killed open research. That's a shame and I personally think we could benefit from more research.
I originally didn't like calling them scaling laws.
In addition to the law part seeming a bit much, I've found that researchers often overemphasize the scale part. If scaling is predictable, then you don't need to do most experiments at very large scale. However, that doesn't seem to stop researchers from starting there.
Once you find something good, and you understand how it scales, then you can pour system resources into it. So I originally thought it would encourage research. I find it sad that it seems to have had the opposite effect.
Exactly like semiconductor wafer processing.
Board games like the Settlers of Catan are good examples of the behavior— concretely the start of the game when everyone trades vs the end of the game when if you suspect someone wins it makes little sense to trade unless, you think it will help you win first.
People are fooling themselves if they think AGI will be zero sum. Even if only one group somehow miraculously develops it, there will immediately be fast followers. And, the more likely scenario is more than one group would independently pull it off - if it's even possible.
Ilya Sutskever (Sep 20, 2017)
> The goal of OpenAI is to make the future good and to avoid an AGI dictatorship. You are concerned that Demis could create an AGI dictatorship. So do we. So it is a bad idea to create a structure where you could become a dictator if you chose to, especially given that we can create some other structure that avoids this possibility.
Nick Bostrom - Decisive Strategic Advantage https://www.lesswrong.com/posts/vkjWGJrFWBnzHtxrw/superintel...
* imagine if Google alone had LLMs. For an innocuous example, the only provider in my LLM client that regularly fails unit tests verifying they actually cache tokens and utilize them on a subsequent request is Gemini. I used to work at Google and it'd be horrible for that too-big-for-its-own-good institution regressing to the corporate mean to own LLMs all by itself
Why do people keep repeating this. The only way artificial intelligence is impossible is if intelligence is impossible. And we're here so that pretty much removes that impediment.
"One thing is certain, is the less reliable gains from compute makes our purview as computer scientists interesting again. We can now stray from the beaten path of boring, predictable gains from throwing compute at the problem."
Isn't Ilya Sutskever who said some months ago that we were going back to research ?
Compute is a massive driver for everything ML. From number of experiments you can run in paralle, to how much RL you can try out, how long stuff is running etc.
ML is pushing scaling on dimensions we haven't had before (number of Datacenters, amount of energy we put into them) and ML is currently seen as the holy grail.
But i'm definitly very very curious how this compute and current progress is playing out in the next few years. It could be that we hit a hard ceiling were every single % point becomes tremendesly costly before we hit a % point of benchmark archievements which makes all of that usable daily. OR we will se a significant change to our society.
I do not think its something in between tbh because it def feels like in an expoential progress curve we are currently in.
There is absolutely NO reason why that PDF shouldn't load today.
You want to make an existing system more efficient, then take away resources.
I also feel like most insiders were fully aware of this fact, but it was a neat sales pitch.
People spend money on this because it works. It seems odd to call observable reality a "pervasive belief".
> Academia has been marginalized from meaningfully participating in AI progress and industry labs have stopped publishing.
Firstly, I still see news items about new models that are supposed to do more with less. If these are neither from academia nor industry, where are they coming from?
Secondly, "has been marginalized"? Really? Nobody's going to be uninterested in getting better results with less compute spend, attempts have just had limited effectiveness.
.
> However, it is unclear why we need so many additional weights. What is particularly puzzling is that we also observe that we can get rid of most of these weights after we reach the end of training with minimal loss
I thought the extra weights were because training takes advantage of high-dimensional bullshit to make the math tractable. And that there's some identifiable point where you have "enough" and more doesn't help.
I hadn't heard that anyone had a workable way to remove the extra ones after training, so that's cool.
.
.
The impression I had is that there's a somewhat-fuzzy "correct" number of weights and amount of training for any given architecture and data set / information content. And that when you reach that point is when you stop getting effort-free results by throwing hardware at the problem.