Comment by drob518
5 hours ago
> It remains unclear whether continuing to throw vast quantities of silicon and ever-bigger corpuses at the current generation of models will lead to human-equivalent capabilities. Massive increases in training costs and parameter count seem to be yielding diminishing returns. Or maybe this effect is illusory. Mysteries!
I’m not even sure whether this is possible. The current corpus used for training includes virtually all known material. If we make it illegal for these companies to use copyrighted content without remuneration, either the task gets very expensive, indeed, or the corpus shrinks. We can certainly make the models larger, with more and more parameters, subject only to silicon’s ability to give us more transistors for RAM density and GPU parallelism. But it honestly feels like, without another “Attention is All You Need” level breakthrough, we’re starting to see the end of the runway.
I see a lot of researchers working on newer ideas so I wouldn't be surprised if we get a breakthrough in 5-10 years. After all, the gap between AlexNet and Attention is All You Need was only 6 years. And then Scaling Laws was about 3-4 years after that. It might seem like not much progress is being made but I think that's in part because AI labs are extremely secretive now when ideas are worth billions (and in the right hands, potentially more).
Of course 5-10 years is a long time to bang our heads against the wall with untenable costs but I don't know if we can solve our way out of that problem.
I think we will see models becoming small reasoning core which don't remember tonnes of facts but can reason with data fed to it or they can search.
The echoes of A.I. winter.
I think in domains like Math and Software Engineering, they are less constrained by training data anyway. They can synthetically generate and validate programs. To what extent that scales into novel insights is a different matter, but I think they dream of the AlphaGo Zero moment at least in verifiable domains.
> I’m not even sure whether this is possible.
Based on what's happened so far, maybe. At least that's exactly how we got to the current iteration back in 2022/2023, quite literally "lets see what happens when we throw an enormous amount data at them while training" worked out up until one point, then post-training seems to have taken over where labs currently differ.
Right, but we played the scaling card and it worked but is now reaching limits. What is the next card? You can surely argue that we can find a new one at any time. That’s the definition of a breakthrough. I just don’t see one at the moment.
> I just don’t see one at the moment.
Did you see the one before the current one was even found? Things tend to look easy in hindsight, and borderline impossible trying to look forward. Otherwise it sounds like you're in the same spot as before :)
1 reply →
better tooling and integration
We pay people to create more high quality tokens (mercor, turing) which are then fed into data generating processes (synthetic data) to create even more tokens to train on
But does that really help, or do you get distortion? The frequency distribution of human generated content moves slowly over time as new subjects are discussed. What frequency distribution do those “data generating processes” use? And at root, aren’t those “data generating processes” basically just another LLM (I.e., generating tokens according to a probability distribution)? Thus, aren’t we just sort of feeding AI slop into the next training run and humoring ourselves by renaming the slop as “synthetic data?” Not trying to be argumentative. I’m far from being an AI expert, so maybe I’m missing it. Feel free to explain why I’m wrong.
That's the problem in a nutshell. There is an art to how you generate the synthdata so that you don't get crappy trained models (especially when mistakes cost XX million dollars).
It's also theoretically why facebook paid 14bn for alex wang and scale ai
> The current corpus used for training includes virtually all known material.
This is just totally incorrect. It's one of those things everyone just assumes, but there's an immense amount of known material that isn't even digitized, much less in the hands of tech companies.
What large caches of undigitized content exists? Surely, not everything has been digitized, but I can’t think it’s much in percentage terms.
The amount of private data that is locked up inside private internal databases is huge. This is especially true of regulated industries. There is a wealth of data - financial data showing how to budget for things, pricing data on various products that are B2B, standard operating procedures at mature companies that have gone through various revisions, designs for manufacturing plants so people don't keep reinventing and making the same mistakes again, and on and on.
The Vatican Library contains roughly 1.1 million printed books and around 75,000 codices, only a small percentage of which have been digitised.
2 replies →