[ajross]

That sounds like spin to me. If there were a clear "quality edge" in "certain business domains" stemming from "exclusive proprietary data", someone would have been exploiting it already using meat computers.

But no, businesses are dumb. They always have been. Existing businesses get disrupted by new ideas and new technology all the time. This very site is a temple to disruption!

Proprietary advantage is, 99.999% of the time, just structural advantage. You can't compete with Procter & Gamble because they already built their brands and factories and supply chains and you'd have to do all that from scratch while selling cheaper products as upstart value options. And there's not enough money in consumer junk to make that worth it.

But if you did have funding and wanted to beat them on first principles? Would you really start by training an LLM on what they're already doing? No, you'd throw money at a bunch of hackers from YC. Duh.

[linkregister]

Frontier labs are paying the same constellation of firms offering proprietary data and access to experts in their fields to train LLMs.

They are neck-and-neck only because they are participating in the arms race. The only other way to keep up is mass-distillation, which could prove to be fragile (so far it seems to be sustainable).

[ajross]

Meh. I think there's basically no benefit shown so far to careful curation. That's where we've been in machine learning for three decades, after all. Also recognize that the Great Leap Forward of LLMs was when they got big enough to abandon that strategy and just slurp in the Library of All The Junk.

I think one needs to at least recognize the possibility that... there just isn't any more data for training. We've done it all. The models we have today have already distilled all of the output of human cleverness throughout history. If there's more data to be had, we need to make it the hard way.

[bonoboTP]

Ok, maybe pretraining is now complete and solved. Next up: post-training, reinforcement learning, engineering RL environments for realistic problem solving, recording data online during use, then offline simulation of how it could have gone better and faster, distilling that into the next model etc. etc. There's still decades worth of progress to be made this way.

[k32k]

" There's still decades worth of progress to be made this way."

That's not true. Moreover the progress can slow to a crawl where it's barely noticeable. And in that world the humans continues to stay ahead - that's the magic of humans. To be aware of surroundings and adapt sufficiently whilst taking advantage of tools and leveraging them.

[linkregister]

This is an interesting theoretical statement that does not survive a collision with reality. The long-tail expert RHLF training is effective. We have seen significant employment impact to call center employees. This does not mean its progress will be cheap or immediate.

[k32k]

I think this is where we are at, too.

But if you say stuff like this on here you get down voted. Why?

[nradov]

The quality edge hasn't shown up yet. If this strategy actually works then the quality improvements will only become apparent in the next round of major LLM updates. There's a lot of valuable training data locked up behind corporate firewalls. But this is all somewhat speculative for now.