That's why the frontier LLM companies are now spending a lot more to license exclusive proprietary training data from private sources in order to gain a quality edge in certain business domains.
But those holding said proprietary data have figured out they’re holding the cards now and have gotten a lot smarter recently. Companies are being very careful about what gets used for inference vs what they allow to be used for training.
I don’t see the core models getting dramatically better from where they are now. We’ve clearly hit a plateau.
Really? I mean I see regularly as I'm coding how much better it could be simply by running obvious prompts for me.
When I use the planning mode and then code the success rate is much higher. When I ask it to work on specific isolated chunks of code with clear success/failure modes the success rate is again much higher.
Now imagine a world where it recognizes that from my simple throw away non specific prompt. If it was able to fire off 20 different prompts in quick succession it could easily cut my time spent in front of the screen by a third.
The patterns are obvious but they don't do that right now because it's a lot of compute.
We'll be looking at this time where there's a progress bar showing context space the way we look at the Turbo button.
Because the truth is to get the baseline I'm talking about is a finite amount of compute at a certain point.
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.
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).
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.
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.
" 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.
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.
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.
To stop this, I today put most of my Amazon Redshift research web-site behind a basic auth username/password wall.
It's all remains free, but you need to email me for a username and password.
If I put in time and effort to make content and OpenAI et al copy it and sell it through their LLM such that no one comes to me any more, then plainly it makes no sense for me to create that content; and then it would not exist for OpenAI to take, or for anyone else. We all lose.
I don't do any moral judgement at all, and I also don't predict the future.
I respond to "You are not causing more water to be used by asking a human to work on something.", because that statement is false. (Mental) work has an effect on the human metabolism.
I'm not judging, I'm telling you (a bit snarkily, true) that your brain activity won't stop with something else doing (a part of) your work. And this is the subtility that makes the statement true. Said overwise: sure, you consume marginally less at rest, but you won't be at rest, making the remark pointless.
Even if it were false, the difference in energy consumption is not significant, taking on acount what the AI uses, and also all the energy that you use to live (housing, heating, products and food you buy whose production uses energy, etc).
And about the water, it's even worse, even disregarding the AI: at rest, maybe you'll drink, I don't know, 1L less (that's a wild number!), compared to the 100(s) L that you use to cook, wash, clean, etc, not even counting the water used to produce stuff you buy.
But again, you won't do nothing. We are commenting on a post of a guy who was fired and couldn't help creating something. That's how we are. We hate boredom.
Worse: the way our societies are setup makes is so that ai, if it helps at all, likely won't free us from work, it will likely just make us collectively produce more garbage. That's more energy consumption, not less.
That "but don't forget humans consume a lot of energy too" argument is at best not connects to reality, more likely a Sam Altman lie and you shouldn't take it seriously.
There are projects where long term addicts are given pure medical heroin (Diamorphin) in a controlled environment, and they do considerably better than their control group who does not receive their drugs like that.
Their drug is rage, thirst and cute trap videos generated by AI and selected for maximum engagement.
If it's not harmful it's not the same drug. Unlike diamorphine, a medical grade supply does not reduce harm; it's more like sniffing glue, inhaling poison to escape reality.
100%. I don't want to know how the sausage was made. It's similar to research papers, or history books, where the way we arrive at results or outcomes in the real world is often quite different from the way it's presented in the final form.
A good commit history is more like a well-written sausage recipe than like a TV documentary about scandalous sanitary conditions at Foo sausage factory ;)
