[rafram]

The core of the training data is public, but the part that actually makes these models smart came from (pretty highly-paid) experts via platforms like Mercor. Claude didn't magically learn to write good code by reading all of GitHub - humans trained it in that, more or less manually.

[rapind]

If you pay me to curate a playlist of musical hits, can you now publish and charge people for access to that playlist (*including the curated material)? Can we do the same with movies? Books?

/edit Added a note to make it more obvious that the material is included in the playlist, just like the material is incorporated as part of curated AI models.

[tanseydavid]

>> If you pay me to curate a playlist of musical hits, can you now publish and charge people for access to that playlist?

If the contract was "work-for-hire" then yes, of course I can.

[rapind]

Maybe I wasn't clear. The playlist includes the material in it. Just like curated AI does.

[datsci_est_2015]

Given the breadth of LLM knowledge, I somehow doubt this. Sure, it’s probably responsible for the quality of LLM insights, but I don’t think anyone was asking experts about e.g. the complex ecological effects of invasive zebra mussels and their provenance in Lake Michigan.

[visarga]

No, they do RLVR (reinforcement learning with verifiable rewards) like everyone else. And probably use claude data too, with human in the loop and tool feedback.

[jaen]

...and the rest of the training data (ie. the entire corpus of copyrighted works) was not written by experts expecting compensation? Double standards.

[Ajedi32]

No, public data is not generally written by "experts expecting compensation".

By the way, I don't expect you to pay me for this comment. You can just read it for free. You're welcome.

[jaen]

Ugh, please don't read strawmen into other's arguments and try to follow the HN guidelines.

Also, how about making proper arguments yourself? The vast majority of the training data isn't generated by company-paid AI experts either.

Notably, books, even though they don't form a large part of the training data, significantly improve performance on some tasks (same way as expert-generated data).

Why do you think the AI labs are so eager about scanning (and then destroying) every book on the planet?

If you removed all copyrighted works from the training corpus, the model would be notably weaker.

[calgoo]

No, but people do upload data with an expectation that the data not being used without their permission (unless they do a BSD/MIT/Public domain like license). Otherwise, the platform AND/OR the user do expect the data NOT to be used for purposes other then what it was intended for. Your comment is still your comment, and the hacker news platform also has a say in this. If there had been an opt-in, then fine no problem, but there was none, they just trained on everything available, including downloading pirated books from the internet.

[Ajedi32]

I think it's unreasonable to post anything on a public forum and then expect to be able to control who reads it and for what purpose.

[calgoo]

Answering here as it wont let me reply: Just because you feel that something that is public, does not mean you can do whatever you want with it. You can't just copy an article from a news site and paste it on yours, that theft. If you dont agree, fine, but that is the law, and ALL the mega corps have been fighting to keep it this way for the last 20 years. If they want to steal everyones info, fine, then lobby to change the copyright laws and no problem.

[sneak]

Copying isn’t theft, we settled this in the 90s.

[pastel8739]

Books?

[Ajedi32]

The vast, vast, majority of AI training data is not books. I wouldn't be surprised if there's more text on HN alone than every book in the history of mankind (most of which are also no longer copyrighted).

[rafram]

I didn't say that.

[thom]

No, you just parroted an increasingly popular talking point, the entire purpose of which seems to be to absolve AI companies of the enormous theft that put them in the position to hire experts in the first place.

[rafram]

Well, I'd never heard anyone make it before, but sure. (I looked into Mercor a bit and know some people who've worked in data generation/labeling, which is what exposed me to that side of the operation.)

It doesn't absolve them of any theft, but it does make the assertion that they should be required to release their models to the public seem, to me, a bit farcical. There are dozens of free and open-weights models that have all trained on exactly the same web crawls and books as GPT-5 and Opus. The proprietary models are better because of proprietary data.

[franga2000]

Cool, then they can train their proprietary models on their proprietary data only.

Even if the other models were trained on the same data, which is unlikely, since they had less time and money to scrape it and fewer lawyers to be able to do something like pirate, the proprietary models are still largely built on the public data and wouldn't exist without it. At the very least, they should release the intermediate model, before training on their proprietary data. Not that that's how that works...

[thom]

I agree that saying that they have now trained on lots of proprietary data allows them to muddy the legislative waters further than they already have. What a happy coincidence!

[noitemtoshow]

I’d suggest you to learn more about how LLM training work. Training on internet data alone will not result in an agent answering your questions.

[mbesto]

> The proprietary models are better because of proprietary data

Source? Otherwise this is pure speculation.

[jaen]

Indeed, that's exactly why I replied - you omitted one side from the discussion.

[freejazz]

So? What about the authors of all the works these companies stole?