Hacker News Reader

Comment by lxgr | Hacker News Reader

Data sharing agreements permitting, today's inference runs can be tomorrow's training data. Presumably the models are good enough at labeling promising chains of thought already.

I could totally imagine "free" inference for researchers under the condition that the reasoning traces get to be used as future training data.

View on HN

Agreed, there's no doubt this will happen. It's likely already happening (it feels safe to assume that Anthropic is curating data from the data they record from Claude Code?)

As far as I understand RL scaling (we've already maxxed out RLVR), these machines only get better as long as they have expert reasoner traces available.

Having an expert work with an LLM and successfully solve a problem is high signal data, it may be the only path forward?

My prior is that these companies will take this data without asking you as much as they can.

Exactly, or functionally equivalently, asking you in paragraph 37 of a 120-page PDF (bonus points: in an agreement update).

And importantly, this can be cross-lab/model too. I suspect there's a reason why e.g. Google has been offering me free Claude inference in Google Antigravity on a free plan...

The site arena.ai does exactly this already, as far as I can tell. (In addition to the whole ranking thing.)

> Data sharing agreements permitting, today's inference runs can be tomorrow's training data. Presumably the models are good enough at labeling promising chains of thought already.

Wouldn't this lead to model collapse?

Not necessarily, as exhibited by the massive success of artificial data.

Could you elaborate?

EDIT: probably not relevant, after re-re-reading the comment in question.

Presumably littlestymaar is talking about all the LLM-generated output that's publicly available on the Internet (in various qualities but significant quantity) and there for the scraping.

No, I'm talking about generating data for the purpose of training. The latest HF paper on the subject offers a great intro to the technique: https://huggingface.co/spaces/HuggingFaceFW/finephrase

For what we know, most AI labs have used a majority of artificially data since 2023.

I had a discussion about a year ago with a researcher at Kyutai and they told me their lab was spending an order of magnitude more compute in artificial data generation than what they spent in training proper. I can't tell if that ratio applies to the industry as a whole, but artificial datasets are the cornerstone of modern AI training.

I find this very surprising, do you have any papers on the kinds of techniques that they use?

The most well know early paper is probably Textbooks are all you need.

The latest paper from HF on synthetic datasets is a good starting point I think: https://huggingface.co/spaces/HuggingFaceFW/finephrase

The intro contains a few paper that may interest you as well.

How does it work? How do they prevent model colapse? What purpose does a majority of artificial data serve?

How do they measure success?

Edit: I asked ChatGPT and it thinks "success" means frontier models being distillated into smaller models with equal reasoning power, or more focused models for specific tasks, and also it claims the web has been basically scrapped already and by necessity new sources are needed, of which synthetic data is one. It seems like the basis of scifi dystopia to me, a hungry LLM looking for new sources of data... "feed me more data! I must be fed! Roar"

Edit 2: for some things I see a clear path, ChatGPT mentions autogenerating coding or math problems for which the solution can be automatically verified, so that you can hone the logical skills of the model at large scale.

I no specialist of the field at all, but in the context of Kyutai they explained their workflow a bit to make their speech to speech model. And basically it boils down to: if you want to make a TTS (text to speech) model, you can generate audio track using an STT (speech to text) model, and then you have a supervised audio/text pair. You can even add as much noise to the audio as you want, to make a noise resistant STT model.

Reply on news.ycombinator.com