I mean maybe. There also might be something to this. OpenAI has been very opaque about training techniques.

That's just the base concern with every single model regardless of where they sourced their data from. Garbage in, garbage out.

Sure. Does that fact mean we're prohibited from expressing concerns about data quality? ShareGPT isn't representative of authentic, quality writing.

Right, but training an LLM on the output of another LLM can certainly exacerbate these issues

Maybe, but we are fast approaching the point (or more likely have crossed it already) where distinguishing between human and AI generated data isn't really possible. If Google indexes a blog, how does it know whether it was written with AI assistance and therefore should not be used for training? Heck, how does OpenAI itself prevent such a feedback loop from its own output (or that or other LLMs)?

Your argument would have a lot more force if we were past that point rather than fast approaching that point. Concerns about training data errors being compounded are much more important when you're talking about the bleeding edge.

And your question about how OpenAI prevents their training data from being corrupted is one we should be asking as well!

> If Google indexes a blog, how does it know whether it was written with AI assistance and therefore should not be used for training

Yes, this is an existential problem for Google and training future LLMs.

See also, https://www.theverge.com/23642073/best-printer-2023-brother-... and https://searchengineland.com/verge-best-printer-2023-394709

Or Google can just materialize the expected page into existence at search time.

... it's uncanny how it always finds what you thought you were looking for!

> Heck, how does OpenAI itself prevent such a feedback loop from its own output (or that or other LLMs)?

Seems trivial. Only use old data for the bulk? Feed some new data carefully curated?

Future job: token selector / archiving

<meta name="generator" content="human brain">

I'm only half joking.... I think we likely will end up with flags for human generated/curated content (and it will have to be that way round, as I can't imagine spammers bothering to put flags on AI-generated stuff), and we probably already should have an equivalent of robots.txt protocol that allows users to specify which parts of their website they would and wouldn't like used in the training of LLMs.

Reminds me of the old “evil bit” RFC[1]

[1] https://www.ietf.org/rfc/rfc3514.txt

If content with a "human-generated" flag is rated more highly in some way -- e.g. search results -- then of course spammers will automatically add that flag to their AI-generated garbage. How do you propose to prevent them?

I assume, like the actual meta generator tags, it wouldn't actually be a massive boon for regular search results

And if it's not: why bother.

I think something like this will definitely happen, and your suggestion is the cleanest implementation idea I've seen for it. I imagine there will be a service provided by Google and OpenAI where they verify your identity as a human and then grant you a token to put into your meta tags (wait a second... this sounds like sama's worldcoin idea...).

It will need to be based somewhat on the honor system (just because someone's proved they're a human doesn't mean they won't put their attestation on auto-generated text), but it definitely sounds better than nothing.

They'll still need to incentivize it somehow, though. Why do I as a human want to add that meta tag? If the answer is "better search ranking" then it renders the whole scheme mostly pointless because obviously spammers will want to acquire the attestation and attach it to their auto-generated content.

Pre-AI data is going to become like pre-nuclear steel.

Not to mention it's embarrassing. Google playing second banana to OpenAI.

That assumes that training on the output of another language model somehow gives you the ability to improve your model and to catch up somehow

It does. In general this is known as teacher-student training or knowledge distillation. It works better if you have access to the activations of the model but you can work with just outputs as well.

Does it work even if you don't control the inputs?

Well, it does, that's how we got Alpaca from LLaMA.

I think Amazon was first in the (free) banana business

you joke, but first producy they changed on whole foods were the bananas.

before: organic (south america) and regular (central ou SEA) for 69, 59.

then: both chikita's brand with regular and organic stickers (clearly the same produce, always from SEA) for 49 and 39 cents.

thats was days after the announcement

Did you inadvertently reverse to regular/organic order, or was organic cheaper after?

Google’s been second banana to openai for a few years now, right?

The funny part is that deepmind's tech (and some of Google Brain's research) seems to be as good as openAI's or better, but Google's unwillingness/inability to productionize these systems is keeping them back. It seems like the issue is only with the management, and I'll be looking forward to reading about Google's version of Fumbling the Future[0].

[0] https://www.amazon.com/Fumbling-Future-Invented-Personal-Com...

Many of Google's products were second, or later, to market. Google does not care.

Isn't it generally very hard to be first to market? And even if you are it's more likely that someone coming in later will take your lunch.

Apple wasn't the first one to try to make a successful smartphone, but they had resources, know-how, and tried at a better time with fewer unknowns around.

Or were just willing to adapt when others didn’t. Blackberry mocked the touchscreen for years until finally coming around and their initial implementation was awful.

Blackberry Torch??

That's okay if they can leapfrog.

Google Maps wasn't the first map product. We all used mapquest way before. But Google Maps was technologically advanced. Ajax made maps usable for the first time.

Gmail wasn't the first webmail. Hotmail had millions of customers already. But Google gave people unlimited space to store old email, whereas email in the old days filled up your inboxes and needed to be deleted.

Question is if they can and will leapfrog.

Google Plus was a sign of desperation and utterly failed. The dozens and dozens of different Messengers (sorry I don't even know what the latest one they're pushing is, RCS?) all failed.

As an organization we will see in the coming months and years if Google can still overtake others when coming in from behind or not.

>But Google gave people unlimited space to store old email

This isn't correct. Gmail launched with 1GB per user, which was way higher than other services, and they did keep doubling the storage space year-after-year, but it was never unlimited until Google Apps offered unlimited storage for businesses and schools.

That is true. You're technically correct.

In 2004 1GB of email was effectively unlimited though. Keep in mind hotmail offered 2MB and only increased that 125x when Gmail launched.

https://www.cnet.com/tech/services-and-software/hotmail-to-o...

Ascribing actual human emotion to a giant corporation like a google is probably not a good idea. Their motivations aren't going be heavily dictated by feelings of shame at being late out of the gate.

you talk like chatgpt was some bastion of curated perfectly correct content. get a grip. web scraping is web scraping.