Exactly. All it takes is one clueless employee to use that data lake by mistake while training. The thing with these models is you can't untrain just 3% of the input, you have to start from the beginning

Unless they can promise its never being reviewed/saved anywhere, it will be used for evil eventually, intentionally or not

> The thing with these models is you can't untrain just 3% of the input, you have to start from the beginning

You think they don't checkpoint the model before feeding in a new slug of data?

That would seem...unwise.

On the other hand, the history of IT is rife with people doing unwise things, so it could be true.

> You think they don't checkpoint the model before feeding in a new slug of data?

They definitely do that otherwise they can't rewind the model when the loss shoots up. Unstable training happens to almost all LLMs, but can be managed by rewinding & skipping a few batches.

And if the first "slug of data" contains something you need to remove? If you need to remove something you would need to roll-back to the last checkpoint that didnt include it.

I mean, sure, but the point of frequent checkpoints is that these issues would be rooted out sooner rather than later.

It would have been caught far sooner than “we have to redo everything from scratch”.

Well, that's why you checkpoint a lot and feed in the new training data a little at a time, rather than dumping in a massive slug all at once.

Right?

Not really. If they are ingesting peoples data to train their model, at what point are they checkpointing? Is data I submit 2 months ago that I would want removed not present in a recent checkpoint or would they have to "unwind" the last 2 months of training to remove it? What about people that would want something removed 6 months prior? What if something in the core model is to be removed?

Checkpointing may help, sure, but it isnt going to allow you to remove a single piece of training data without headache, and potential substantial retraining.

If that 3% from the GPs comment comment is uniformly distributed throughout the training history, the only way it can be reliably removed is to retrain from scratch.

> If they are ingesting peoples data to train their model, at what point are they checkpointing?

Multiple times per day if they're not incompetent. And I don't think they're incompetent.

Are you implying that removing the diff between checkpoints achieves the same effect? Ive never heard of this, but I suppose it may be possible.

I suppose the "ghost" of the removed weights would also have shaped subsequent training though...

Interesting idea...

Microsoft enters the chat.

They'll host a managed OpenAI model in Azure for you.

I may trust that slightly more. Or Amazon’s bedrock. Or Github. But wait, do I see copyrighted code on GitHub Copilot, which is owned by Microsoft?

My background is in building distributed systems for self-sovereign ownership and make them easy to use and available to everyone. Like https://qbix.com and https://intercoin.org

You should have the software infrastructure of Facebook and Twitter but choose where to host it.

In addition, I prefer Wordpress, Discourse forums, GitLab to GitHub, Redmine to FogBugz, etc.

When it comes to YOUR art, your code, your content, your relationships, you should be able to run it on your own servers.

Any analysis of your data should be done locally, with local models. You should have the open source software and the weights, and you should be able to choose which hosting company to trust, or host on-prem. Villages should be able to do this without needing server farms in California. This should be obvious stuff. But we the people need the software!

But hey, the documentation and teasers and trailers should go on YouTube and TikTok.

It isn’t even about them scraping all your content and training on it. It’s about not giving Twitter all your followers and YouTube all your hours of video production and content so they can give you pennies for being “an influencer”. Own your community!

Someone had to build it, and in the Web2 space nearly everyone sold out to venture capitalists. I was sure that in the 12 years we built Qbix and 5 years of Intercoin someone would make a better open source alternative to Big Tech and Big Finance. Nope. They all either sold out, or have a solution that doesn’t compete on features (eg Mastodon). I would say the closest is Matrix!

I mean, are you suggesting (for example) that someone who wanted a career in short-form video essays should skip YouTube and host their own videos? I don’t see how that could possibly work in general.

You don't see how, say, someone who wants to teach a class and collect tuition might not want to put all their content on YouTube and get a pittance?

And now replace teaching with pretty much any other content. A musician giving a concert, TED talks, etc.

Well, the concern I have is about alternatives. Joanna Videoessayist might not have the technical know-how, the business acumen, or the capital to build her own platform — she just wants to make great video essays. Eliminating platforms like YouTube wouldn’t make Joanna better off, it would just cause her to return to her marketing job.

No one is eliminating YouTube. We're making a better open source alternative.

If it's good enough, she will leave YouTube just like her mom left AOL and embraced the open Web. Why did content creators leave MSN, CompuServe, et al ?

They should treat youtube as another marketing channel (as a showcase on the street is), but not as their main source of income or backup for example.

- They can run a website with their own personal brand and collect emails there. - They can backup their videos in other services in case youtube close their account ...

You mean you don’t trust Sam Altman, a man who explicitly changed his company to do the opposite of what it was founded to do, to keep his word?