Post

related comments from Alpoge:

These issues are precisely the ones answered by paying for an Enterprise account which will assure you - at least legally speaking - that your data / interactions are not shared or used for any training. However the average consumer if not going to pay for one of those so all bets are off.

This is where provenance splits in two: what evidence shaped this run, and what user data shaped the model behind it. We can make the first auditable in an agent workflow; the second is still mostly opaque. That boundary matters once the work has real value.

Realistically it would benefit them more to use the data and lie about it than it would be to not use it. There are currently zero repercussions in this industry regarding stealing someone else's work. In-fact, the core premise of LLMs is training off stolen data.

Good point. But isn't that the easiest and best way for a frontier lab to gain access to extremely high-quality data? And how could one prove that the data was used? (Besides, who could even afford such a lawsuit?)

The issue there is that we know they aren't trustworthy because they stole all of the training data. The damage will have already been done by the time they disclose, if they disclose, and for that to happen, most of the time, you would have to be colluding with your competition to learn about it!

First it is TOS. Then dark patterns. Then outright lies.

To be honest, I don't think that competition really works in this area. Well at least not regarding this dimension. The only real distinction between frontier labs seems to be cost of inference. Privacy, environmental impact, collab with agencies etc. seem to bother nobody.

But ALSO - now the models ask to import my browser cookies. Soon they will slurp my address book. They save my IP address and geo-locate me. Etc. "Low"-tech mining too. And this will become more prevalent as they are pressured to figure out network effects and lock-in. Do not let's be naive!

discrete parametrization of the process described by navier-stokes equations is the art

I would assume they use LLMs in addition to regexes

It does help that API keys and secrets benefit from being regex-discoverable for auto-invalidation by GitHub among others and many providers make their keys trivially identifiable for secret detection as a result.

Regarding the brainstorming question: There is another angle, and this is not about competition but the model provider itself. How big is the likelihood that the model provider will use my data to claim results for him as a PR stunt claiming my intellectual achievement as GenAI model capabilities?

Another apparent instance. Kind of like Apple copycat apps but much worse. bsky.app/profile/gro-...

swillison Gmail, but I have a very poor email track record! Ping me here too if I seem to have missed something

Another practical angle is "if i consume my tokens through a third party like Google Vertex or AWS Bedrock, do I have the same uncertain exposure than if I subscribe to the provider's consumer APIs?" People might well be voting to have more exposure to that in exchange for cheaper tokens.