Remix.run Logo
▲ ozgung a day ago

Paranoia means unfounded suspicion. In this case, it’s completely sane to assume that AI companies will steal your data.

▲zone411 a day ago | parent | next [-]

In November 2026, even publicly available models operating with small budgets are better than top mathematicians at developing proofs. AI labs' internal models, backed by much larger budgets, are significantly better still. It would be extremely silly for those labs to steal any math-related user data. So it is paranoia.

▲bigfishrunning a day ago | parent | next [-]

Whose data did they use to train these models? did they compensate everyone for the data they used?

Big AI has committed piracy on a massive scale, and they've gotten away with it.

▲dash2 a day ago | parent | next [-]

I think the argument "they trained models on publicly available data, I think that's stealing, so therefore I am certain they are going to steal everything else, including training on customers' data when the customers have explicitly opted out of this, violating their contract and permanently destroying their own reputation if they get caught" is not a good one, and should probably stop being made.

I don't even think the argument works if it starts "they committed genuine piracy which was commonly known to be illegal at the time".

▲rcxdude 20 hours ago | parent | prev | next [-]

Just to be clear, the difference between the older models and the newer models in terms of mathematics ability is not from having trained on even more papers, but by reinforcement learning on the act of proving theorems. Anthropic now spends about as much on this training phase as on the 'pretraining' with the masses of scraped or scanned data. This doesn't mean it's not still essential, but it's far from the entirety of the result either.

▲WarmWash a day ago | parent | prev | next [-]

I feel like using the term "Piracy" to describe training on public data is a bit of a stretch.

▲lixtra a day ago | parent [-]

Because public data is the wrong term. Published, copyrighted data is the correct term.

Now we can debate if the model weights are a derivative work of that data, more or less than my old hard disk that learned from a dvd with a ripper and is happy to share that knowledge with the internet over bittorrent. The latter is often described as “piracy”.

▲WarmWash 19 hours ago | parent [-]

It's the not the copying that is illegal though, it's the sharing. I haven't been able to get any free copyrighted content out of an LLM though

▲bigfishrunning 8 hours ago | parent [-]

It really depends on whether an llm is considered a definitive work in regards to the GPL; in my opinion it is.

▲suddenlybananas a day ago | parent | prev [-]

I agree but that's a separate issue.

▲akkad33 11 hours ago | parent | prev [-]

Why would it be silly? Don't they do it with impunity

▲zamalek a day ago | parent | prev | next [-]

Especially given that they already have, and at a massive scale.

▲cubefox 12 hours ago | parent | prev [-]

That's crazy. That would require implementing a huge expensive scanning infrastructure that tries to filter out interesting math claims from hundreds of millions of LLM queries. Finding a needle in a haystack would be far easier. It's completely deluded to think that LLM companies would do such a thing rather than just using their own internal models that are already better than most mathematicians anyway. Not to mention the legal trouble if this ever leaked.