upvote
Yeah, you can’t put anything important to Claude or OpenAI as scientist anymore. And block internet for anyone reviewing.
reply
Use Bedrock....

"With Amazon Bedrock, your content is not used to improve the base models and is not shared with any model providers."

" Users inputs and model outputs are not shared with any model providers."

https://aws.amazon.com/bedrock/faqs/

reply
Bedrock and such don’t address grandparent’s core point though: most humans still collaborate with other humans and that is a weak link. You almost certainly want to at least gut check some ideas with others and do some sanity checks and that’s where the weakness emerges. Those people are doing you a favor and using their own time to do that. In the past that might be pulling up reference books as they read through. Today that might be a copy paste onto an LLM, or large swaths of it.

So unless you operate in a vacuum or you run in an environment that operates in a vacuum, your ideas can easily leak into training data. Now, how likely they are to get surfaced is a question but when you’re working at the very edge there aren’t likely a lot of competing ideas embedded in training data would be my guess. So its more likely such ideas in frontier mathematics might actually resurface easily.

It’s not like write an analysis to… Robert Frost’s Two Roads Diverged poem where there’s countless of these and maybe a unique take might get buried as noise.

So you need to work in off grid spaces really. I work in a fairly information secure environment and I know for a fact people occasionally use their personal GPT account with sensitive corporate information out of convenience. And some may not be aware of what they may or may not be doing and it’ll likely remain that way for several years.

reply
If you have a solution to the Riemann Hypothesis before anyone else does, then no, don't put that into Bedrock, or any other online LLM service that promises to keep it secret. It's just too valuable to take the risk that the promise falters.
reply
I repeat, for the millionth time. You should not trust Amazon. I dont care if you think their reputation depends on it. They do not care about that, they have one of the worlds strongest moats. Nobody is going to stop using AWS because the AI that everyone already expects is stealing or data, is stealing your data.

They do not care, and its ultimately deniable. Also everyone in AI literally offloads all guilt by just saying they got hacked now and nobody thinks otherwise.

You cannot trust Amazon, its wild how many people think you can.

reply
If they respect their terms of service there is no other way to use foundations models than via Bedrock

If they dont, then you can win the biggest settlement in history, and retire as rich as your common billionaire.

reply
You seriously think thats how that goes?

Also pretty easy to deny outright and do it, with hardly anyway to prove other than maybe a whistleblower. I have a hard time believing that Anthropic isnt heavily involved in Bedrock, they arent just sending Amazon their weights files and an instructional manual like its an Ikea coffee table.

I am 1000% sure Anthropic can access your data on AWS Bedrock. (im pretty sure they changed all new contracts as well to state this for auditing "cyber risk")

reply
You make good points, but I would say if their internal controls are good enough for the NSA or NATO: https://aws.amazon.com/federal/top-secret-cloud/

It should be good enough for the GRH.

reply
The NSA or NATO can sue their ass, or spill something about their execs, or inflict very real damage in another way. A maths institute can just complain or cope.
reply
So you are saying they are selective in how they enforce their terms of service? :-)

From the stuff I only know... from watching YouTube...Anthropic trains Claude on AWS: https://youtu.be/KJFnCx0DFsg

Are they not worried their secrets get stolen?

reply
Why would I trust them?
reply
You dont have to trust them, I say trust but verify. But their incentives are more aligned on keeping their terms of service, than giving a freebie to OpenAI and others, especially as they seem to have fired all their AI people...
reply
“Not shared with any model providers” != not shared with data brokers who resell it to model providers
reply
Quite frankly, I don't believe _any_ of the model providers when they say such a thing, AWS included. And the attitudes expressed in the Anthropic IPO thread towards copyright, theft, and training fair use did nothing to dissuade me that anything of value is always fair game, one way or another, for training.
reply
This would be easy to test, with synthetic data, and I hope somebody does an Arvix paper about it.
reply
how many people hold this level of information discipline?
reply
well if the goal is SCIENCE involuntary sharing with competitors is kinda useful right? If the goal however is fame
reply
Stop assuming that its naively in the training set. Its clear to me that Anthropic will maliciously steal. The entire foundation of their business is founded on this practice.

Edit: Why am I being downvoted, I would love for someone to explain how Im wrong here. Anthropic is built on theft. Im convinced that Anthropic has actually deployed bots to downvote negative comments at this point.

Seriously, the downvotes for such an obviously true statement is bizarre. There's zero reason the defend Anthropic ever on this topic. They literally stole the entire world's work, why wouldnt they steal this work.

reply
I disagree with you, and I detest anything that tries to appeal to emotion such as saying "maliciously steal", for the same reasons that I don't like the MAGA crowd or populist politicians.

If you are serious, state your case.

reply
Describing court-sanctioned “fair use” as “theft” is an extreme ideology. You are fine to have that opinion, but you shouldn’t be surprised when people disagree, and the courts also disagree with you.
reply
Courts are fallible and corruptible, calling a common sense understanding of how AI is trained “extreme ideology” is bad faith
reply
So that moves the goalpost to copyright being flawed in general I guess, since the same judiciary decided its precedents.

Maybe we should have had this discussion earlier. Why is there a system that Disney uses to cease-and-desist kindergardens over wall paintings that feature a 100 year old character, whose creator is dead? Why is the law more relaxed (read much shorter protection) in patents than in copyright, despite the former often requiring a higher upfront investment?

reply
Conflating court decisions with ethical behavior is bullshit.

The court approved massive fines and lawsuits against mothers and teenagers over pirated MP3s.

Courts have both said that small individuals should be put in a lifetime of debt but the highly funded companies favored by the current administration are acting in fair use.

reply
deleted
reply
> Describing court-sanctioned “fair use” as “theft” is an extreme ideology.

I could not disagree more vehemently. It's completely fine to disagree with a court decision, especially when so much money is involved.

What is extreme is painting anyone who disagrees as extreme.

reply
Wild for you to call me an extermist.

The courts say a lot of disgustig behavior is legal, especially when it comes to business. Why would I base my opinion or ethics off of US fair use law?

reply
i mean if they did, it would be very easily verifiable to cross-check?
reply
How so, Anthropic hasn't released anything. They front ran the announcement in a very suspect way thay leads me to believe there's even more espionage going on than mere Chat log intrusion.
reply
Their whole MO is stealing IP so that tracks
reply