logoalt Hacker News

Betelbuddy • today at 7:52 PM • 6 replies • view on HN

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/


Replies

Frost1x • today at 8:05 PM

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.

baobabKoodaa • today at 8:03 PM

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.

testtestmetest • today at 8:01 PM

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.

➕ show 1 reply
Phelinofist • today at 8:02 PM

Why would I trust them?

➕ show 1 reply
mcmcmc • today at 8:15 PM

“Not shared with any model providers” != not shared with data brokers who resell it to model providers

jackcviers3 • today at 8:01 PM

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.

➕ show 1 reply