New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Comments
I dont mostly. I also dont censor API keys and whatever.
I pay Google Gemini (incl in my Workspace) and Googles ToS clearly state my chats will not be used for training - This is not a simple sentence, if you violate this the EU will fine you BILLIONS.
So not muc worried.
Other LLMs... yea, well, not as solid ToS by far. Deepseek owns your ass, as does ChatGPT.
Sounds very safe to me
The EU takes 10% of your INCOME (NOT PROFIT) as fines for such cases. So i am not worried.
I did it xD
Things like passwords and API keys should be redacted from the logs in the first place or someone fucked up or incompetent.
The only log I came across with plaintext password in log in decades was from Starwind who requested logs without mentioning removing plaintext password or making private upload. But I've seen enough of their buggy shit that I'll never use their software.
select a word, vscode -> cmd + shift + l , replace word
Only the first tier companies really log anything (OpenAI, Anthropic, Google). The trick is, Claude is on AWS and Google Cloud, OpenAI is on AWS and Azure - all can support a zero logging stance. Just don't send info to OpenAI or Anthropic directly, use their second tier providers (or their public clouds).
Then looking at open weight models, even better "no logging" policies from providers. I have yet to find a provider that just hosts GPU/LLM's that actually logs.
I used to care a lot about this - I looked into all of this - no one cares, and no one wants to care (keeping logs seems like a liability right now). Plus logging this amount of data is going to be hard regardless.
You should still care.
This is sadly true. But wait...I do care. Probably somebody else cares as well.
I mean, no one cares to log. Logging is bad when someone sues you, it's better to say you didn't know, because you really didn't know. Plus logging is expensive at that scale... I just don't see where the money is...
So if you want to log users stuff... I don't know, that feels weird, why?
Bit of an odd assumption given that they obviously log. How else would they know about those "jailbreaks"?
Edit: iirc one of the companies is still under order of a judge in NY to log everything, I didn't follow it too deeply.
Edit 2: it was NYT, federal stuff. Don't care, just worth mentioning that it wasn't NY.
Edit3:
On a related note, how the fuck did they recover the chat history of that dude who sued the bank because of that alleged affair with the boss?
Ah gotcha, misunderstood initially.
Big corpos disagree
they are logging every sing fart and are pretty comfortable with it.
Every word of what you've just said is false. Everyone logs because when something breaks (it will) you want to know who did it, what they did, and how you're legally culpable. If you're big enough, eventually you will work with governments who will expect you to have logs (unless you happen to live in a place without such regulations). Logging is also typically profitable because even data that seems useless is, with a big enough data broker market, useful to someone (advertisers and criminals). I think the safe assumption is to assume everyone logs everything until proven otherwise by a police raid where the cops return empty-handed (like what happened to Signal and Mullvad). Buying more storage is relatively cheap compared to an extremely long legal battle where you now have to prove you're not culpable for something you're lacking logs for.
Delete buttons in webapps almost always never actually delete anything, they just hide it from public view while keeping it on the database. So they didn't "recover" anything that wasn't already there to begin with.
Prompt logging is different from auditing and gathering infra telemetry. This is a discussion on prompt logging...
I've also built that exact infra telemetry - it's expensive to just store, querying becomes insane.
Regardless, you really think Azure or AWS is going to be selling prompts, when they say they don't? Or going into the other providers - most have strict ZDR because they might be handling actually sensitive data.
I do not trust companies to honor their words about data privacy because we've seen time and time again that those words are completely hollow and are proven false for everyone to see when they eventually get hacked. A generic host like Azure or AWS probably lacks the observability to see something as specific as individual chatbot prompts, I was referring more to the people running the chatbots, like Anthropic/OpenAI/Google/MS. Those companies certainly are logging chats for training purposes, and their "anonymization" of that data is opaque to us, so I'll assume they don't anonymize anything, or if they do, they keep unredacted logs somewhere else. The stuff governments/big corps require to be truly ZDR is likely stored on separate infra to ensure no cross-contamination, but you'd have to be silly to think they'd not give up all that sweet, money-making data on everyone else. In every single court case we've seen so far, "deleted" chats always get brought back, so until I see a hack or a police raid proving otherwise I don't see any reason to give them undeserved trust.
I mentioned that. Those are the only companies that log - the point was just use a public cloud instead, the one's that actually operate at HIPAA compliance.
There's no evidence of that. These models are so massive, it would have to be worth millions (billions?), which it's not.
The only actual certification level that I know of that might require separate infra is for DoD/Gov work, which is already in a mirrored public cloud with Google, Azure, and AWS.
All the medical/legal traffic is definitely not in a third public cloud - we would know about this.
But ignoring the logic of scaling to these levels. This would be purely extra engineering complexity with zero upside.
Random data isn't profitable, it needs to be structured data with high quality. Random prompts is not that.
Really, think about this data harvesting, there are some real conspiracy theories here.
Let's assume companies like Amazon and Google really care about profit.
What's the cost to store, say, a year's worth of prompts? Let's say that's a few PB of raw text (I expect a lot more, given how LLM APIs work, lot's of duplication and little edits, but maybe that can be optimized, whatever). Say that costs a few million, easy stuff for these companies. Oh, and you have to pay engineers to do this work building custom software for this all - let's call that a few more million easily.
Now this is unstructured data of no guaranteed quality. No one is going to buy that block of text. So Amazon and Google would have to throw servers at this data to maybe make it structured enough to actually sell to a data broker. More custom software, you are likely hire data scientists - like, this is a real hard problem to solve - make random data maybe be useful to someone?
So, I don't know, say the cost of this logging, retention, and indexing was $20 million per year. Not a lot of money for them, so easy to do.
Who would by this? The companies that would buy this are buying people for a few cents per profile. How much would this massive dump be worth? Basically nothing.
Okay, so maybe not data brokers. Maybe this very rich AI companies, maybe they want it for training data?
And that's just a no - they've downloaded the internet years ago, and that will be the best data they ever get - at least until they make synthetic data to replace it (which, they are, because that is so much better from a legal standpoint).
The next step of AI won't be consuming more data. It will be hiring a lot of people to filter their current dataset, the dataset that is guaranteed to not have AI slop in it, and to make new data that has a guaranteed level of quality.
Databrokers already operate at Petabytes scale a decade ago. At that time
hyperscalers made live demo of their massive data processing capability. These databrokers also has ties to the training data vendor for AI labs.
Distillation is a valid technique and it requires input, intermediary, and output
You can run LiteLLM as a proxy with content masking. Have been planning to set this up but haven't got round to it yet.
Redact everything before sending.
That's a bitLOT naïve for someone who claims he is a sociopath. Dont' you think?
I always thought sociopaths were not naïve but I can see now that I was wrong. You seem to be a little too careless.
You have to be very careful when you deal with industrial scale real sociopaths, sociopaths MUCH BIGGER than you can ever imagine.
Compared to OpenAI, google, meta, and all the LLM guys you are an angel. YOU SHOULD BE VERY CAREFUL WITH THEM!