alt Hacker NewsPool together with some friends and buy an H200 or two to run unquantized open source models with abliteration/heretic transformations.
You need to be able to use these models for the real world and not for some imaginary world where everything is safe and nice and happy all the time, while at the same time intensely surveilled in the name of CYA and the latest panic about whether speech THAT ISN'T EVEN BETWEEN TWO PARTIES is considered "wrong".
I'm a free speech fan that acknowledges there are lots of boundaries of free speech (fraud, perjury, blackmail, defamation), but the one thing that all of the boundaries have in common is that a second party must be involved for them to make any sense at all.
Maybe the courts will uphold this, maybe they won't, but don't take the risk!
Replies
Gotta find some rich friends in order to buy and host a H200 or two.
A middle road is to use open-weight hosting providers, maybe non-US ones if you’re in the US.
This is not even an option in the UK. Communication felonies (dangerous speech, threatening behaviour) only demand potential audience.
You can trust another party and do this by renting a few H200s. You cannot pool together with some friends without trusting another party.
You end up just weighing up the difference in trust between a vendor and a friend against the level of disinterest that they might have in your affairs.
Would be better if your friend is lawyer and take $1/yr(or higher) for the service so that attorney-client privilege can be applied.
edit: ah, future crime cannot be protected.
That's definitely something I'd consider if I had cash to spare for H200's! Unfortunately I think for most of us the price of self-hosting has to be 2-5x lower still.
Preach! I am so tired of Anthropic’s safety team being the arbiters of what is right and wrong. If you so much as hint at impropriety you can have future sessions flagged ad infinitum when it comes to specific topics.
You don't need an H200 for that kind of use case.
A cheap second hand 10ish year old card like my radeon rx570 with 8GB of ram is plenty enough to run a small uncensored model with llama.cpp if all one wants is chitchatting with a clanker.
We are not talking about heavy coding use cases here.
Way ahead of you bud, mac studio m5 ultra 256gb version is coming soon.... ordered 2x of em just in case. expensive as fuck but its a hedge against all this bullshit and more
Lead time is like a year tho no?
I agree, but someone will say fake child porn.
I'm not saying fake child porn should be allowed or not-allowed, just showing there exist possible exceptions and rationalizations for them even without two parties.
It’s been quite exciting that Qwen 3.8 Flash Next has come out: it really is similar to Opus 4.5, 4.6 for coding. Remarkably intelligent, and runs on a single DGX Spark, which I paid $4000 USD for
Sadly they’re now twice that price, which is a shame because I really want a second one!
I have a heretic modified version of it too, for when I want to use it for security and so on. Quite interesting