I appreciate your thorough reply but I really think you’re contradicting yourself in several places here. Prepare for a long reply.
Having a sustainable cloud provider (like Hetzner) vs a random no-name 2 year old startup is definitely not a 
vs 
comparison.
The age of a company doesn’t automatically make it more or less trustworthy. By that logic we should never trust any new privacy tool or service and just stick with the old giants forever. Mullvad was “new” once. Proton was “new” once. Pretty much project Privacy Guides recommends was a “random no-name startup” at some point. Age doesn’t equal trustworthiness, what matters is what they actually do. Dismissing something solely because it’s young is not a privacy argument, it’s an appeal to tradition. I thought that was something we are actively challenging here 
You’ll never know if they do snoop or not anyway
So your argument is essentially “trust nobody, verify nothing, just give up and use a VPS”? Because you literally just admitted you can’t verify what Hetzner does with your traffic either. So how is that really different from trusting a provider that has actually published their security architecture and privacy policies? At least some of these AI startups are publishing security docs and opening up to scrutiny. Hetzner isn’t exactly letting you audit their hypervisors either. Not that this matters as Hetzner isn’t an option anyways but more on that later.
the AI landscape is a wild west and nobody cares about your privacy
“Nobody” is doing a LOT of heavy lifting in that sentence. TEE implementations, confidential compute, E2EE approaches like Confer’s passkey-based encryption, zero-data-retention API policies. Afaik these are real, technical, verifiable privacy mechanisms that exist right now. Maybe I was also just drugged and am hallucinating. They’re not perfect, sure. But you know what? Nothing is. But saying “nobody cares” just because the big players are bad actors is like saying “nobody makes a secure phone OS” because Samsung and Xiaomi are awful. GrapheneOS exists. And in the AI space, privacy-focused approaches exist too. You just have to actually evaluate them instead of blanket-dismissing everything.
Don’t be fooled, Microsoft, Google, Anthropic or anyone else really, do not care and are equally just awful. They are all lying and equally bad
This kind of blanket nihilism is actually counterproductive to the privacy community imo. If everybody is “equally bad” then nothing matters and there’s no point evaluating anything. This is exactly the attitude that benefits the worst actors the most. There ARE material differences between companies. Lumping together a provider that uses TEE + accepts crypto + publishes no-log policies with OpenAI (which is literally court-ordered to retain all chat logs indefinitely) is not useful analysis. It’s doomerism dressed up as skepticism.
Not everybody needs a befy model for their needs but if you do have enough of those needs, then going open model + VPS (if not self-host) is still the most sustainable and privacy-respectful approach by far.
I actually agree with this in principle. But the keyword is “if you do have enough of those needs.” Not everyone does, and not everyone has the technical skill or budget to spin up a VPS with GPU passthrough. For those who need something more capable than what a small VPS can run, dismissing every cloud option as “equally bad” doesn’t help them make informed decisions. It just leaves them with zero actionable advice.
And about the “just use a VPS” suggestion. Let’s do the actual math on that. GLM-5, lets just call it the current open-weights SOTA, is a 744B parameter model that needs ~1.5TB of VRAM at full precision, or ~241GB even at aggressive 2-bit quantization. Hetzner’s best GPU server (GEX131) has 96GB VRAM and costs €889/month. Afaik you can’t split LLM inference across separate physical servers, so you’d need a specialized GPU cloud with an 8x H100/H200 node. We’re tlking $15,000-25,000/month, not exactly the “affordable VPS” you’re casually recommending. And if that is affordable for you then I seriously want your salary. You could run it on CPU with RAM offloading on a high-RAM dedicated box, but at 1-2 tokens/second it’s essentially unusable. And sure, you can run a 7B or 13B model on the €184/month GEX44, but then we’re back to what you yourself called “quite unreliable for basic tasks” because small open models are exactly that. So the realistic choice for most people who actually need capable AI for work is: spend thousands on hardware for a self-hosted setup that still can’t match frontier models, or use a cloud AI service with the best privacy practices you can find. Dismissing the second option as “equally bad” doesn’t help anyone make that decision.
Are those your thoughts? Or you starting to think it now that those companies shoved enough ideas into your mind with lots of marketing?
Come on mate. That’s just condescending. We don’t need any of that here. Yes, those are my thoughts. I use AI for work daily. Not because marketing brainwashed me, but because my workflow genuinely benefits from it and in some cases it’s pretty much required to reach goals. You can dislike the AI industry (I do too in maaaaany ways) without pretending that everyone who finds it practically useful has been manipulated.
And, I say this respectfully, you’re telling me all these companies are “equally awful” and that I’ve been brainwashed by marketing… while you stream on YouTube, speak at Google tech conferences, and run your whole streaming setup off a Mac Studio? You’re actively participating in Google’s and Apple’s ecosystems — two of the companies you just called “equally just awful”. You’re even considering running Cursor (an AI-powered code editor) on your streaming rig (which kinda tells me you use it on another computer) while telling me that thinking AI is useful means I’ve had ideas “shoved into my mind” by marketing.
And then there’s the contrast with how you treat startups. You dismiss Chutes as a “random no-name 2 year old startup” that can’t be trusted, but then you shared shared urban-privacy.com which is a company selling anti-facial-recognition clothing with zero testing data, zero peer-reviewed validation, and crazy high prices that other people rightfully called out. And your response was “let’s be patient and not kill them already” and “let’s assume that their intentions are honest.” So new startups deserve the benefit of the doubt and patience… unless they’re in the AI space?
Burning an entire forest to know what’s the temperature outside
Great metaphor honestly. But the solution isn’t to pretend forests don’t exist or that nobody should ever check the temperature. The solution is to find the most efficient and privacy-respecting way to do it. Which is literally what I tried with this thread before it turned into a lecture about how nothing matters because everyone is bad.
The “everything is equally terrible so don’t bother evaluating” approach helps nobody. It’s the privacy equivalent of “don’t vote, all politicians are the same.” Real threat modeling involves nuance not nihilism.
, rebrand under a new name and milk some people elsewhere with some over-the-weekend “product”.