*Do not add AI to the program. Take it off your roadmap, the whole nine. Focus on actual functionality NOT features that compromise privacy and data integrity. Oppose the backslide into the inability of humans to THINK and form their own ideas.
”As soon as we started thinking for you, it really became our, civilization.”
Innumerable relevant minds have tried to warn us. Don’t be ignorant of the implications. AI will not give you competitive edge. The same way an excessively improved calendar will only waste your time in area where you’re already massively behind.*
HOW COULD IT BE DONE?
ᅟRemove it from development until the full program is released and you find find yourself scrounging for new ideas and features.
REAL WORLD USE CASES
None.
RECOMMENDED ALTERNATIVES
AI exists in Grammarly and any other number of external AI assistants that users already have ready access to. Unless you’re introducing an option as secure as Lumo, leave it aside.
An AI provided by a platform that’s at the forefront of data privacy? That never records questions or answers and even the hosting service provider can’t see your conversations?
If you’re going to make that statement, provide an AI model you’d use as an alternative for outright comparison. Which, unto itself is irrelevant. The specific request is to not shoehorn in unsecure AI when even more basic functions are higher on the community’s priority list.
Never thought I’d find myself living in a time when the progressive option was to reject AI and outright recognize that it poses more danger than it mitigates. Not the least of which is compromising a person’s ability to think and search for accurate and reliable information.
models that you host on a machine you control are much more secure and private than a model run in a TEE. We plan to provide all of them, but TEE is the last one in terms of potential compromise. I don’t talk about marketing claims, but about the reality that you can verify.
Running AI locally means the model weights, prompts, and outputs never leave your device. You can monitor and audit the whole stack yourself. Self-hosting on your own server comes next: still under your control, with some exposure depending on how you secure it.
TEEs, like the Intel SGX enclaves used by Proton Mail for search, are useful in cases where you need cloud resources but want to protect against a curious server operator. However, TEEs rely on opaque vendor hardware and centralized attestation services, and history shows they’ve had multiple side-channel vulnerabilities. This makes them a weaker guarantee compared to simply running the model locally or on your own trusted machine.
That’s why in Anytype our priority is local and self-hosted first. TEEs can be offered for convenience in some cases, but they should be understood as a trade-off, not the gold standard. If you care about verifiable privacy, nothing beats running things entirely under your own control.
You know Proton didn’t make their own LLM, right? They are using multiple different 3rd party LLMs in the background to answer your questions in Lumo. What’s end-to-end encrypted is only the chats between your device and Lumo and your chats stored in Lumo, but those same questions forwarded/sent between Lumo and the LLM are not end-to-end encrypted. Even Proton says it themselves in their blog: Lumo security model: How Proton makes AI private | Proton , saying “So we need to provide the LLM with cleartext messages.” So no, your conversations with Lumo are not only known to you, the LLM needs to read them in order to answer your questions and have contexts to the conversation with you. Homomorphic encryption is how Lumo can truly be private, but that idea has never worked out in practice and does not exist in any practical applications, let alone in Lumo. Of course Proton wouldn’t tell you this upfront in their marketing; they have a lot of half truths in their marketing if you look a little longer, and it’s kind of disgusting of Proton as company preaching transparency.
There are a number of AI models that run fully local on your device or self hosted that are so much better than Lumo in terms of privacy, since your questions and data do not go anywhere else.
The only part that I agree with you is that LLM shouldn’t be a priority at this moment before other features, but I have no place to make that decision. In fact, chatting with LLM via the upcoming chats feature is already being tested internally by Anton and the team.
I believe the AI hype is 90% marketing bullshit, but the one thing generative AI is actually useful at is summarizing/rewording text, so it would make sense for a text-based tool like Anytype to add AI.
I also agree with Anton that a private, local genAI would be really nice and that it fits with Anytype’s local-first, privacy-preserving philosophy.
My main concerns are:
That adding AI into Anytype takes precedence over other far more critical work like feature parity with mobile apps, automatically download all files, or cross-Space search and object sharing. This is a little bit like what happened to Chats: it’s a fine feature, I don’t have anything against it per se; but it should not have been implemented before other critical work to make Anytype more usable and performant.
That Anytype uses AI models that have been trained on stolen material (like all the major models). While tech ethics isn’t explicitly stated in Anytype’s mission and goals, it’s been for me a major reason for adopting Anytype. So adopting AI models that spits in the face of every ethical considerations would be a big disappointment.
AI is not our main priority right now—it’s more of an experiment. We’re exploring how it can strengthen our system and what’s the best way to integrate it. When it comes to models, we believe users should have the freedom to choose.
Chats may seem like a detour, but in reality, they’re built on next-gen infrastructure that unlocks faster, higher-quality features—things like transclusions, formulas, reminders, and other top requests that were previously too costly to build. They also align with our long-term strategy, opening the door to new use cases.
The result: better product quality, a more polished UX, faster development, and stronger revenue potential.
Pure personal knowledge management is a tough, unsustainable category. With chats, Anytype is uniquely positioned: the only option for small teams who can’t use the cloud but need a simplified, secure mix of Slack and Notion.
Proton does not utilize any third-party API providers for Lumo’s AI assistant; all AI models are self-hosted in Proton’s own European data centers. The primary models powering Lumo are Mistral’s Nemo, Mistral Small 3, Nvidia’s OpenHands 32B, and OLMO 2 32B from the Allen Institute for AI.
AI is not just a marketing gimmick—the true intention behind massive initiatives like OpenAI’s $500 billion Stargate project, including collaboration and competition with firms like Palantir, is about large-scale data acquisition and the future of pervasive surveillance systems. By offering free AI tools to the public, these organizations accumulate vast datasets that can be used to train even more sophisticated models, enhancing their ability to analyze, monitor, and predict human behavior as part of national security and intelligence objectives. That’s why these services are free: public participation directly feeds these systems, accelerating the development of comprehensive surveillance and control infrastructure.
Do you know Socrates believed writing was detrimental, fearing it would weaken memory, create a false sense of understanding, and prevent the dynamic, interactive dialogue necessary for true knowledge acquisition?
Tools don’t kill thinking. Calculators didn’t kill math. AI removes drudgery. It frees time for judgment and design. If thinking declines, fix governance.
I’m a creative freelancer, I run my own studio. We choose where AI helps and where it doesn’t. We don’t delegate goals, only subtasks. Humans stay in the loop. We set objectives and hypotheses. We define acceptance criteria. We review impacts. Accountability stays human.
The question is not “AI or not.” It’s “where, why, and with what guarantees.”
I actually kept up with Lumo since it was first released a month ago. And I am very aware that they host these 3rd party LLM models on their end. Not sure what you are getting at.
The point that @pandora was making was that 1) they don’t want AI, ever. 2) but for some reason they don’t mind Lumo even though they don’t want AI; and the only rationale I can think of is because they think Lumo is as private as it gets. My comment was to explain that Lumo is not as private as they think it is compared to other solutions mentioned. It’s not a full end-to-end encryption like intra-domain Proton Mail emails, Proton Pass, or Proton Drive where Proton can’t see the text/files. In Lumo’s case the LLM will be reading the messages or files (even those uploaded directly from Proton Drive) in cleartext on Proton’s server.
This is why Proton says “conversations are saved with end-to-end encryption in your chat history” rather than just saying something straightforward like “the chats are end-to-end encrypted”, because they can’t. It’s fine if people are okay with that, but most people don’t know this, and this is one of the examples of half truths by Proton.
Sorry, that really wasn’t my intention to come across as attacking. I’m mostly just trying to confirm and verify what people are saying, especially since on Reddit, a lot gets said without much verification. I might’ve come off a bit too harsh here—my bad!
I’m biased, so I agree with removing AI as it doesn’t contribute anything to my workflow and appears as just a security/privacy concern which is what AnyType is supposed to address. The notes I take are already summarized, already organized where I want them to be, and I know where to find them if I need to refer to them. Polishing and adding the core fundamental features is more important and more valuable than introducing AI at this point in time.
If AI can be implemented as a local, self-hosted option though then I wouldn’t really care otherwise. But if it is going to be a 3rd-party solution (ex. using OpenAI’s API), I would strongly oppose it and probably even demand an option to disable it altogether. I’ve dabbled in it a fair amount in Notion and Capacities, and its impact is minimal.
I guess it just depends on how complicated one’s workflow is to even need AI in the first place
Neither would I. Bad option. It was just the most secure thing I could think of at the time.
Edited post. NO AI. Period. Not in a privacy focused program.
As another measure, if someone writing a sensitive project that needs proof of provence can use a program where no AI can be added to it, ever, that’s one more boon for the program.
Bingo. I just ran into this with a writer having to prove the provenance of their work–that it wasn’t AI generated.
A program where it can be CERITFIED that AI was never used to create any of the space’s content, perfect. Love it. Makes proving the provenance of my work that much easier.
Ai is a problem. It’s not a solution. To say nothing of how destructive AI is to the communities whose water and power supply these data centres are jacked into. They destroy whole communities and when a town is approached with a proposal the town council is often forced to consider approving the project under an NDA so they can’t even VOTE on whether or not they want something built nearby and attached to their supply infrastructure.
Ai is bad for the planet, it’s bad for our ability think for ourselves, it can compromise an academic or author’s claim to authorship of their work.
for recognizing it!