It's totally valid if the models want to pack words tight during their thinking process, as long as the final conclusion (which is the interface to user) is written in HUMAN LANGUAGE, then I don't care whether the model thinks in alien language
In 2026, applications, third or even first party, don't need to have full-disk access, and are not given either. They see a jailroot environment. I give full disk access to the terminal app, and a handful of others. 90% of them, nope.
At least that's the case in macOS, I'm pretty sure Windows can do that too. Linux of course has had such capability since forever, but I guess most distros you need to manually take care of it.
This is implemented the wrong way around. Each program should only have access to its own folders by default, with it being possible to grant additional access. Also, I don't believe Endpoint stuff is included in the normal Windows license.
Maybe it isn't built-in, but most Windows user I've worked with, including myself, have been using Sandboxie for probably two decades at this point, probably hard to find any Windows software that is more ubiquitous than Sandboxie in developer circles.
Sandboxie is essentially a giant pile of fragile hacks on top of a Windows API that does not want to be used this way. Does it seem like it works most of the time? Sure. Has it had bypasses? Also yes. I've used it in the past but I don't truly trust it.
Yes you can sandbox Obsidian on the OS. The point they're making is nearly every third party program ships Without sandboxing. There's nothing special about Obsidian here.
This hardening is enabled by default with Gatekeeper. That includes Homebrew apps, unless you disable it.
When an app tries to access something outside of its sandbox, you get a notification asking to approve or deny. Full Disk Access I think needs to be explicitly given on System Settings (Privacy & Security -> Full Disk Access).
That's probably all the hardening the average person needs. BlockBlock because most malware tries to get persistence. Little Snitch or LuLu for fine-grained whitelisting of network requests for any apps that have plugins (e.g. you give Documents permissions to Obsidian, plugins inherit that, but they can't exfiltrate if you only allow requests to trusted domains).
I've never tried to do this or similar in Windows (obviously easy in unix-like environments) but I'm going to bet it's far more trouble than it's worth for 99% of users
On macOS at least those 99% of users are probably installing from the App Store, where apps are sandboxed by default and need to explicitly ask for access to paths outside that sandbox. Even when not installed from the App Store a permission dialogue is popped if an application tries to read from sensitive paths like your photo library.
Does that help in this case though? I think the worry is that a rogue Obsidian plugin does bad stuff with your Obsidian vault, not just do stuff to the rest of the computer. But that vault/those notes live in the same sandbox as the (rogue) 3rd party plugin, which doesn't help with that, they really need to be isolated away from the notes themselves.
Anything that reduces the blast radius helps. There should still be a focus on further hardening. Most value comes from exploits that enable pivots. Attackers will focus on other vectors that enable broader pivots because immediate high value notes only exist for a limited set of users.
Broadly, I think other open source solutions are lacking in (1) integration of external knowledge into the chat (2) simple UX (3) complex "agent" flows.
Both internal RAG and web search are hard to do well, and since we've started as an enterprise search project we've spent a lot of time making it good.
Most (all?) of these projects have UXs that are quite complicated (e.g. exposing front-and-center every model param like Top P without any explanation, no clear distinction between admin/regular user features, etc.). For broader deployments this can overwhelm people who are new to AI tools.
Finally trying to do anything beyond a simple back and forth with a single tool calls isn't great with a lot of these projects. So something like "find me all the open source chat options, understand their strengths/weaknesses, and compile that into a spreadsheet" will work well with Onyx, but not so well with other options (again partially due to our enterprise search roots).
Looking forward to your progress! Just checked the paper and it says the underlying backbone is still DETR. My guess would be that SAM3 uses more video frames during the training process and caused the dilution of sparse engineering-paper-like data.
1. The chat context is always provided, and that introduces a bit of uncertainty - when the chat history mentioned something the model is always inclined to connect with it.
2. When I tried to set each context to an empty string, the model doesn't show any evidence of remembering concepts. I told it 5 times that I love cats, and when asked about its favorite animal, its output remains "honeybee" and "octopus".
I can’t decide if I’m skeptical of the entire concept or not. I guess I believe it will do something to the network to add this EMA of vectors in, so I’m surprised you didn’t get at least a change in animals after talking about cats. But, I’m not clear that reweighting logits at the end is super useful. I guess this is supposed to be in some way a realtime LoRA, but then what do you have except a super-undertrained LoRA, trained just off whatever conversations you’ve had?
Looks pretty good but all CJK characters are displayed as questionmarks (???) and switching to a CJK native font does no help. So my user folder is now crowded with ???????? which makes it hard to navigate :(
