There are a lot of problems solved by tiny models. The huge ones are fun for large programming tasks, exploration, analysis, etc. but there's a massive amount of processing <10GB happening every day. Including on portable devices.
This is great even if it can't ever run Opus. Many people will be extremely happy about something like Phi accessible at lightning speed.
Positive results in mice also known as being a promising proof of concept. At this point, anything which deflates the enormous bubble around GPUs, memory, etc, is a welcome remedy. A decent amount of efficient, "good enough" AI will change the market very considerably, adding a segment for people who don't need frontier models. I'd be surprised if they didn't end up releasing something a lot bigger than they have.
The lack of integrated sandboxing in windows compared to android/iphone is still frankly unacceptable. I've become increasingly paranoid about running any application on Windows (not that your average linux distro is even remotely better) and yet Apple and Google seem to be far, far ahead in user permissions (especially with GrapheneOS, god bless that team) and isolation of processes.
Consumers and businesses deserve better. It's crazy to me that in 2026 Notepad++ being compromised means as much potential damage as it does, still.
The sandboxing on mobile platforms puts the OS vendor in a special position to enforce a monopoly on apps and features. Apple enforces it aggressively, while Google only reluctantly so far. It also prevents the user from exerting full control of the system. Apple does it by locking things down directly, while Google punishes you for owning your devices with attestation.
There has to be a better way. I think Linux's flatpak is a reasonable approach here, although the execution might be rather poor. I want a basic set of trusted tool that I can do anything with, and run less trusted tools like GUI programs in sandboxes with limited filesystem access.
Those are policy decisions not really connected to the sandboxing technology. They control what sort of signing the system will accept and make it so that it only runs things they approve, and they only approve things that are sandboxed a certain way. The exact same sandboxing could be used with a system where an admin user can decide what gets to run and what kind of sandboxing is required for each thing.
> I've become increasingly paranoid about running any application on Windows (not that your average linux distro is even remotely better)
Linux excels over Windows in the area of security by a wide margin, I have no qualms about running an app on Linux versus Windows, any day of the week.
No, Windows has consistently been ahead of Linux for many years in terms of average-user desktop security, from binary hardening to designs like secure desktop, because average Windows users do not typically have curated software selections, so you assume the worst. (When I wrote the original "binary hardening via compiler flags" RFC for NixOS over 10 years ago, almost everything in it was already done on Windows and had been for years.) It's still not ideal; macOS takes it even further and actually allows things like "storing secrets on disk in a way that can't be read by random programs" because it can e.g. make policy decisions based on code signatures, which are widely deployed. None of this exists in pretty much any Linux distro; you can literally just impersonate password prompts, simply override 'sudo' in a user's shell to capture their password silently, copy every file in $HOME/.config to your evil server, setuid by its very definition is an absolute atrocity, etc. Linux distros make it easy for people to live in their own chosen curated software set, but the security calculus changes when people want to run arbitrary and non-curated software.
You can make a pretty reasonably secure Linux server by doing your homework, it's nowhere close to impossible. An extremely secure server also requires a bit of hardware homework. The Linux desktop, however, is woefully behind macOS and Windows in terms of security by a pretty large margin, and most of it is by design.
(In theory you can probably bolt a macOS-like system onto Linux using tools like SCM_RIGHTS/pidfds/code signatures, along with delegated privilege escalation, no setuid, signature-based policy mechanisms, etc. But there are a lot of cultural and software challenges to overcome to make it all widely usable.)
> Linux excels over Windows in the area of security by a wide margin
No, this is wrong but might be true if you are talking about Linux package manager vs. Random Windows .exe on internet. But if you are talking about Secure Boot, encrypted disk, sudo etc. Windows is more secure but it looks like https://amutable.com/ will make Linux more secure like Windows.
Edit: Some insecure things on Linux: Dbus (kwallet etc.), sudo, fprint, "secure boot".
And executable you run has access to any file in your home directory, including SSH private keys, secrets in config files, browser cookies, passkeys—all of it. That includes the thousands of npm modules installed as a transient dependency of at least one tool you use that brings node as a dependency.
Windows at least has a proper ACL system; on Linux it just takes a single compromised executable to loose everything.
Threads like this one make me feel at home. Last night I spent an hour trying to figure out a way to adjust tailscale to allow me access to containers on a MacVLAN on my NAS when I connect in away from home. Claude's an excellent tool to help me make informed decisions. I find the knowledge needs to be double checked more than some domains (I'm a big fan of requesting Claude search online for information before using its discourse as a basis for any decisions) but I still feel like I'm learning the WHY and HOW because I can still ask.
I share a lot of the same hesitations as others in the thread - using a giant US-based tech giant's tool for research as well as another US giant's tool to manage access, but it's really a game change and I'd be unable to find the time to do everything I want if I didn't have access to these otherwise.
I'm not even a software guy by engineering, my network is already complicated enough that learning and correctly securing things otherwise would simply just not be feasible with the time and energy I'd like to dedicate to it.
I set up tracking all my bank statements, pay slips, Costco (and other multi-category store) receipts, and other things in Fava (beancount's webui) over the last month. I spent way too much time developing a plugin to help categorizing transactions, making a full api and python scripts as well to let AI make edits for me (dry runs before clearing).
I stumbled upon Maybe (https://github.com/maybe-finance/maybe) and the community fork (https://github.com/we-promise/sure) a couple days ago and almost invested too much time to migrate to it. That project still doesn't have split transactions. A fresh intuitive UI is great but Beancount's format is so simple it's hard to beat.
I started writing code using LLM for this purpose and then stumbled upon ‘sure’. I did use ‘maybe’ couple of years ago and it had few issues and I never bothered.
Last couple of days, I tried and moved few of my accounts to ‘sure’, so far it is not bad. There are few UI/UX issues and the entire investments working in silo and not viable on your main page. But the good thing is, developers are actively working on fixing those issues and improving them.
AutoLISP is still my most fluent language, pleasantly surprised to see anything in it on HN. There's something fun about its idiosyncrasies, but I am genuinely so glad for modern IDEs, linters, tooling in just about every other environment. AutoCAD has severely neglected it, despite some large businesses built exclusively upon it.
Still your most fluent - that says something about how the language shaped thinking. The neglect is real; Autodesk clearly wants everyone on .NET, but there's a lot of institutional knowledge and working code out there that just... works.
Part of why I built this: preservation. If AutoLISP fades from AutoCAD entirely, at least the workflow can live on in the browser.
I cargo culted my way into AutoLISP in the R12 era knowing nothing about lisp or funtional programming (BASIC and a tiny bit of Pascal was all I'd done by then). Just using notepad without any assists like highlighting matching parentheses and no deeper theoretical knowledge was tough, but I could see vague outlines of a world of mathematical elegance just out of my grasp.
"Vague outlines of mathematical elegance just out of grasp" - that's exactly it. Notepad, no paren matching, just counting brackets by hand. The elegance was there, we just couldn't quite see it yet.
I've been transcoding my media collection leaving my PC on overnight over months, it's great. My biggest issue is client support for native playback of AV1, naturally.
For what it's worth, AB-AV1 [1] is a pretty awesome tool written in rust which compares random samples from a file at different parameters based on their VMAF score [2] (algorithm from Netflix for human-perceived visual likeness), choosing optimal parameters to save as much space as possible with the loss you're willing to stomach, on a file-by-file basis.
Small plug: I made a nice little python GUI wrapper for ab-av1 [3].
IIUC, it's more about the client hardware that determines ability to play without transcoding. You'd have to check the mix of devices you have connecting to it and make a judgement call.
First in-the-wild reference I've seen to some of my favourite books. I feel your watch as Goblin makes more sense if it's stuck around with you for a long time and generally works but is a bit of a pain to use. Thanks for the share.
On the normal food blogger pages it is fluff to make room for ads or to get people to misclick on the ads while trying to get to (or back to; the scroll changes feel intentional too) the actual recipe bit. Super effective, especially with people using mobile and trying to scroll with one finger while cooking.
A ton of these sites were bought up and they all look the same now and run the multiple floating ads, especially a wide banner at the -bottom- which is a perfect misclick monetizer.
I thought it's because there's no copyright possible on recipes, but there is on the fluff writing...
I don't know if it's to prevent mass theft of content (oh no, the thieves would have to go through the text and cut off the stories. Although nowadays the biggest IP thieves have built systems to automate this..)
Show me something at a model size 80GB+ or this feels like "positive results in mice"
reply