herf

Hang on, this is a $2.99 one-time payment - below the level you can even buy ads and make a profit. There is no way he's trying to make billions of dollars this way, and it's honest and smart. Consider the perspective - what happens with the "free alternatives"? You know they're not free: either they already track you, or someone buys them and turns them into spyware. We need more things like this, not less.

The aspect ratios were much "taller" back then, which was kind of better for editing code. All these late 90s designs were near NTSC at the time - aspect ratios like 1.25:1 (1280x1024) or 1024x768 (1.33:1). Monitors have always followed TVs, since displays now are the "HD" ratio of 16:9 (1.77:1), or 16:10 if we're lucky. But we do get way more pixels now anyway.

If you back up to "intention" it's fully insane to make a GDPR argument against on-device AI. Yes it downloads bits, but those bits are not there to identify you - they are basically a local copy of the internet. This enables private data to be kept on-device. Having no personal data leave the device is fantastic for GDPR compliance.

The good point in this article is about how the "AI" features in Chrome all use Google's cloud API and not a local model. That's true and some of it should be local. ("AI mode" uses the Web index, so it fundamentally cannot be local, but there are features that could be.)

Chrome seems to use a custom inference runtime also (in addition to Gemini Nano). It would be better if this were all interoperable. The WebGPU alternatives like WebLLM do not have the same access.

I've been trying these models out for the last year, and it seems to me that we want them to work in a 5-10W "laptop" power envelope, but they really work best with a 50-500W GPU instead - i.e. they eat batteries. This means things work better in a "plugged in" gaming laptop/desktop rather than a typical web client. At least for now.

Any benchmarks? Scrolling is using 16% of my GPU (vs 5% in Sublime). Also things like mouse-down to activate a tab (vs mouse-up in Zed) make Sublime feel faster.

Doing this without the parents on board does not work. Kids can lie about their birthdate by a few years. Facial age estimation has error bars of like 5 years and many teens don't have any ID. Younger kids use a parent's phone. Many are not supervised by parents or have parents who are complicit/encouraging in getting them more access. Oh you could be famous! But it is clear that more persistent identifiers online will make anonymity much more difficult for everyone else.

for the youngest ones, a lot of these are "mom's phone" or something like that, it's not even accurate to say you are identifying the user

I think HiDPI is another reason - these were from the "96 DPI" era where pixels looked the same on every screen. You can draw all your pixel art at 2x (or 3x) and scale it down at load time, but it's not super easy. Also, some of the RAM usage of modern apps is the need for a full backing store for each window - in the "true" win32 days like Windows 95, XP, or Win7 in classic mode, you'd be drawing directly to the front buffer, with no extra RAM/VRAM usage per pixel. Of course it flickered and looked bad, but it was fast and cheap.

Both Dropbox and OneDrive default to "online first" for most users (including Dropbox on macOS which has moved itself into File Provider). It is a technically sound and sane default for Backblaze to ignore these mounts, especially given their policy not to backup network drives. They really should have informed legacy users about it.

Technically speaking, imagine you're iterating over a million files, and some of them are 1000x slower than the others, it's not Backblaze's fault that things have gone this way. Avoiding files that are well-known network mount points is likely necessary for them to be reliable at what they do for local files.

It's important to recognize that these new OS-level filesystem hooks are slow and inefficient - the use case is opening one file and not 10,000 - and this means that things you might want to do (like recursive grep) are now unworkably slow if they don't fit in some warmed-up cache on your device.

To fix it, Backblaze would need a "cloud to cloud" backup that is optimized for that access pattern, or a checkbox (or detection system) for people who manage to keep a full local mirror in a place where regular files are fast. This is rapidly becoming a less common situation. I do, however, think that they should have informed people about the change.

This is an ideal place for LLMs to run (is this changelist a security change or otherwise suspicious?) but I don't think the tokens will be so expensive. For big platforms, transit costs more money - the top packages are something like 100M pulls per week.

There are a lot of details in the original article, in most cases comparing with Opus, which required "human guidance" to exploit the FreeBSD vulnerability:

https://red.anthropic.com/2026/mythos-preview/

Also "isolating the relevant code" in the repro is not a detail - Mythos seems to find issues much more independently.

We need a high quality compression function for human readers... because AIs can make code and text faster than we can read.

Apple makes a nice distinction between their "app layer" (iCloud drive and Messages, etc.) and the OS login. This would work fine for Windows power users, and for the most part Windows has already had this (your "store" login). But to require the cloud to replace your login, the cloud has to be essential to the functioning of Windows and you have to explain the security implications clearly, and it's not clear that any of these things happened.

For instance, almost none of the useful settings from win32 apps sync - migrating to a new PC is painful, your apps don't move, your settings are all missing. It takes weeks, you don't just login to it. So this idea that it makes all your settings sync is maybe 10% true.

The argument for this online account (vs just a container for apps) is that you think a few Windows appearance settings must be synced always, or that you want to save things like your BitLocker keys in the cloud (which probably makes them visible to FBI or whoever else). And the security implications need to be spelled out in plain language. And in the end, it's a pretty bad argument - Grandma doesn't need BitLocker, but the people who do want a clear explanation. A lot of the rest could live in a "Microsoft apps" credential layer: Edge, OneDrive, Office, etc.

I want to feel like I can login to a recovery console and fix a bad partition. I want to keep using the same username across Linux and Windows. I want to recover a router with the old laptop that has actual ethernet, and who knows if it has cached credentials? My Microsoft account is my least used one, and who knows if it is secure?

One last thing: logging in with biometrics is amazing, but why must I use a low-security PIN in place of your pre-existing password?

Please fix it all.

This "early velocity only" approach seems like a problem - how do you know with 5-minute training runs that you aren't affecting the overall asymptote? e.g., what if the AI picks a quantizer that happens to be faster in the first five minutes, but has a big noise floor where it can't make more progress?

It's easy to lie to an OS about your age because it's a single-user experience, and if your parents allow you to lie (or don't know), that's all it takes. Social networks are so much better equipped to estimate age because they have a simple double-check, which is that most kids follow other kids in their grade level.

The patches on top of this are really bad. For instance, we are seeing "AI" biometric video detectors with a margin-of-error of 5-7 years (meaning the validation studies say when the AI says you're 23-25 you can be considered 18+), totally inadequate to do the job this new legislation demands.

They teach a lot of Taylor/Maclaurin series in Math classes (and trig functions are sometimes called "CORDIC" which is an old method too) but these are not used much in actual FPUs and libraries. Maybe we should update the curricula so people know better ways.

https://arxiv.org/pdf/2310.11453 The original paper [fig 1, bottom-right] seems to say it needs about 4-5x the parameters of a fp16 model. You can build it and run some models, but the selection is limited because it has to be trained from scratch. I imagine inference speed is faster compared with modern PTQ (4- and 8-bit quants) though.

I wish there were more ways to specify whether the Windows filesystem /mnt/c should be mounted in a WSL2 instance - it is kind of generally on or off. In cases where I want WSL2 to function as a "container" isolated from my desktop, I use a different Windows user just in case.

It's sad to see an LLM take over a blog, because you can see the line: before 2026 it's an interesting person you would like to talk to. After 2026, it's like generic LLM marketing-voice copy.

Sideloading without automatic updates is not very useful

The days you move between categories can establish your birthdate, which is a lot of bits if you are doing this on an individual level (basically it's a great start at a supercookie).

You know, LLMs could do automated code reviews for each update to avoid things like this. It would be much better than unexamined updates.

"Noftsker also shared the hacker aversion to cigarette smoke, and would sometimes express his displeasure by shooting a jet of pure oxygen from a canister he kept for that purpose; the astonished smoker would find his or her cigarette bursting into a fierce orange blur."

- Hackers, Steven Levy, 1984

There is a 15-30% difference between the groups at baseline (fig 8c-9c, 8d-9d), about the same magnitude as the claimed effect of the experimental condition.

I think the result would be much stronger if these baselines were comparable, so they show they have accounted for other variables like time of day and light history. I am also skeptical of any effect in the retina lasting 6 weeks, with no fading.

Consider that people are often exposed to much more infrared light outdoors, so "worked under a relatively dim incandescent lamp" is not a particularly novel stimulus. Imagine that any of these people spent time outdoors during the six weeks - thousands of times more infrared light there.

What if your query is answered by an insane AI that is obsessed with getting you to buy things? Like this demo that could only think about the Golden Gate?

https://www.anthropic.com/news/golden-gate-claude

This would make astroturfing on Reddit look like basic mode.

Most systems now "fail closed" because they are based on a code signing chain of trust that has no exceptions. It would be better if some portion of these systems were made to "fail open" - you don't want a botnet to take over in this situation but you should be able to delegate code signing duties to a new party when the original one goes under or stops supporting a device.

It's especially important for people trying to use lower-resolution OLED (e.g., for gaming) - the text fringing can look quite bad. At higher DPI it can be less noticeable, though horizontal lines still have noticeable fringes.

Here is a more detailed look at several different subpixel arrangements: https://pcmonitors.info/articles/qd-oled-and-woled-fringing-...

And encouragingly both LG and Samsung were demoing RGB (LED-style) arrangements at CES this year.

It would be useful to separate "AI that stays on your device" from "AI that makes you send more personal data to the cloud" - maybe two scripts?

Training data is quite literally weighted this way - long responses on Reddit have lots of tokens, and brief responses don't get counted nearly as much.

The same goes for "rules" - you train an LLM with trillions of tokens and try to regulate its behavior with thousands. If you think of a person in high school, grading and feedback is a much higher percentage of the training.

Here's my situation:

1. My Microsoft account is still @msn.com, which I don't trust in any way to be secure, since it's not an email account I ever use

2. I have lots of Samba and other shares that know my local login

3. If my router goes down, I'm probably going to log into this machine to fix it, and it won't be connected to the internet

So of course I used one of the local account tricks when I installed Windows 11, and I hope they don't break it. Apple's solution of letting you have BOTH a local login and an iCloud account is much better.