This is the root of the problem. Reverence and social deference for "faith". On other words: "Belief without evidence"

Everyone alive since the Pandemic and subsequent Disinformation Wars knows that "faith" is a toxic and stupid idea which is incompatible with modern global civilization.

Religion needs to be viewed with great suspicion. Science and the people who discovered real things about the world need to occupy that place of reverence.

(LLMs) a technology LeCun has repeatedly called "a dead end when it comes to superintelligence."

He keeps saying that but AFAICT the block diagram for his JEPA solution is the same thing just predicting next floating latent space token instead of discrete word-token. Which is very powerful and cool but I mean its not like he is getting rid of backprop or convolution or even attention really. He should stop attacking his older work and just be a rock star.

You could say by llama4 the genAI division had lost its legs. But the Metaverse division was way ahead of them...

I don't want people all over the world connecting to my bedroom. If I wanna host a website I pay an extra $8/mo for VPS

One word drives darknet markets: Drugs That is enough of a usecase

Somehow the lawyers will get the majority of the settlements.

This is bad news if the Pixels are now eSIM only.

GrapheneOS only supports pixels and many of its users prefer a permissionless SIM.

Maybe they'll add support for some other device, perhaps one of the privacy phones.

I started PC programming in 1982, had to teach myself asm almost immediately in order to write a hardware interrupt handler for the serial port.
Back in those days code size was important, and for any driver/TSR type program it was simply crucial. Here in Norway we needed to load the KEYBNO.COM program which took over the keyboard interrupt and provided the official Norwegian layout, including the 'æÃÃ¥ÃÃÃ...' keys for our 3 extra letters in the alphabet.
Over the years, Microsoft/IBM had many version of this driver, the final one which also did text more font layout changes (at least for the default 25x80 mode) was up to 60 kB. This was large enough that lots of US-developed engineering/DB applications simply didn't fit, so I wrote a replacement:

All keys (including Ctrl- and Alt- modified combos) were handled, along with font remapping for 25x80, 43x80 (EGA only) and 50x80 (VGA only) screen layouts, using a total of 704 bytes. It became so popular that Compaq/HP stole it to give to their customers, then when we caught them red-handed (they had not figured out how to get rid of my startup Copyright message) they refused to pay but promised to not do it again.

Half a year later we caught them again, they had started telling customers that they had to keep it a secret.

Terje Mathisen

ah, I will check out memory hole on some of my systems.

At least on 6.1 you have to be below 50% RAM usage too.

I found this in a RHEL doc that pointed to a kernel README that looked old af but said the same thing.

I have a few systems that run an app on solar during the day at 80% RAM and I had to stop the service before suspend to get it to work.

Yet it worked for a couple months in disk hibernate but then stopped and only memory sleep would work. On a Debian Bookworm stable kernel, so who the heck knows what broke (wasn't me!).

Battery usage overnight is different enough with many machines that I wish hibernate to disk worked reliably.

Anyway if I have 16GB RAM and a 36GB swap it seems bonkers to me that it is by design only working if less than 8GB of RAM is committed.

The subsystem is quite brittle and everybody seems to know.

Most package management systems require us to figure out which card we have, figure out which package supports it, and install that.

Really we wanted "install the package that supports my card".

Apparently this current problem highlights this disconnect when a package no longer does what it used to but the package system blindly updates it anyway.

Being 2025, surely somebody in the past 30 years has floated a meta package management system to handle this mapping? Or an apt plugin? Anybody here know that history?

I mean, we even have nvidia-detect for their cards to do the actual probing work.

Granted arch is rolling and rolling gonna roll, but we can have software that makes this work correctly.

NIST have always been the world leader in creating ever more accurate clocks, the current masers work in picoseconds and below, so allowing the reference to drift by 4.8 microseconds means that precision dropped by at least 6 orders of magnitude.

If allowed to propagate to the GPS control clocks, this would have been enough to totally destroy the navigation system since a clock that is off by 4.8 us corresponds to a position error of 1500 kilometers. (OTOH, USNO has its own large ensemble of atomic clocks, so they don't depend short term on NIST updates.)

Full disclosure: I worked with the NTP Hackers (network time protocol) team for 20+ years, so I'm probably a bit more interested in precise timekeeping than most. I have personally soldered together 4 or 5 GPS-based reference clocks that would deliver 25-35 ns RMS precision.

I voted against Biden in the Primary. But as POTUS he surrounded himself with competent people AND LISTENED TO THEM.

Yes he and they made sime bad choices (Afghanistan withdrawal, tepid Ukraine support) but not strategic damage and coupled with some very good decisions (block China access to GPUs, energy policy)

If Dems sweep the Midterms this has a high chance of happening.

REMEMBER: ELECTIONS HAVE CONSEQUENCES

This guy either socially engineered his way through a line, analyzed a weakness in the line, or time-traveled from the '90's not realizing we've set up an incompetent but totalizing police-state control grid to interpose every tiny aspect of our lives.

To be fair, "pay on board" is less applicable to airplanes than trains because seatbelts are important in turbulence.

That said, the lack of capacity is widely acknowledged to be a feature of wildly incompetent management.

We just heard they've started a new project to rewrite the air traffic control system for the umpteenth time (and billions and billions later) to hopefully allow for more frequent landings and departures. I fear it won't be specified for AI-assist takeoffs and landings and will be obsolete before it's done.

Better make some more 8" floppies.

There's also a local LLM for link previews of dubious accuracy. And another for suggesting and naming tab groups.