I'm sticking with stone knives and bear skins.

We can bust on Microsoft all day and all night, and they deserve it, but the fact that their ability to find and fix these problems has greatly increased is a good thing. Software is incredibly complex, and no software more complicated than "10 GOTO 10" is free from the potential of security problems. Microsoft's QA has gone downhill in recent years, but now it's getting better apparently (even if it's after the fact). They are not going away, so this makes all our lives better.

I wrote a file searching utility in C++ about 30 years ago. About 15 years ago I ported it to Python and found it to be faster than using the Windows directory enumeration APIs, and I've been using that ever since. It just searches filenames, but that's all I need.

"NOTE: Experiences vary by region." says to me that Microsoft doesn't even know what their changes will do.

Can they factor a number larger than 21 with QC yet?

Perhaps they could take a clue from software from 40+ years ago and make top-posting a configurable option.

But Microsoft has pretty much completely given up on making their software configurable for user preferences.

Hey, I resemble that remark.

"Dr. Sam Becket never returned home."

You forgot to misspell his name.

Not if they keep spending 10 figures a year trying to make VR happen.

How long can Meta survive without shipping a product people actually want?

Most likely the next step change in AI improvement will come from the next generation learning from today's LLMs and diffusion models. But not being trained on their output. More likely they will be evaluated by current AI.

AI advancements have a predictable pattern. First we mimic how humans do something, then we use those AIs to evaluate the next generation which learns how to do it without mimicking humans. LLMs have been trained on human output and work by trying to determine what a human would do in any situation. The next generation will most likely learn how to think and speak without ever seeing a word of human generated text. Now that LLMs exist to evaluate their output we have the necessary building blocks to design the generation of AI that can really produce content far better than any human.

We are living in the era of early-to-mid 90s chess engines, where AI learned from human moves and brute forced its way to barely beating the best in the world. It took 10 years from the point where chess engines could compete with top humans players until chess engines were effectively unbeatable. And those unbeatable engines were trained by other AI, not by looking at human games. It got to the point where looking at how humans played would have just made it worse.

That's why they compared it with YouTube shorts, which is basically YouTube's TikTok replica.

"Please protect us from child-harm lawsuits since child harm is pretty much our business model."

This is very important. There's a light-year of difference among a typical scholarly article, a physics paper, a math paper, or some kind of incomprehensible humanities bafflegab that no sane person could comprehend. The former, if it's not too technical, should be readable to the average undergrad. The second and third might not be because there are so many specialized concepts and so much specialized language. The latter (and I'm not indicting everything coming out of the humanities, but a lot of it) is incomprehensible because it literally doesn't make sense.

Did they take into account that most countries drive on the right? If you're walking the perimeter of a room counterclockwise, it might be because you are staying on the right.