BearOso

Yeah, I was wondering why the two results were so similar, despite lack of detail in the prompt. Both have an array of "miniature" cities on a game-like map, all of them drawn surrounded by a circular border. Something sketchy is happening here. Either more prompt or intervention is going into it than the author claims, or it's copying some preexisting portrayal.

I dunno, it could be about money, too. $100,000 budget per employee, per month. Why were they willing to spend that much on AI usage? I hope said employees also make that much in salary.

It's not faithful. The AI took the original image and assumed the ! was supposed to be elevated like that. This project is low-effort AI generated.

Edit to add: This GitHub user has a over a hundred repositories full of random slop. It's clearly being used to generate activity for a resume or something.

The S is misproportioned. M should have the middle go all the way to the baseline. The middle stems on 3 and 4 are too long and too short. The punctuation is all different sizes. This is just me glancing at the normalized comparison image. It's really obvious.

Purely rhetorical--with sugar cane being so dominant, why are we turning corn into corn syrup again?

Inverse square law is proportional, so the distance isn't fixed. It depends on the base sphere of the original measurement, which can be quite large. If the size of the data center property is 100m and the flux at the property line is 70dB, it's only 1/4'd at a doubling of the radius, 50m away. So that only attenuates to ~65dB.

It's also not a uniform medium. Sound gets reflected by houses, the ground, resonance, so it can constructively interfere and have a extra few times the expected amplitude. This can make the sound quaver and become really annoying, too.

Apparently the noise emanates from massive fans for evaporative cooling on the roof. It's not coming from inside, which is why it carries so easily.

Extra zero. Noise level ~10m outside a facility is 90dB. Even traditional server facilities can hit 80-90dB inside. I've read reports from people living near data centers having almost constant 60-70dB noise levels at their houses, which, yeah, it still shouldn't fly.

The disparity is that older data centers were for storing and serving data. They have a lot of bandwidth and storage, but aren't consuming much energy. Plus, they are there to provide quicker access to data, to provide a service to actual people.

New AI data centers are obviously compute-oriented instead. That means they consume a lot more power, like hundreds of megawatts. For perspective, on average a U.S. household uses only 1.2KW. Because of this, cooling needs are significantly increased, and while water usage is often overblown, noise volume is a major problem. At 100m, it often exceeds the 90dB threshold for hearing loss.

Then there's the fact that millions of dollars of GPUs are inside these things, so security means lots of bright lights at all times of day.

These are just problems with the physical presence of a data center, not to mention the electric bill hikes, land value loss, environmental effects, etc. And there's no obvious positives to one going up. No new jobs. No tax money because it's often elided with incentives from bribing local officials.

With all the negatives and seemingly no benefits, ask yourself who wants these data centers, and why? If the only way to improve AI is using more and more compute like this, what is the end-game?

In reality, it is a big scam to trick the public into thinking AI is dangerous and needs to be regulated. Then the government can limit it so only the preferred companies can continue to develop it.

It's because both frontier companies are worried about plateauing and there's no barrier to competition. They both have upcoming IPOs, too.

4-5 orders of magnitude is huge. Assuming an order of base 10, it's 10000x-100000x. So a call to grep may return in 1s on a typical PC. That means a GPT call takes equivalent energy of 10000-100000 PCs to do the same in 1s. That's a difference that can't be equalized with scaling. It would require a revolutionary breakthrough.

I also don't understand where the idea that frontier models are getting better efficiency comes from. The results are certainly improving, but that comes from feedback and multiplexing requests, which cost more.

As the number of samples approaches infinity, the percentage share of all possible results approaches equality. A uniform random distribution is like the mathematical identity of statistics.

The two articles by the author are focused on the sharpness aspect. What's missing are the dilation/contraction due to beam intensity and resultant gamma. You can take away the scanlines and phosphor filter and still get a more "classic" image with just a modified bilinear filter.

A particularly good example of this are the whites of Mario's and Yoshi's eyes in SMW. They look much better with dilation. (No pun intended)

I kind of like the idea that as the edge of the universe is observed, it expands to fit that observation. "Observed" being in the same sense as the uncertainty principle. But there's also the idea of the universe or space as a kind of mesh that's stretching, not really growing.

When I think about things at this scale, it really makes me feel minuscule.

The Navier-Stokes thing? They had hundreds of segmented groups of 10000 agents running trying to solve that. I can't imagine the expense. For what little publicity it got, it wasn't worth it. There are also reports they cheated by using data from a couple human researchers, but I don't think that one's true.

It is kind of funny, because they're partly up in arms because the Mastodon thread linked to your slightly ambiguous comment.

Summary:

- The Motorola flagship costs too much and he's worried it will no longer run on the Pixel 7a that he got for $200 or that there will be any cheap options.

But they said already there would be cheap options, and Motorola has a Moto G series that it would likely use.

- Worried about other stuff that existing Motorola Android phones do, like privacy violations.

...that definitely won't be in GrapheneOS, they wouldn't allow it.

- More things Motorola does as a company that are bad.

The current option being Google means, despite that, Motorola is probably better.

[delayed]

In Jobs's time, 2007-2011 there was still a lot of overall pushback against PRISM, and also from Apple. I remember reading that they reviewed and only ended up complying with 3 requests. It's not nothing, but Google and Microsoft were far more eager to offer up their data.

He had a kind of decent plan to improve things, but when AI popped up he panicked and went crazy. Same thing happened to Zuckerberg, who abandoned his VR crap and dumped into AI at a massive cost.

The AI companies' marketing plan, FOMO, worked--just on themselves and competitors.

Following through on the top comment, surely you realized when you decided to forgo the back half of the test that, had the program compensated for it, you were consciously leaning into the ADHD diagnosis, right?

That's the thing about modern chatbots--they are nothing new technologically. They're just taking the same old neural networks and brute-forcing everything with ridiculous computing power and energy consumption. Autonomous agents just feed everything back into a loop that's even more wasteful.

If burning dead organic matter is modern technology, then I guess the pinnacle of science was in the stone age when we learned to make fire.

We all reduced CFC use and stopped ozone depletion.

The problem with climate change is that there is a huge incentive to not fix it, and the culprit is not just rich oil corporations. Many countries subsist solely on oil profits and have no other natural resources to sustain themselves, i.e. Saudi Arabia. Those countries have influence over others like the US.

There are alternative options already available to replace fossil fuels for most things. In first-world countries, the only thing stopping a transition is the loss of corporate profits.

Furthermore, legacy x86 BIOS will always boot in 16-bit real mode and require transition to 32-bit protected mode. Even NT is briefly 16-bit on those platforms, and Windows 95 only remains so just a little bit longer for backwards compatibility sake. It was less of a chimera than most of us thought.

It's starting to look like win32 was the pinnacle of GUI API design.

[delayed]

People don't want to pay for it. It's fine when you can buy a monthly plan and get lots of usage at the company's loss, but that's not sustainable.

Once everyone has to pay enough for the usage to just break even, it's going to be a shitshow. And what's worse, they'll have to make up for all the debt they're accumulating now, so even pricier.

Basically, individuals will only stay at the current massive discount. Corporations that are paying token costs right now are dialing back as they realize it's not worth the expense.

We are going to have an AI crash, and the only thing left are going to be the efficient on-device or local models, which are already starting to bridge the gap.

Not just significantly weaker. The Librem 5 uses the Cortex-A53 low-end ARM designs from 2012, and that particular NXP chip is not among the better implementations. The 3GB of RAM is anemic and becomes a greater restriction as time goes on. Many features simply haven't been implemented (most of the sensors) and those that have been that are technically difficult (i.e. camera) provide only a bare experience.

If we're honest about it, the Librem 5 is not something to use as your main phone, it's a toy to experiment with.

I like how this article pokes fun by not only using the tropes it describes sarcastically, but then adding subtext that AI never would. It also uses alliteration, grouping synonyms that use similar sounds, another playful literary device that LLMs don't get.