The company describes this generator as a solid state device, but the diagrams show the reliance on fluid/flow of hydrogen between the hot side and the cold side for moving some protons around. That seems to be something in between the semiconductor-based solid state thermoelectric generators that are already commonly understood and some kind of generator with moving solid parts.

It still seems like a low maintenance solution to have a closed loop of hydrogen, but that seems like a potential maintenance/failure point, as well, to rely on the chamber to remain filled with hydrogen gas.

The inventor/founder at the center of the article, Lonnie Johnson, was on the team at JPL that designed and implemented the thermoelectric generators (heated by radioactive decay from plutonium-238 pellets) on the Galileo spacecraft sent to Jupiter. So I would expect that he’s more familiar with the thermodynamic and engineering challenges than even a typical expert.

The PR fluff put out by the company mentions that the theoretical basis for this specific type of generator was worked out a while ago but needed materials science to advance to the point where this type of generator can be thermodynamically and commercially feasible.

Looking at how this generator is supposed to work, it’s interesting in that it does rely on the movement of fluid, but is supposed to be a totally closed loop, to be a bit different than the pure solid state, semiconductor-based Seebeck generators that are already well known.

The other area talked about in this article is that they believe that it can be effective with lower temperature differentials than any previous technology, which might make a huge difference in whether it can be deployed to more useful places and thereby make it economically feasible more easily than prior concepts.

In the end, if these generators can output some electric voltage/current, it might just take on similar generation characteristics as photovoltaics, which could mean that hooking these up to the grid could draw on some of the lessons learned from the rise of grid scale solar.

Specifically, desktop RAM is slabs of silicon, placed into little packages, soldered onto circuit boards in DIMM form or similar, to be plugged into a motherboard slot for RAM.

The AI demand is for the silicon itself, using advanced packaging techniques to put them on the same package as the complex GPUs with very high bandwidth. So these same pieces of silicon are not even being put into DIMMs, so that if they fall out of use they’ll be pretty much intertwined with chips in form factors that a consumer can’t easily make use of.

There’s not really an easy way to bring that memory back into the consumer market, even after the AI bubble bursts.

More along the lines of a “pizza finder” service that scours different menus and shows the pizza options at a bunch of places, whether those places exclusively offer pizza, specialize in pizza with some other options, or just offer pizza as one of several options. It would be perfectly reasonable for such a service to only return results related to pizza, without any implicit suggestion that each place it returns only has pizza available.

Financial analysts were sounding the alarm in October. On October 7, Bloomberg ran an influential article about the circular deals:

https://www.bloomberg.com/news/features/2025-10-07/openai-s-nvidia-amd-deals-boost-1-trillion-ai-boom-with-circular-deals

That built on earlier reporting where they described the deals as circular, as the deals were being announced. Each of these reports notes the financial analysts at different investment firms sounding the alarm.

From there, a robust discussion happened all over the financial press about whether these circular deals were truly unstable. By the time Gamers Nexus ran that video the financial press was already kinda getting sick of the story.

Whatever the hell these trading algorithms were doing on November 20, they definitely weren’t ahead of the curve on investor knowledge and belief.

it’s like AI companies went from buying 10 memories as usual to 1000000,

I mean, they basically did. OpenAI announced a few months ago that they reached deals to buy 900,000 wafers of DRAM per month, representing 40% of global production capacity. For a single company. There are several other companies competing at those scales, too.

Plenty of the AI functions on phones are on-device. I know the iPhone is capable of several text-based processing (summarizing, translating) offline, and they have an API for third party developers to use on-device models. And the Pixels have Gemini Nano on-device for certain offline functions.

but I don’t think most are designed to control people’s opinions

Yeah I’m on team chaos theory. People can plan and design shit all they want, but the complexity will lead to unexpected behavior, always. How harmful that unwanted behavior is, or how easy it is to control or contain, is often unknown in advance, but invented things tend to develop far, far outside the initial vision of the creators.

And LUKS was already widely available on Linux as alternative.

Yeah, I found LUKS and LVM to be more intuitive for creating encrypted partitions, and had that on my daily driver by around 2009 or so, so I never really felt the need to try Truecrypt.

Being able to point a camera at something and have AI tell me “that’s a red bicycle” is a cool novelty the first few times, but I already knew that information just by looking at it.

Visual search is already useful. People go through the effort of posting requests to social media or forums asking “what is this thing” or “help me ID these shoes and where I can buy them” or “what kind of spider is this” all the time. They’re not searching for red bicycles, they’re taking pictures of a specific Bianchi model and asking what year it was manufactured. Automating the process and improving the reliability/accuracy of that search will improve day to day life.

And I have strong reservations about the fundamental issues of inference engines being used to generate things (LLMs and diffusers and things like that), but image recognition, speech to text, and translation are areas where these tools excel today.

AI drives 48% increase in Google emissions

That’s not even supported by the underlying study.

Google’s emissions went up 48% between 2019 and 2023, but a lot of things changed in 2020 generally, especially in video chat and cloud collaboration, dramatically expanding demand for data centers for storage and processing. Even without AI, we could have expected data center electricity use to go up dramatically between 2019 and 2023.

The analog dials were an illusion. That information has been processed digitally for at least the last 25 years.

What I’m saying is if YouTube is sharing $10 million of revenue with channel owners in a month that has 1,000,000,000 total views across YouTube, that’s a penny per view.

Then, if the next month the reconfigure the view counts to exclude certain bots or views under a particular number, you might see the overall view count drop from 1,000,000,000 to 500,000,000, while still hitting the same overall revenue. At that point, it’s $0.02 per view, so a channel that sees their view count drop in half may still see the same revenue despite the drop in view count.

If it’s a methodology change across all of YouTube, a channel that stays equally popular as a percentage of all views will see the revenue stay the same, even if the view counts drop (because every other channel is seeing their view counts drop, too).

Isn’t that the formula? They take all of the revenue, set aside the percentage they’ve set for revenue share, and then divide that among all channels based on viewer counts. Dropping viewership for all channels proportionally means that the same amount of revenue will still be distributed to the channels in the previous ratios.

Most 4k streams are 8-20 Mbps. A UHD runs at 128 Mbps.

Bitrate is only one variable in overall perceived quality. There are all sorts of tricks that can significantly reduce file size (and thus bitrate of a stream) without a perceptible loss of quality. And somewhat counterintuitively, the compression tricks work a lot better on higher resolution source video, which is why each quadrupling in pixels (doubling height and width) doesn’t quadruple file size.

The codec matters (h.264 vs h.265/HEVC vs VP9 vs AV1), and so do the settings actually used to encode. Netflix famously is willing to spend a lot more computational power on encoding, because they have a relatively small number of videos and many, many users watching the same videos. In contrast, YouTube and Facebook don’t even bother re-encoding into a more efficient codec like AV1 until a video gets enough views that they think they can make up the cost of additional processing with the savings of lower bandwidth.

Video encoding is a very complex topic, and simple bitrate comparisons only barely scratch the surface in perceived quality.

Article is paywalled for me.

Does it describe the methodology of how they use the transmitter and receiver?

What specifically are they transmitting? Is it actually wifi signals within the 802.11 protocols, or is “wifi” just shorthand for emitting radio waves in the same spectrum bands as wifi?

Yeah I’m with you.

“Using this technological advancement to improve health care is good”

“Not in countries where health care is publicly run”

“What” is the correct response here.

“The only difference between the two emails was the link,” the memo said. “ActBlue delivered. WinRed got flagged. That is not a coincidence.”

It could also be that winred is more often associated with spam because emails with winred links use a style more associated with other actual spam. Like if spammers use words like Trump a lot to try to scam victims, and a lot of those emails get flagged as spam, then the word Trump itself becomes more highly correlated with spam. And since the word Trump is highly associated with winred links, maybe winred gets caught up in the rule set/heuristics that associate Trump fundraisers with spam.

How does being a dick to users get back at site admins you don’t like?

That’s another round trip, and you still have to use JS to identify þe browser.

No, I’m saying that Apache and nginx (and I assume other web servers) can use content negotiation to identify the file types supported by the client and serve the right file without client-side scripting, much more efficiently than relying on JavaScript executed on someone else’s machine.

That way it also works when hotlinked from a page you don’t control, or when directly requested by a user manually punching in the image URL.