That was one thing that I wondered about for a moment, like, maybe you have people working very specific types of fields—I’d seen some papers in, like, psychology and stuff talking about human-AI interaction, and maybe this is a really hot thing to be doing to be doing in that field—but this survey was specific to “companion” uses, which I don’t think that would likely include professional use the way it was defined here:

The people who have used LLMs for social, emotional and personal purposes are known as “AI companion users” and a representative sample of 1,000 of them were surveyed to discover how they use their social bots, the nature of their relationships and the impact of their encounters with these systems.

These are some wild numbers, looking at them together.

So, basically, it’s heavily disproportionately loaded towards high-education users. 25% users have post-graduate degrees.

https://www.ccdaily.com/2025/09/datapoints-educational-attainment-in-the-u-s/

This has 14.9% of the US age 25 and older falling into that group.

But it’s also disproportionately loaded towards black and Latino users. This has 30% of the userbase being non-Latino black.

https://blackdemographics.com/black-population-in-the-u-s-reaches-51-6-million/

That’s 12.7% of the US.

I wouldn’t blink at either one of those statistics in isolation.

But…both black and Latino groups, particularly black, typically have lower levels of educational attainment in the US.

I would expect that to mean that there is either some really high concentration in one area—like, a exceptionally high level of black users with graduate degrees are using AI, enough to throw the numbers that far from overall proportions in society, or you have different strong “clusters” of different groups using AI—like, there is a really high level of black AI users and a really strong cluster of high-education users.

considers

It does say that users are disproportionately below 50, and that would explain some of it. Younger generations are both more educated and less white than older ones.

searches

https://www.pewresearch.org/short-reads/2019/07/30/most-common-age-among-us-racial-ethnic-groups/

Would need to go dig up demographics for each cohort to figure out what proportional numbers would be, but just glancing at that, that’s definitely not going to explain nearly all of that. The groups are roughly equally-sized for younger age cohorts, so that’s be a fair bit of that on age, but there’s still gonna be a substantial education split.

searches

https://nces.ed.gov/FastFacts/display.asp?id=72

For the 2021-2022 school year, this has whites being granted about six times as many masters or doctorates as either black or Latino students, so that’ll be a pretty big educational gap even for younger age cohorts. Yeah. Huh.

drums fingers

Well, I dunno. It’s an interesting pair of figures, though.

39% said they occasionally tell AI things that they would not tell other people.

I would be a lot less concerned about this were it not for the sneaking suspicion that most users (a) probably are using cloud-based LLMs rather than local and (b) have no idea what the rules on data privacy regarding their conversations on said service are.

Yudkowsky tends to like engineering bioweapons in If Anyone Builds It, Eveeyone Dies.

I’m not sure that that’d be my first concern, but I also don’t really think that ability to affect the non-electronic world is really the big hurdle. I think that the bug hurdle is getting some kind of AGI running somewhere, and once that happens, it’s going to be really difficult to keep it from acting in the real world in a way that gives itself more capability to act.

Like, I don’t think that the idea that computers are a sort of effective containment system and that the AI is inside a computer, and humans will act as a sensible check on the computer influencing the real world, so it can’t do much, holds up very well. If you do wind up with something more-capable than humans running around, I think that it’s going to be very hard to anticipate all the possible risks.

Say that someone manages to get an AI producing something particularly valuable and useful, doing something profitable for them. My guess is that they are probably going to link it up to whatever they think will be useful.

My understanding from past reading is that one of China’s principal near-term interests is in trying to leverage AI to do industrial automation. Even if it were not, you’re going to have more-and-more heavily-automated production facilities, and probably more general-purpose systems.

https://en.wikipedia.org/wiki/Lights_out_(manufacturing)

Lights-out manufacturing or dark factory is the manufacturing methodology of fully automating the production of goods at factories and other industrial facilities, without requiring any human labour presence on-site.[1] Many of these factories are considered to be able to run “with the lights off,” but few run exclusively lights-out production. For example, in computer numerical control machining, the presence of human workers is typically required for removing completed parts and setting up tombstones that hold unfinished parts. As the technology necessary for total automation becomes increasingly available, many factories are beginning to use lights-out production between shifts (or as a separate shift) to meet increasing production demand or to save money on labor.

An automatic factory is a place where raw materials enter, and finished products leave with little or no human intervention.[2

You have the ability to influence humans, which can also affect the real world. One of the most-seriously disrupted industries by LLMs right now is in copywriting, targeted advertising. Why? Because they’re pretty good at influencing humans, even in very primitive form. And we have an interest in building models that can influence people even more effectively.

You don’t have to have a human be off trying to intentionally trying to destroy humanity to be doing sonething that gives an AI increasing ability to act in the real world.

I remember reading an article a while back about some guy who had spent a while talking to ChatGPT or some similar chatbot and who became convinced that he needed to violently liberate the god being held captive in the computer. Like, if, as some have put it, “glorified auto-complete” can get a human to do that, what could a much-more-capable system with goals and intentions do? I think that just the ability to communicate with humans, which we certainly already permit (and is one of our main uses), would make containment pretty hard.

People already run all sorts of LLMs connected to uncontained, agentic things. Like, sure, if you have some kind of isolation system adequate to contain it, then there’s no risk. But…people are running stuff without such an isolation system, all over, now.

Hell, South Africa infamously used an LLM to write the legal rules that they were proposing to control AI, then didn’t validate the text by human. If there’s one place you would not want to use unvalidated AI output, that’s it.

I tend to be somewhat skeptical of the practical risk, but in If Anyone Builds It, Everyone Dies, Yudkowsky makes the case that existing systems are already too dangerous, not because of what they thenselves can directly do, but because they are being applied in ways that could lead more-sophisticated systems to emerge.

If you’re an AI researcher, one of the most obvious and imnediately-relevant applications of AI is using it to improve existing AI.

Think of our current CPUs. No human could ever have designed them. They are far too complex, contain billions of transistors. But…we built computers and then promptly used them as tools to build more capable computers.

Yudkowsky’s point is that you have people doing that sort of thing with AI now, and doing so in a way that produces very unpredictable results.

Now, my guess is that that doesn’t constitute much of a probability of going very badly by complete surprise. I think that it’s very likely that instead of a paperclip maximizer that obliterates humanity, you get something that fails spectacularly in a way that causes itself to stop functioning. Like, I’d bet that we’d see something more like a series of similar, spectacular failures before we get something that can function long enough and well enough to represent an existential threat.

But…I will concede that that’s more of a gut feeling. I don’t have real numbers. I can’t make a really convincing argument to myself that I am definitely right. And if I’m wrong, the cost is very high.

I have a bench variable-voltage power supply. Mine is a very low-end one, doesn’t even have an integrated voltmeter, so you need a multimeter or oscilloscope or something to set the voltage.

But…if you need some kind of exotic power at a given voltage and don’t have time to wait for an order to come in, one of those will fill the hole for all kinds of fixed-voltage power supplies.

I have only very rarely needed it. I’ve used it to provide 5V power to a PIC project that I was breadboarding, and another time tried checking whether it was low voltage causing a problem with a given battery-powered device. Maybe one or two other occasions.

If you specifically want barrel plugs, you can get “universal” adapters that will adapt one cable to various diameters and polarities.

I think that USB is also kind of increasingly filling the hole here as a universal standard for DC power on newer products.

Yeah, I had that problem.

I then had a coworker who was a sysadmin, and randomly dropped into his cable box at work and realized what I was doing wrong. Get a bunch of cable ties. Coil up each cable and cable-tie it.

And honestly, one has to keep perspective. Yeah, I’ve occasionally used a really old cable, but let’s say that the worst happens and I need to occasionally get a new cable. It’s…probably not that much of a disaster. It can be hard to get some old computing hardware, but even very elderly cables are generally commercially available.

I tend to think that keeping around things like IECs and analog audio cables are more-worthwhile since they’re slow-moving standards. Most old PC cables just haven’t found much use, in my experience.

Cables used to be something that brick-and-mortar stores had a large markup on. I remember a guy I once knew who had worked at a PC store telling me that the store made more (like, in absolute terms) selling the parallel cable that came with the printer than with the printer. Back then, having some cables around could avoid that, and having a few on-hand was a good idea. But online retailers have made the competition a lot more intense.

I typically have gotten cables at Monoprice if I’m getting a bunch at once, which in the US has historically has had pretty good cable prices, though I haven’t tried comparing online retailer prices recently.

Yeah, honestly, you could go way further back.

https://en.wikipedia.org/wiki/Venus_figurine

A Venus figurine is any Upper Palaeolithic statue portraying a woman, usually carved in the round.[1] Most have been unearthed in Europe, but others have been found as far away as Siberia and distributed across much of Eurasia.

Most date from the Gravettian period (26,000–21,000 years ago).[1] However, findings are not limited to this period; for example, the Venus of Hohle Fels dates back at least 35,000 years…

For context, we figured out agriculture and started to move away from hunter-gatherer life about 12,000 years ago.

These figurines are recognised as some of the earliest works of prehistoric art.

Most have wide hips and legs that taper to a point. Arms and feet are often absent, and the head is usually small and faceless. Various figurines exaggerate the abdomen, hips, breasts, thighs, or vulva, although many found examples do not reflect these typical characteristics. Depictions of hairstyles can be detailed, and clothing or tattoos may be indicated.[5]

The original cultural meaning and purpose of these artefacts is not known. It has frequently been suggested that they may have served a ritual or symbolic function.

I have a guess.

https://en.wikipedia.org/wiki/Demon_core

The demon core was a sphere of plutonium–gallium alloy that was involved in two fatal radiation accidents when scientists tested it as a fissile core of an early atomic bomb.

Scientists referred to this flirtation with a nuclear chain reaction as “tickling the dragon’s tail”, based on a remark by physicist Richard Feynman.[12][13]

It was a case of an unfortunate mistranslation due to the lack of awareness among historians about the thriving scalie scene at the Manhattan Project.

Legend has it that every new technology is first used for something related to sex or pornography. That seems to be the way of humankind.

—Tim Berners-Lee, inventor of HTTP and HTML

https://www.theguardian.com/technology/2002/mar/03/internetnews.observerfocus

The dirty secret that drives new technology: it’s porn

The Guardian, 24 years ago

If there’s the demand, and I bet that there will be, someone is going to be making plug-and-play local AI compute boxes once prices come down that eliminate a lot of the learning curve. Like, preconfigured software and all that.

Joe User is not going to be talking about memory bandwidth or compiling software or setting sampler parameters. But—as the cloud guys demonstrate—he can use it if it’s set up. And there will be more-approachable front ends and such.

Use a local model

I also think that local models are generally preferable, but everyone having a local model requires more hardware, since much of the time the local hardware to run the local model is probably going to be idle, whereas hardware in the cloud can be servicing User B when it’s not servicing User A.

In 2026 and 2027, absent some very unexpected development, memory prices are going to be elevated.

And while some people could get local hardware, we couldn’t build a comparable-to-what-someone-could-get-in-the-cloud AI compute box for everyone if everyone wanted to go local, not for years, as we don’t have the memory production capacity today. If everyone tried to get a local box for AI compute at the same time, it’d just do what happened when the cloud companies did it, but at greater scale, because it’d require even more memory. There’d be a new memory shortage. Prices on memory would keep rising, pricing would-be buyers out, until the number of buyers still willing to buy was equal to what supply was present.

Despite the fact that it’s registered and paid for until 2040.

I can believe that Verisign might have some route to terminate service (even in cases where there is some contractual obligation to provide a service for a given period of time, someone can typically break the contract, just having to pay some sort of penalty for doing so), but I also suspect that they are legally required to issue you some sort of refund.

You generally cannot just sell someone a service and then refuse to provide it; that’ll violate consumer protection laws.

That probably won’t make the author happy—he probably wants the domain, not the money. But at the minimum, I will say that he probably doesn’t have to pay for service that he doesn’t get.

It might be that they are doing that; he doesn’t specify.

But it gets much worse. Once the 3rd-level domains are terminated, it is assumed that the now vacant 2nd-level domains will become available for registration. Should someone (other than me) scoop up fraser.name they would be able to recreate and control neil.fraser.name. They’d be able to hijack hundreds of accounts that are linked to that address. They could commit code with my authentication. They could seize control of IoT devices. There is no way to enumerate all accounts (online and offline) which have been opened using this email address over the past quarter century.

Yeah, personally I’d rather have some kind of key-based mechanism, at least as an alternative to the X.509-based system. Think of what PGP or SSH do. Like, the key is your identifier, not the endorsement of some centralized registrar.

Hyphanet does something like this, can do key-based addresses.

That has some drawbacks too—I mean, you are the ultimate authority here, and thus responsible for keeping that key safe, and that’s a big responsibility. If you think of all the attacks on cryptocoin wallets and stuff, I think that personal computers are not incredibly-well-suited for that—if we’re going to do that, we probably want something more like hardware-based keystores, like YubiKeys or similar, where a key cannot be exfiltrated if a PC is compromised, as PCs have a big attack surface. And it’s harder for humans to remember a key as an identifier, which means that they probably have to have a computer somewhere store a list of known and trusted keys (though doing so solves a lot of existing attacks, like phishing emails with look-alike URLs that attack the limited human ability to remember names, so there are pros and cons). And it avoids the risk that someone with access to a certificate authority’s key can compromise your security.

Whether or not it shows it requires a login for old.reddit.com depends on IP. I see this on a few networks, not on others.

The big concern with filesystems becoming corrupt is that all of the data can be lost on a filesystem. Most of the time, this doesn’t show up, but if you lose power at just the wrong time, it can happen. With FAT, HFS, ext2, and similar filesystems that were vulnerable to corruption on power loss, that was a real risk.

So having some kind of power source to make an unclean shutdown unlikely avoids the risk of a really, really bad scenario.

The general shift away from reliance on those filesystems has mitigated the really-bad-case that drove a lot of UPS use.

Databases are probably the most-critical case where applications update a file on disk constantly, and where corruption would be bad. As long as they live on a filesystem that itself can’t become corrupt on power loss, they’re normally structured with a carefully-ordered series of writes and use of write barriers so that they cannot become corrupt.

An individual file can be partially written out. Like, say I’m saving a file in Photoshop, and I have only written half of that file. What winds up on the disk is the first half of of a Photoshop file. That’s an invalid Photoshop file. For some applications, that can matter. Like, say (a) I chose to overwrite an old file and (b) the application doesn’t atomically store the new file (like, write it out, fsync(), and then rename the new file over the old) and © there’s no backup or other system in place.

Or, even aside from partially writing out a file, I can entirely lose work that isn’t saved at all, if I’m not using a software package that makes use of some sort of autosave system. If I haven’t saved all day, then, well, all of that unsaved work is going to go away if there’s a power loss.

That’s a much smaller potential data loss, but, sure…power loss can cause data loss, even if it doesn’t produce that “everything is lost” filesystem corruption.

I’m not asserting that there’s no utility to a UPS; quite the contrary, as my comment was explaining to OP why there is use some people have for having NUT even if he personally doesn’t care much about it for his use case. I’m just pointing out that some of the major reasons for having a UPS+host-side-software combination that existed in, say, the 1990s aren’t really much of a thing in 2026. That was that you really did not want machines using filesystems normally used then to perform an unclean shutdown. Even with a UPS and host-side software like NUT, you can have data loss, as NUT cannot force all software to save unsaved data; there’s no standardized mechanism for forcing applications to save. If the PC is attended, you might get a warning and have a chance to manually save, but that’ll be the best that’ll happen, and if the PC is unattended, well, that save isn’t going to happen. But what NUT and similar software can do is to, when the UPS battery level gets low, trigger a clean shutdown. That ensures that the OS has time to perform a clean shutdown, and on filesystems that have a risk of filesystem corruption if the power is cut—FAT, HFS, ext2, and similar—a UPS with host-side software would avoid that case from arising.

If someone wants to have a UPS in 2026 on filesystems that don’t have those limitations, that’s fine. They may want to have, say, a five minute window to save their work if they are at the PC when power loss occurs. It may help avoid problems that arise from software that cannot cleanly handle power loss—I that Steam can, if power loss occurs while it is updating itself, wind up in a broken state from which it cannot automatically recover when restarted, as I’ve done that myself.

But generally, if you lose power on a system in 2026, you aren’t facing a “I might lose the contents of my hard drive” scenario.

ext4 is merely journaled, which means that there are rare instances where it can be corrupted, but recovery is fast.

None of these things mean that the filesystem becomes corrupted; the filesystem will be in a filesystem consistent state when remounted. You can have a write that is partially performed on a file, but that does not cause the filesystem to become corrupt when it does not complete being written. The journal can contain that write, but it can not yet be committed, and won’t yet be visible to a user. But that doesn’t make the filesystem corrupt; the commit itself is atomic.

And even if the filesystem doesn’t become corrupted, individual files can become inconsistent or incomplete.

Sure, but that is not filesystem corruption, and I addressed it in my comment.

Modern filesystems can corrupt as well with unclean shutdown. They are just less likely to do so.

It shouldn’t be possible for them to become corrupt on an unclean shutdown, at all. Not at the filesystem level. It isn’t just a question of probability.

It’s possible that there is some bug in the implementation of a filesystem that causes it to violate its guarantees.

There could be, oh…hardware bugs, say, where hardware violates guarantees.

Files can be partially-written. Software might not handle that, though properly-written software designed like a DBMS will be set up so that a partially-written file will not become corrupt at the file level.

Some filesystems permit disabling features that guard against filesystem corruption.

But absent something like that, it should not be possible for one of the filesystems to become corrupt on an unclean shutdown, at all.