Delegating multiple levels down is allowed by DNS spec but almost never done, and causes problems. For starters, some registrars just straight up did not handle the first.last.name case at all. Which is particularly bad for .name because originally, you could only register third-level domains. Only later did .name added second-level domain registration.

This creates an obvious collision problem: if John Doe registers john.doe.name first, and then Jane Doe wants to register doe.name, what happens?

You might think that, obviously, both registrations stick. But this means John Doe owns a subdomain of Jane Doe's namespace, that Jane Doe cannot get rid of or otherwise control. When someone looks up records on john.doe.name, doe.name is never queried; .name skips it. This causes all sorts of problems for cookies: John Doe can now set cookies on Jane Doe's website, and Jane Doe's website always sets cookies that John Doe's website can read[0].

The time-reversed scenario (John Doe registers after Jane Doe) is even worse. Obviously, top-level domain operators should not be allowed to project subdomains onto other people's registrations like this. Like, imagine if Verisign decided you could register subdomains on someone else's domain, that they did not control and could not get rid of, if you paid them a little extra? There would be all sorts of shenanigans.

The alternative resolution is confiscation: when Jane Doe registers doe.name, John Doe loses his website. This is more in line with how you thought it was working. And, indeed, what Verisign is planning to do is basically this en masse, dropping all the third-level delegations. Probably because the operational headaches of managing two different conflicting delegation schemes was too much.

[0] This is actually a problem for any DNS suffix that is shared across multiple websites; there's a Public Suffix List that exists specifically to keep you from, say, writing cookies to .com that hit every website. But doe.name isn't exactly a public suffix: it's Jane Doe's personal website, one that just happens to have a suffix collision with another website because of a stupid idea someone had.

[0] Wikipedia also mentions the existence of cloning tools that would let you copy any AMPS subscriber ID and program another phone with the same number, stealing the subscriber's account. I suspect AT&T was more angry about that.

There is some standardization in that, if you issue geometry and sample textures normally, you will be relying on hardware units for rasterization and texture sampling, and those do have a relatively limited set of options for how they work. However, those options encompass most if not all of what 3D engines could ever want or need, even the early 3D ones that had all sorts of hacks. The kinds of things GPUs can't actually rasterize in hardware are things like NURBS or metaballs, but this doesn't matter because the industry settled on triangles[0] as the medium of real-time 3D really early on.

A lot of the look of old 3D games comes down to them doing things the lazy way, and I'm proud to say the lazy way is still very much supported - so long as you turn off features that didn't exist later on and limit yourself. I'm not as familiar with old DOS games, so I'll use the PS1 as an example. The "PS1 look" is very much a result of its incredibly limited rasterizer, whose limitations can be trivially emulated in modern hardware:

1. Don't allocate or bind a depth buffer. You will need to Z-sort your geometry every frame. Part of the "PS1 look" is that some games did it wrong; if you do it correctly you'll be burning performance for no visual benefit. Pick a scheme that breaks in a visually interesting way. Likewise make sure to not allocate any multisampling since the PS1 didn't anti-alias either.

2. Restrict the precision of vertices in your vertex shader by rounding them. The PS1 did vertex math with incredibly low precision and it makes vertices wobble.

3. Restrict the precision of UVs in your fragment shader. PS1 UVs were 0-255, so you want 1/256 precision.

4. Aggressively discard geometry that clips the edge of the viewport. The PS1 did not have proper polygon clipping; you just couldn't render off the end of the viewport, so games had to discard geometry and subdivide polygons to avoid things disappearing. This is another "if you do it well, nobody will notice" thing.

5. Turn off perspective-correct attribute interpolation; this will make your UVs warp and distort at certain angles. You do this with the `noperspective` attribute in GLSL.

6. Disable bilinear filtering when sampling textures.

7. Allocate and render to a low-precision color buffer. Lots of games would change the color depth to match the available VRAM. This is probably the hardest aspect to emulate unless you are actively counting VRAM usage and sticking to realistic values.

8. Bitcrush your textures before uploading them. The same reasoning applies here. Also, don't upload any mipmaps.

9. Lighting was done through vertex colors, which were calculated on the CPU. You don't have to limit yourself to CPU lighting but you do have to make sure lighting is done in the vertex shader.

10. Optional: implement the early model PS1 texture banding bug in your fragment shader.

This will give you a rendering pipeline that is good for basically all purposes except bit-accurate PS1 emulation. Of course, if you were doing bit-accurate emulation, nothing stops you from issuing a screen-filling quad and writing a GPU shader that does software rasterization on the GPU. Or launching a compute shader that accesses the framebuffer arbitrarily - although you do have to worry about memory coherency across warps and workgroups. Texture sampling can also be done "in software" if you so wish - i.e. by doing unfiltered reads to the texture which are glorified array lookups, and then filtering those reads as you wish.

[0] There was a brief moment where quads might have been able to win - artists like to work with them, and Nvidia was shipping quad hardware. But the advantage died very quickly, and quads can be easily converted to triangles - not so much the other way around.

This is a Chinese character that can refer to goats, rams, and several other kinds of "medium-sized animal with thick wooly coats and horns". English has no direct equivalent; we have to be too specific when we talk about these animals.

A model with a small number of neurons has to compact its concepts in order to fit within the parameter count while reducing loss. OP's article shows that the compaction is usually generalization: "goat" and "ram" get compressed into "羊". The canonical example of superposition, an image model where "car" feeds into a bunch of "dog" features, was shown to be actually generalization. In a sense, the model was so small that its optimizer didn't have the parameter budget to distinguish between "car" and "dog"; it has a single set of "thing that moves fast" features.

So basically, really small models are merging features into semantic buckets that are so generalized they don't make sense to big human brains unless we really, really squint. The buckets chosen happen to be what has the lowest loss, of course.

Secure Exclaves are special isolated processes that run on the same application processor as the regular OS. They're a very new OS feature and AFAIK Apple has barely talked about them until now.

Exclaves are isolated because Apple made a proprietary CPU extension called Guarded Exception Levels. It lets you separate an ARM exception level out into two privilege levels: guarded and unguarded. The guarded level can have different page permissions from the unguarded level, and switching between the two modes is extremely fast. Apple uses this to lock down iOS: its kernel, XNU, has it's entire virtual memory system hollowed out and replaced with a thing called the Secure Page Table Monitor (SPTM). XNU lives in EL2 (ARM hypervisor) but SPTM lives in GL2 (guarded hypervisor). In this configuration, SPTM is the only thing that can allocate memory, and it controls all memory allocations in the system. This is specifically why jailbreaks have become far harder to make[0]: you basically need to hack SPTM to get anything.

XNU, strangely enough, lives in both EL2 and EL1 (ARM kernel); but Apple also uses GL1 (guarded kernel) for a second, funnier OS called the Secure Kernel (SK). SK has its own exclusive resources and userland. Processes that run under SK cannot be tampered with or monitored by XNU because SPTM isolates them and their kernel from XNU. This is used for isolating any particularly security sensitive thing.

For example, code signing enforcement happens in an Exclave[1]. XNU can't run unsigned processes because it needs SPTM to bless the pages. But SPTM is supposed to be a very small attack surface; there is a long history of using memory safety bugs in the verification code to gain higher levels of privilege, so putting that code in GL2 would be bad. Instead, it gets put in an Exclave, where:

- A compromised XNU cannot tamper with the verifier, because SPTM won't give it write access to the pages.

- A compromised verifier doesn't compromise the whole system, because it's not part of SPTM/GL2.

- A compromised verifier can't interfere with other exclaves, because they're isolated by SK.

The only thing a compromised verifier can do is fail to reject an unsigned program, which would let you run specific unsigned code in EL0. You would not suddenly get kernel mode access from a verifier failure.

Anyway, what makes this relevant for HomeView is that all the microphone data processing can be isolated in an Exclave from the rest of the OS. So if the CIA wants to force your HomeView to listen to you, it needs to either compromise SPTM (very hard), or find a (non-memory-related) logic error in the listener exclave.

[0] This is also morally the same scheme as the Xbox 360 hypervisor that went decades without a softmod

[1] Morally speaking, it's an Exclave; it lives in GL0. But technically the code signing enforcement is in a different SPTM domain from the Secure Kernel, called the "Trusted Execution Monitor" (TXM). I would not be surprised if Apple migrates it over to an Exclave in the future.

I'm sorry, what?! Communications is a valuable thing. How is all that AI slop going to hit the Internet if there's no routers?

There are some wrinkles to this. If CXMT's RAM was somehow inferior, then the market could sustain two different prices for two different products. But it's difficult to separate the market like this without just making unusable trash. The one separation we do have - DDR5 vs HBM - works against the consumer's favor. One layer down in the supply chain, RAM manufacturers have to buy silicon wafers to etch circuits into, and HBM takes up treble more wafer space than DDR5 per gigabyte. In fact, this is specifically the scarcity that allowed Sam Altman to engineer a RAM crisis, by buying up a bunch of wafer supply that he could then redirect to more AI-optimized HBM.

We can also infer from all of this that CXMT - while producing lots of memory - does not have the scale necessary to actually quench the inflated demand. Either because their yields are shit, or because China wants to run a business and made the same calculation as the other RAM producers that the AI bubble would pop before they could get additional equipment online to service demand.

I know this because BP coined the term "personal carbon footprint" in order to tell well-meaning liberals like me that we were personally responsible for global warming, and not the oil companies spending to make sure the only practical mode of transportation was individual vehicles burning petrochemicals for energy. Or, prior to that, the oil companies giving Greenpeace money to go heckle nuclear energy and set back the development of clean energy by decades[0].

Likewise, when we do eventually Nuremberg these oil fucks, they are going to throw this exact same argument back at us. They are going to argue that they were just responding to shareholder demand, that fiduciary duty in a for-profit corporation mandates CEOs and board members commit crimes against humanity if it would maximize profits. It doesn't, of course. It'd be insane to argue that corporations are legally obligated to break the law if it makes them a buck. Which, of course, means that there's already a historical record of Nazi-collaborating business owners who used the shareholder value argument specifically to get out of a hanging at Nuremberg. So let's not use that argument. The moral and legal culpability must be concentrated primarily in the people who had the power and made the decisions at the time, and not laundered through the mechanism of the market to an accountability sink.

Besides, are you really going to put everyone in America with a 401k on trial? That would actually be a really good pretext to putting America under martial law and stripping Americans of their rights[1].

[0] To be clear, PV solar is so cheap now it makes any other form of energy look expensive and useless - nuclear included. The point is specifically that we emitted 40 more years of carbon than we needed to because of this.

[1] Which, to be clear, is already happening, has been happening for decades, and is always happening unless you make continuous efforts to revalidate your rights through democratic means outside of merely voting.

The underlying issue is not if we can build the network (we can) but getting people to actually use it. For example, if we wanted to repeat the Japanese success of IC cards here, we'd run headfirst into the problem that our public transit networks aren't big enough to be a credible Visa/MC alternative:

- Most public transit systems large enough to issue their own fare cards and readers are also regional monopolies, because there's not enough transit demand to sustain multiple companies with different routes. If I'm in Pittsburgh, all the buses and trolleys are run by PRT. In the Wasatch Front, it's UTA. Those agencies have little interest in becoming banks; they operate the fare cards mainly to keep fare payment easy, and they don't need to coordinate with anyone because everyone already joined into a single large transit agency.

- Metro areas with multiple transit agencies often have political differences that make coordination difficult. For example, in Long Island, NY, Nassau County's NICE bus system had legacy Metrocard fare payment that wasn't upgraded to OMNY until a month or two before Metrocard was completely ripped out and shut down. A rare exception to this would be the San Francisco Bay Area where there is an insane amount of political fragmentation and somehow they all wound up taking Clipper.

- A lot of Americans just never touch trains or buses enough to actually need a transit card.

- A lot of transit agencies are just surrendering to Visa & Mastercard and taking credit cards now anyway, even though transit fares are exactly the kind of microtransaction that is ill-suited for those networks.

Ironically, the best bet for an American-run payment network would actually be to nationalize E-ZPass[0]. In fact, in 2008 Congress passed a bill specifically mandating a unified toll payment system, but nothing came of it because the bill had no actual teeth. The main problem with this idea is that E-ZPass transponders won't fit in your wallet; you'd have to launch a separate form factor for an "E-ZPass Card" and at that point you run into all the same problems I just mentioned with making a unified transit fare system.

[0] E-ZPass is an RFID transponder system for toll payment that is very well-adopted along the east coast.

I mean, every time ISPs, NAT, or IPv6 is mentioned you have a LOT of people who are really angry they can't just netcat a random port on their friends' machine to send files to it.

At inference time, there's a whole new legal question: are the specific outputs of the model substantially similar to anything that the LLM had access to? Even if AI training itself was infringement, a specific generated output of that model could be noninfringing anyway. And vice versa, AI training being fair use doesn't form a blanket defense against a specific AI copyright infringement. If you ask the model to create an art program, it could still generate something substantially similar to the code of an existing art program.

In the case of regurgitation[0], the fact that an LLM was involved is immaterial. The only thing the law cares about is if a copy was made somewhere along the chain and if the alleged copy was substantially similar to the purported original. The only thing LLMs do for this analysis is make access to the original far easier to allege, because LLMs are trained on everything.

[0] Inclusive of both model weights and context regurgitation

Now that being said, Adobe has plenty of non-standardized file formats that they absolutely could do AutoCAD-style shenanigans with.

As seen on the main Artcraft repo. There's other tools that are MIT-licensed, although given that people are noting that this is AI-created the notion of copyright on this is suspect to begin with. Remember kids: all AI slop is born public domain because machines don't have a soul for the purpose of copyright law. And probably also the 13th amendment. Don't upload your brain, kids.

Anyway, let me get up on my soapbox and talk about these "have your cake and eat it too" licenses, that try to be source-available while privileging one particular vendor as the Blessed Owner with special monetization powers. To put it shortly, this doesn't work, because it mixes all the bad parts of FOSS with all the bad parts of proprietary software.

The business model of proprietary software does not work with even source-available terms. Proprietary software is a game of becoming essential to your customers and then taking them for everything they've got. All the successful proprietary companies are, effectively, selling legacy software that was useful to a group of companies a long time ago, but is now so deeply integrated into their IT systems they don't even know they have it until the licensing police show up. "You can use it for free but can't sell it" is anathema to this: it is specifically the "use" part that you're worried about.

Likewise, being able to modify the software - even only for internal uses - means that the proprietary owner has to compete against their own customers for ongoing payments and revenue. Again, if you've dug your talons into a big org to the point where you're indispensable, you don't want your customers making their own forks of your code. You want them paying $$$$/mo for enterprise support.

What FOSS is doing is basically promising to never do these things. If your goal is to make money, then adopting FOSS is like a cat going vegan. And the half-measure source-available licenses don't help, because you wind up conceding all the things that make you money just so you can claim ownership over things most people weren't interested in.

also, from their CLAUDE.md:

> Clean-room. Never read, disassemble or copy anything inside the Illustrator bundle (names/listings only). Never copy Adobe icons, artwork, presets or wording beyond feature names. Behaviour comes from public docs and black-box observation of the running app with synthetic documents only (screenshots by window id, stored under plan/illustrator/screenshots/, never committed). Never copy GPL/AGPL code (Inkscape, lib2geom…).

There is no such thing as clean-room vibe coding. Every line of FOSS code has been trained on. If you use any AI coding model, you are generating code that is going to be similar to copylefted code. There is no "MIT/Apache only" model out there you can use.

Also, no, switching to GPL doesn't help. Actually, it is my opinion that you cannot meaningfully comply with GPL while vibe-coding, because you have changed the definition of "corresponding source code" from a bunch of .rs files to a number of prompts being handed off to a cloud model. Remember: according to GPL, the source code is not "the thing you put into the compiler", it's "the thing the developers prefer to edit to alter the program". Otherwise, you could say "the source code is this uncommented ASM listing" and sidestep the GPL entirely. What we're doing with AI is the same idea, except instead of going down one level of abstraction, we're going up one.

1. Normal people assumed the CDC et all would contain the outbreak early, or that it would burn out, like what happened with SARS

2. World leaders brushed it off for a variety of subreasons[0] interesting to political scientists but, for the purposes of this discussion, all boil down to "but I don't WAAANA contain a pandemic."

The underlying problem is that in order for humanity to actually deal with a catastrophic risk, the risk needs to be both plausible enough to the average person as well as have a solution whose costs are not too high. For COVID, by the time the risk was clearly known, the cost to contain it was "refrain from human socialization and remain at home for an indeterminate amount of time plugged into the Metaverse™".

Now, let's look at AI extinction risks:

1. People are aware of them (I've watched Terminator!) and the risks are plausible. However, the connection to currently existing AI is not. As far as the general public is aware, AI is that thing that tells them to eat rocks when they Google old The Onion stories and floods their social media timelines with realistic-looking pictures of Shrimp Jesus.

2. The purported solutions to extinction risks require extreme concentrations of power: you need national control of AI research, bans on large GPU deployments, bans on training on publicly-available copyrighted data, some kind of military effort to render Chinese AI labs inert or dead, etc. Some of these may be attractive to some people[1] but the whole package taken together seems like an obvious power grab, if not outright invocation of other non-AI extinction risks. Like, at some point, if the AI wants to kill us, it just has to nuke its own data centers (or the data centers hosting a competing model) and hope the old Cold War nuclear retaliation systems take the bait.

If someone said, "Hey, your guinea pig or pet rat is going to eat you tomorrow unless you engineer a pathogen that eradicates all rodents from this planet and inject it inside yourself", you probably would tell them to pound sand, even if it is at least theoretically plausible that such a thing would come to pass.

[0] Xi Jinping censored initial discussion of the pandemic as fake news. Donald Trump thought it was going to only affect China. California and the UK Tories were partying in violation of their own lockdown rules. Japan took the excuse to shut down tourism for three years and massively restrict immigration but was, from what I'm told, constitutionally prohibited from implementing any domestic lockdown rules.

[1] I personally would like to see a moratorium on new data centers and an explicit revocation of the EU Text and Data Mining copyright exception

I mean, think about it: what would it mean if ls had a separate sandbox identity from the shell that spawned it? It would mean that any process on the system not entitled to read files from disk could get that entitlement by just fork/execing ls and parsing its stdout. That's not a security boundary that makes sense. ls doesn't read files - ghostty reads files, and ls is just its deputy.