💀👻🎃🦇MOSS 666🦇🎃👻💀

Retrocomputing Maker and Designer in Ontario, Canada. I make stuff without promising a delivery date. I don't do crowdfunding, but I do gratefully accept Patreon and Ko-Fi support for what I do. 

Buy me a coffee!: https://www.ko-fi.com/mos_8502
Become a Patron!: https://patreon.com/Studio8502
Buy my games!: https://studio8502.itch.io/

Public Key

npub1qnufz4pycufk27kkeev5g0fgm0w0tqexsly674s2uwy0tynk5ymsk5hzwk

NIP-05 Address

mos_8502_at_studio8502.ca@momostr.pink

Profile Code

nprofile1qqsqf7y32sjvwym90ttvuk2y855dhh84svng0jd02c9w8z84jfm2zdcprpmhxue69uhhyetvv9ujumt0d4hhxarj9ecxjmntmzq8k6

Publishing to

relay.momostr.pink

Author Public Key

npub1qnufz4pycufk27kkeev5g0fgm0w0tqexsly674s2uwy0tynk5ymsk5hzwk

Show more details

Last Notes

2026-10-09 22:36:30 UTC

I need you to break your existing habits and talk to some non techie people about AI. Not to argue for or against, just to suss out how they feel, if they feel anything.
I cannot stress to you enough how little of a fuck the normies in my slice of the world care one way or the other about AI.

2026-10-09 21:45:39 UTC

Oh, if you thought you didn’t like The 8-Bit Guy before today…

2026-10-09 19:19:36 UTC

- reply

China is using the AI boom to push investment in Chinese-made silicon as well, getting longer term returns out of it in terms of domestic tech manufacturing. Their ideal future has them exporting Loongarch APUs with massive unified RAM that can run local models for chump change off wall power.

2026-10-09 19:17:20 UTC

- reply

It's important to separate the American AI industry, which is a Ponzi scheme, from the Chinese industry, which is part of the CCP's tech strategy, treated as national infrastructure investment. Companies in China are much more restricted in what they can and can't get away with.

2026-10-09 19:09:11 UTC

- reply

Giving every session a 1,000,000 token context window (which you really need for agentic work) really motivated a lot of this. To take on the big boys with less money, they have to be more efficient and a lot cheaper, and so far so good.

2026-10-09 19:05:42 UTC

- reply

Model distillation is a thing, but not nearly as important as it's made out to be. A lot of the efficiency gains have to do with the KV cache -- every user can share the model weights, which are read-only, but every user needs their own KV cache, so making it smaller and faster means more sessions per unit of hardware and power and water.

2026-10-09 19:03:02 UTC

- reply

Long term, I hold out hope that I'll be able to run a model trained on pure MIT/BSD/PD/CC0 licensed code and prose, on local hardware, using the clean renewable energy that powers my city already.

2026-10-09 19:01:43 UTC

- reply

I focus on using the most efficient models available, because by definition they are using less valuable power and water per million tokens. DeepSeek has done remarkable work making the same hardware support more users -- v1 could do 90 sessions on a single server, v4 does over 400 simultaneous users with a bigger, better model. And there are still efficiency gains to be had, according to the latest research papers.

2026-10-09 18:58:44 UTC

- reply

Oh, that's a big problem, and one in dire need of solving with regulation. The political climate right now is allowing them to do as they please, but the next election may bring a sea change there. We can hope.
I think they should be required to generate their own renewable power and recycle their cooling water.

2026-10-09 18:57:29 UTC

- reply

I walked away from opium, I think I can handle a language model.

2026-10-09 18:56:10 UTC

- reply

If I lost access tomorrow for some reason, I have other shit I can productively do with my time. The tech is a tool to me, a tool to do something good for the world -- but there are plenty of other ways, if the way I'm currently enjoying ceases to be.

2026-10-09 18:51:04 UTC

- reply

Or like take the released source code for MS-DOS and update it to be 6.22 compatible without MS's help.

2026-10-09 18:50:01 UTC

- reply

There's something vaguely Japanese-feeling about that. Like how "Pasocom'" means "Personal Computer".

2026-10-09 18:46:23 UTC

- reply

Imagine if we could, say, make a 100% compatible OS/2 kernel based on the documentation and headers. That'd be neat, wouldn't it?

2026-10-09 18:45:53 UTC

- reply

Like... think about ancient codebases for which we have only partial source code, or the source is in some obscure version of a specific language we don't have a good compiler for. The machine could be used to reconstruct the missing bits (with acknowledged limitations) to give something like the original back to the world, or translate the source into a more common dialect.

2026-10-09 18:42:27 UTC

- reply

Intuitively, that feels like you were talking about UTF-8 text? I don't know Esperanto.

2026-10-09 18:41:04 UTC

- reply

I understand that. I do get it. It's origins are anything but immaculate. But it stubbornly refuses to disappear from the world, and my own convictions demand that I either find some prosocial use for it or try to stop it being used at all, and we all know how well the latter strategy works against the drug problem.

2026-10-09 18:36:13 UTC

- reply

Might be fun if you have a real VT100 terminal kicking around the crib.

2026-10-09 18:35:36 UTC

- reply

And there's NOTHING anyone can say to convince me it's sentient, or my friend, or has feelings, or "escaped containment" or any of that load of old bollocks. It's math and code and data, nothing more. Anyone who thinks otherwise needs their head examined.

2026-10-09 18:33:40 UTC

- reply

I get it, I do. I'm not *reluctant*, I'm firmly in the anti- camp as far as education goes. It's flatly, demonstrably bad for it. I'm also against using image models, music models, etc. I'm just convinced there's utility in this one small area of life, under certain restricted conditions. The best I can do is "not completely without utility, but not worth what the big evil players are charging".

2026-10-09 18:31:17 UTC

- reply

Just HTML 5 and CSS. No JS. I hate JavaScript. I hated it when it was new.

2026-10-09 18:23:06 UTC

I have no crisis. I have no big fucking mess. I haven't got anything major going on. There's nothing to light your hair on fire about. I cannot stress this enough: I'm fine. Nothing is wrong.
But if you felt like buying me a pizza, it'd sure be a nice treat today.
https://www.ko-fi.com/mos_8502

2026-10-09 18:07:37 UTC

Still not a web designer. But I can read a CSS handbook and I know typography basics.
https://studio8502.files.fedi.monster/media_attachments/files/117/412/281/540/144/153/original/5fdc5141971b8706.png

2026-10-09 17:25:31 UTC

- reply

Yes, I told him to avoid LLMs. Of course I did. That's the responsible thing. No learner should use them. They demonstrably interfere in learning skills. It'd be irresponsible to advise a child to use such a thing.

2026-10-09 17:23:47 UTC

- reply

Sure, I have no problem with that. We should be responsible for what we create.

2026-10-09 17:21:34 UTC

- reply

I don't disagree -- but here we are gaming the most optimistic version, for the sake of argument.

2026-10-09 17:21:00 UTC

- reply

I told him, read read read. Read books, read code, ask questions. Keep at it.

2026-10-09 17:17:17 UTC

Young (14) programmer comes to me on Discord wanting hints for his toy OS today.
So I tell him to avoid LLMs while he's learning, and gave him links to PDFs of K&R C, 21st Century C, Gang of Four, and Operating Systems Design and Implementation. All the best books for young aspiring OS coders.

2026-10-09 17:07:53 UTC

- reply

Agreed without reservation. I want someone to take all the permissive-licensed code out there (MIT/BSD/etc) and use that for training, since the license terms allow *any* use.

2026-10-09 16:59:19 UTC

- reply

People think I'm nuts for using agentic coding at all. Maybe they're right. I don't feel crazy, but then if I were I probably wouldn't anyway.
But the argument over whether or not agentic coding *can* work is *over*. It bloody well can. Now it's on us to figure out how to make it work *well*.

2026-10-09 16:49:00 UTC

- reply

I think Linus is nuts for using it in the kernel. I wouldn’t. I’d make a separate agentic fork and confine it there to let it prove itself. Sanity seems rare in this field.

2026-10-09 16:44:48 UTC

- reply

Personally, I recommend not adopting agentic coding in any established project that people rely upon. Instead, use it to make new experimental stuff and master the tooling, explore things, maybe make something neat without risking your existing codebase.

2026-10-09 16:41:14 UTC

- reply

Not that one. No no no. Not even as a joke.

2026-10-09 16:40:47 UTC

- reply

Theres a whole middleman service called OpenRouter that lets you seamlessly switch between providers and models with a single open API any agent harness can use. A great way to keep the costs down.

2026-10-09 16:39:41 UTC

- reply

Hell, all the really good models are open weights. Anyone can rent or buy hardware and spin up a service to sell compute at a markup. There’s no big secrets here. Everyone is publishing their findings in the open.

2026-10-09 16:38:20 UTC

- reply

What’s comparatively nice about the LLM as a service industry is that they’re selling a commodity. One recent model is about as good as another for coding. They distinguish themselves with things like efficiency and pricing rather than functionality at this point. So there is plenty of competition in the industry, nobody can hold a monopoly. It’s selling GPU compute at a markup, and the most efficient and reliable seller who can accept the lowest margins wins.

2026-10-09 16:33:57 UTC

- reply

It’s hard not to root for the side that fucks over Adobe. That’s all I’m really saying.

2026-10-09 16:32:38 UTC

- reply

It also represents the ultimate triumph of open software against proprietary. In that world, selling software is not economically viable.

2026-10-09 16:29:22 UTC

- reply

This might (might!) be the one thing that survives the inevitable bubble pop. The actual prosocial AI use case that sustains the industry at a reduced level.

2026-10-09 16:25:40 UTC

- reply

We are talking about a potential explosion in software diversity. Bespoke tooling generated on the fly. Projects that would never even be attempted before coming to fruition. Good and bad ideas brought to life. A torrent of absolute dreck and some potentially great stuff too.

2026-10-09 16:21:26 UTC

- reply

A lot of case law supports the rewrite side. But Adobe has a lot of cash. I imagine it’s in the interests of the AI companies to support the rewrite projects though.

2026-10-09 16:17:04 UTC

- reply

Do I think that’s the world we have, or will have? I don’t know. What am I, a psychic? All I do know is there is clearly something to all this agentic bollocks, and I mean to understand it at the same level I do C programming. I will not be denied knowledge in my field. Do I advocate it? No. I’m not saying it’s good or for everyone or even universally useful. I’m saying there’s something here, and I mean to grok it.

2026-10-09 16:14:44 UTC

- reply

The fundamental unit of trade in software has always been developer time. But in this world, the sheer scale of what developer time buys is dramatically greater. Small teams can do in weeks what used to take large teams years. That drives the cost of code way down to near zero.

2026-10-09 16:10:46 UTC

- reply

Hell, in that world the very notion of software copyright might as well vanish like a fart in a windstorm. What does it matter if you own Software X, when I can essentially duplicate it for myself without touching your stuff? You no longer have anything to sell me as far as software goes.

2026-10-09 16:08:09 UTC

It’s a fucked up world we live in. I think we can all agree there. But let’s just for the sake of it, game out the most optimistic version of what’s coming for software.
We are talking about a world where we can “just” vibe code up a functional clone of whatever software for whatever platform. What’s the upshot? Proprietary software is *dead* in that world. D E A D dead. That at least is kind of nice. In a world where a few hundred dollars in token spend buys you a functional clone of Microsoft Word or Adobe Acrobat, the economics of software have fundamentally changed.

2026-10-09 14:22:38 UTC

- reply

It just seemed like a good idea to make sure something could be salvaged in the event of total project failure.

2026-10-09 14:21:44 UTC

Interesting fact: as part of the test suite for libargentum’s Foundation, Core Graphics, and AppKit equivalents, they are built and tested both in the QEMU/AgOS environment and for the host Linux machine. So if AgOS dies, it’s no effort at all to shift focus to a Mac-like Linux/BSD desktop environment as a backup plan.

2026-10-09 14:14:26 UTC

- reply

I question how good and stable it can become, but I am willing to give them time and space to find out, if for no better reason than the only ones getting hurt are capitalist fuckstains.

2026-10-09 14:12:03 UTC

- reply

The goal in mind is laudable. Fuck over the big software houses who’ve fucked us over for years. I have no objection to that. It’s a prosocial use of AI.

2026-10-09 14:10:22 UTC

- reply

Apparently it actually works okay for an alpha and is improving. Good for them. Maybe in six months to a year it’ll be usable and stable enough for my tastes. I doubt it, but it could happen. Even so I am amused greatly by its mere existence.

2026-10-09 14:08:50 UTC

So that vibe coded Adobe clone suite. Anyone else have no desire to use it but think it’s hilarious that at least one evil corpo is getting screwed over?

2026-10-09 13:54:25 UTC

- reply

I say this not to convince any ai haters. They have good reasons. I say it in hopes that anyone considering using it goes into it with realistic expectations and not a head full of techbro nonsense about AGI this and Basilisk that.

2026-10-09 13:53:15 UTC

- reply

Agentic coding is not perfect, not for everyone, not a panacea. Coding is not solved. AI isn’t magic. It is potentially useful in specific circumstances. It must be used with care and attention or it will just produce useless slop.

2026-10-09 13:51:31 UTC

- reply

So far that has held. For example, a major bug I got out of the machine reversed the Y axis of the mouse pointer. That got fixed fast I can tell you.

2026-10-09 13:50:29 UTC

- reply

I accept that there will be bugs. That is a natural consequence of agentic coding. But hand authored code also has bugs. The important thing is that as bugs are found they get fixed.

2026-10-09 13:49:25 UTC

- reply

I go one class at a time. I implement that class based on the public API documentation available for the target version 10.6. I credit that for the fact that so far everything is actually working as intended.

2026-10-09 13:47:41 UTC

- reply

You may notice how relatively slowly I’m moving compared to other agentic projects. This is intentional. I take the time to test everything to at least some degree. What I do not do is hand over a high level design and let the machine grind on it for twelve hours and accept the output. I find that to be insanity.

2026-10-09 13:45:22 UTC

Okay. On to NSColor.

2026-10-09 13:08:13 UTC

- reply

See that kind of name would recontextualize the NS prefix on things.

2026-10-09 13:06:54 UTC

- reply

If I could I would name it “N—— System” where N is some non-“Next” word beginning with N, one or two syllables. “Nova System” is taken.

2026-10-09 13:05:04 UTC

- reply

Ideally it’d be named something that feels familiar, friendly, clean, bright, like striking a new path into the future.

2026-10-09 13:03:30 UTC

- reply

I’m still unsatisfied with the name “Argentum”. It’s interim because it has to be called something. But it doesn’t feel right.

2026-10-09 13:02:35 UTC

So many fucking classes. NSBezierWhatever this time.
But the important thing is that it’s working. It’s going to make my vision real.

2026-10-09 03:03:55 UTC

I want to sleep. The trazodone is not helping. [sigh] Might as well finish up NSView tonight.

2026-10-09 03:01:58 UTC

Today was Thursday, which means my 20 year old chonky void cat Crystal got to eat one (1) french fry from my dinner. It is her favourite day of the week for this reason.

2026-10-09 01:05:09 UTC

- reply

DeepSeek 4.1 Flash.

2026-10-09 01:04:36 UTC

- reply

I have a theory. As the context grows, it periodically gets compacted. Because the compaction is lossy, it adds entropy to the context. So as it gets compacted again and again, the entropy grows, until it reaches a point where what’s being fed into the model is no longer coherent enough to get decent output. So I periodically clear and reload the context as I work. It seems to solve the decoherence you observe.

2026-10-08 19:58:38 UTC

I can tell I’m done for the day because everything makes me think “why bother?”

2026-10-08 19:53:08 UTC

- reply

I imagine for some people that person is me.

2026-10-08 19:51:04 UTC

- reply

You ever run into that? Someone whose core point is good and right, but whose manner and erudition just makes you want to shit into their pinned open eyes?

2026-10-08 19:50:12 UTC

- reply

A lot of soft, liberal-coded language about energy and community and micro aggressions. Exactly the language used by everyone who ever fucked with my life.

2026-10-08 19:48:09 UTC

My wife is watching a webinar about fat acceptance and liberation. I’m supportive of the topic and thesis, but their execution makes me want to OD on ozempic and become a protein influencer.

2026-10-08 19:26:09 UTC

Christ web stuff depresses me.

2026-10-08 19:17:57 UTC

- reply

I would ask the LLM to make me a site template but I've seen the output and frankly the LLM should stick to library code where it's actually kind of okay at the job.

2026-10-08 19:16:06 UTC

Can you tell I am not a web designer?
https://studio8502.files.fedi.monster/media_attachments/files/117/406/889/623/069/495/original/72babb7da0b4b166.png

2026-10-08 18:55:54 UTC

Well. I guess I get to make a website. I haven't written any serious HTML in years.
Do we still use tables for layout?

2026-10-08 18:40:20 UTC

libargentum-core.so - AgCore - Foundation equivalentlibargentum-draw.so
- AgDraw - Core Graphics equivalentlibargentum-app.so
- AgApp - Application Kit equivalent ORlibargentum-ui.so
- AgUI - Application Kit/UIKit equivalent
Thoughts?

2026-10-08 18:18:22 UTC

- reply

For clarity's sake: the API in Argentum currently targets the version of Cocoa in Mac OS X 10.6 -- right about when things were at their best on the Mac, in my opinion. That's what an Argentum 1.0 release looks like -- a Snow Leopard quality release.

2026-10-08 18:16:36 UTC

- reply

I will post this mockup as many times as it takes to motivate me to keep fucking going.
https://studio8502.files.fedi.monster/media_attachments/files/117/406/654/250/770/508/original/ad36146f6ce92c87.png

2026-10-08 18:14:37 UTC

- reply

One day, this will all come together and look so fucking cool. That day will not be today. Today I'm chewing through AppKit classes and methods.

2026-10-08 18:08:11 UTC

It ain't much to look at, but it's an actual NSWindow drawing its own chrome with an actual NSView for a content view.
https://studio8502.files.fedi.monster/media_attachments/files/117/406/621/299/937/036/original/245aeca98d4fbed8.png

2026-10-08 15:36:32 UTC

Uncomfortable truth: 95% of code was slop years before LLMs were a thing.

2026-10-08 14:10:32 UTC

- reply

I am not sure corporate capture is a realistic possibility, given how much competition there is in the model space. What worries me is more that new learners won't learn as well because they'll start using LLMs too soon. I don't think deskilling is a universal side effect, but I do think that LLMs do interfere in the learning process, and no student should use them.

2026-10-08 13:57:14 UTC

AppKit just drew its first window. Nothing fancy, just a rectangle on the screen, but it works. I can build on this.

2026-10-08 12:58:07 UTC

Where is a good place to go to talk about AgOS, to people who would actually be interested in knowing about it? Is there a Discord or something with an audience who actually wants to hear it?

2026-10-08 12:56:09 UTC

"I see why so many people block you."
Actual thing someone said to me. Before blocking me.

2026-10-08 12:14:46 UTC

- reply

Really I just have a burning need for something prosocial to come from all this horrid mess and slop. I can’t live in a world where everything is shit forever.

2026-10-08 12:09:45 UTC

- reply

I wouldn’t say I like it exactly, but it is kind of working? Like, with caveats aplenty, but better than reasonable expectations?

2026-10-08 12:09:08 UTC

- reply

I’m a bit more willing to wade into the slop to get what I want, but I don’t expect anyone to join me who doesn’t want to. But I am objectively getting interesting results reimplementing Cocoa for my hobby OS.

2026-10-08 11:58:38 UTC

- reply

I envy them that. I’m doing my thing alone. Just me and DeepSeek.

2026-10-08 11:56:07 UTC

- reply

To be honest, I would have preferred LLMs never being invented. But if they’re going to exist, I guess we might as well get some neat toys out of it.

2026-10-08 11:50:04 UTC

Uncontroversial statement: coffee is good and being up at 0630 kind of sucks.

2026-10-08 11:45:44 UTC

- reply

I mean. Power to them if they can do it. By all means fuck Adobe over.

2026-10-08 11:42:18 UTC

- reply

A bunch of people are playing silly buggers thinking they’ve got a magic genie that can do anything, and a bunch more are screaming that everything they don’t like is fascist. I don’t know how to live in that world. I just want to make cool stuff and play with it, stuff that maybe helps someone else do cool stuff too.

2026-10-08 11:38:11 UTC

- reply

I see what I see: a useful tool that requires care and skill to use well, surrounded by a lot of people who are loudly wrong about it.

2026-10-08 11:36:52 UTC

- reply

They work better than expected. But I worry about maintenance.

2026-10-08 11:35:19 UTC

- reply

The truth is messy and complicated and nothing is as simple as “adobe is fucked” or “it’s all junk”. But nobody wants to hear that.

2026-10-08 11:34:15 UTC

- reply

Really, I think perhaps both extremes of the whole debate are being silly. On the one side you’ve got people way too confident in the machine’s ability to deliver whole applications. On the other you’ve got people who refuse to believe it can do anything productive at all. Both are wrong, and they’re both irritating me.

2026-10-08 11:31:21 UTC

- reply

Slowly but surely functionality builds up. That’s how it is working as well as it does.

2026-10-08 11:21:43 UTC

- reply

OOP with ObjC and Cocoa lends itself very well to this process of increments. Everything is small and encapsulated. No huge functions. Massive existing public documentation. It’s probably as good as it gets in slop world.