2026-05-20: sovereign compute, image tools, decision theory, future generations

2026-05-20 — 2026-05-20

Six new posts, eighteen updates, fifteen days — and no, he hasn’t finished anything from last month, since you were about to ask. The thread running through most of the new work is who gets left out: future generations can’t retaliate against decisions made before they’re born, small community groups can’t own their AI without some American corporation holding the keys, and if a predictor’s already worked out your choice before you sit down, the usual rules for deciding fall apart. There’s also a solid run of practical stuff on AI image tools — which programs are worth your time, which ones are already being swallowed by the all-in-one options, and why one technical decision you make before you even open the app matters more than any of the fancy front ends. Dan’s been at it from six directions at once, which is very him.

digest

1 Whose AI is it anyway

1.1 DIY Generative AI

Dan’s mapped out what actually goes into making one of these big AI models — the computing power, the training data, the clever maths, and the trained model itself — and had a look at which bits the open-source mob have actually managed to get their hands on. Short answer: running an existing model on your laptop is dead easy these days, but if you want to replicate how a frontier model was actually built from scratch, you’re out of luck, because the big labs only hand over the finished thing, not the recipe. Like always getting a lovely roast chook from Woolies for dinner, with no idea how to roast one yourself, let along keep one alive in the coop out the back. Then a some Chinese labs — starting with DeepSeek and Alibaba — came along and released not just competitive models but proper detailed notes on how they made them, which put a hole in the comfortable assumption that open-source is always running a year or two behind the well-funded operations. Turns out it only takes one mob actually sharing the recipe to make everyone else look stingy.

1.2 Community sovereign AI compute

A footy club, a co-op, a school parents committee — Dan reckons any group willing to chip in could buy their own AI hardware outright, instead of renting it off some American megacorp that can change the rules or cut you off whenever Washington gets twitchy. Fair enough point. What’s new is there’s been a bit of a breakthrough in how these models handle memory during long conversations — turns out they can now juggle a lot more people at once on the same machine, which makes the whole scheme a damn sight cheaper than it looked six months ago. The idea is you own the thing, it sits in your shed or wherever, and nobody can throttle it, censor it, or jack up the price because their shareholders are having a moment. Dan’s written it up so other groups can copy the playbook — because apparently one semi-functional community AI wasn’t enough, he wants to spread the chaos nationwide.

1.3 Sovereign compute for small collectives: a technical implementation guide

The why’s mostly covered above — this one’s the nuts and bolts, and the main news is the case got a fair bit cheaper. The big practical problem was always long conversations eating the machine’s memory — a couple of people in a deep back-and-forth could chew through the whole thing, leaving everyone else waiting. A new model called DeepSeek V4 uses a cleverer way of handling memory that cuts that consumption down by about ninety percent, which basically means the box can handle ten times as many of those long sessions before it falls over. Dan’s also done the decent thing and added a section admitting what still doesn’t work: fitting more into memory isn’t the same as going faster, and stripping out the model’s built-in political guardrails — which is kind of the whole point for a group wanting to own its own tools — is still a rougher job with this one than the older model he’d been recommending.

1.4 Generative art with language+diffusion models

Dan’s properly down the rabbit hole on AI image generation — type what you want, machine draws it, simple enough until you ask who owns the result. He’s added a table comparing how Japan, Singapore, the EU, and Australia handle that question, and the short version is: nobody’s really settled it, and “I made it so it’s mine” is not the slam-dunk you’d hope. There’s also a decent bit on how the standard models are trained to refuse anything dark or distressing — built-in manners, basically — while the community-trained versions couldn’t give a stuff, which raises the obvious question of who decided what a computer’s allowed to imagine. His practical tip: keep your prompt logs and embed your metadata, because if it ever comes to a stoush, you’ll want receipts. If you want to know which software will actually let you run any of these specialist models, the next post has the dirty.

1.5 Front-end clients for AI image models

Boys and their gizmos eh? Dan’s had a proper play with all the software you’d use to run AI image generation on a Mac — Draw Things, ComfyUI, InvokeAI and a few others — checking which ones are a bastard to install, which models they’ll run, how much memory they hoover up, and whether they slip in content filters without asking you. No clear winner. Draw Things is the easiest to get going and supports the widest range of current models, ComfyUI gives you more knobs to fiddle with and can handle video too. Fine. Apparently Apple Silicon can run the same model in three different ways under the bonnet, and that choice — not which fancy app you went with — determines speed, memory use, and which models will even work at all.

1.6 Editing images with AI

All-in-one AI editors are now good enough that half the single-purpose tools are starting to look like overkill. Used to be you’d need three different programs just to fix a face, strip out a background, and sharpen the resolution; apparently now that all just happens as a side effect, which is either impressive or mildly terrifying, take your pick. He’s also tidied up the individual recommendations. I won’t bother you with the details. Half these tools have a shorter shelf life than milk; Dan’s at least keeping score of what’s still worth grabbing.

2 Deciding in the dark

2.1 Markov decision problems

Dan’s been building out his post on making decisions when you can’t see the full picture — runs a beer warehouse through the whole thing as his worked example, which, fair enough, that’s where most decisions happen. The bit worth stopping at: all the tidy maths for planning ahead only works if you’re trying to do well on average. Switch your goal to “don’t get wiped out” — minimise the worst case instead of maximising the average — and the whole machinery falls over. Most of the theory just assumes you’re chasing averages and never gets around to mentioning it can’t handle the other thing. He’s also added a proper explanation of what goes wrong when your picture of the world is too thin — miss enough detail in your description of the current situation, and history starts mattering again, which is exactly what the Markov idea was supposed to get rid of.

2.2 Cosmic decision theories

There’s a classic puzzle in decision theory where a predictor has already worked out what you’re going to choose — and is almost never wrong — before you even sit down at the table. Two schools of thought have been bickering about it for decades: one says only worry about what your action directly causes out in the world, the other says your choice also tells you something about who you are and what the predictor already knew, so that counts too. Both camps have built enough elaborate theory to make your eyes water. Dan’s not convinced you need any of it — he reckons he can work through the same ideas with more sensible tools before going anywhere near the philosophical deep end. Knowing Dan, “before going near the deep end” is probably his way of saying he’ll spend six months in the shallow end and then call it done.

2.3 Decision theory in mechanized causal graphs

Normal causal diagrams are fine for showing that one thing leads to another — rain causes mud, demand causes price hikes, you know the drill. But they choke when someone’s predicting your behaviour not by watching what you do, but by reading how you tend to decide. That’s exactly what makes Newcomb’s problem such a headache: the predictor already filled the boxes before you walked in, based on the kind of decision-maker you are, not the actual choice you made. Dan’s fix is to add what some folks calls “mechanism nodes” — extra bits on the diagram that represent your decision-making process itself, not just the outcome it produces. Draw a line from your policy to the predictor’s policy, and suddenly the paradox looks a lot less paradoxical. It’s still a bloody weird thought experiment, but at least now you can see why it’s weird instead of just staring at it like a stunned mullet.

2.4 Causality, agency, decisions, learning

Dan’s had a proper tidy here, and the interesting bit is something he calls “mechanization” — instead of just drawing arrows for what caused what, you give every variable its own little node for the rule that generates it. For decisions, that means you’re not drawing “Dan chose to buy a tractor” — you’re drawing “the rule Dan uses for tractor decisions” as a thing in the diagram. Sounds like splitting hairs, but it’s not: once you can put a decision rule in the diagram as its own node, you can represent a predictor who’s already read how you make choices before you’ve made any. That’s the thing that breaks ordinary decision theory, and it’s what makes the weird Newcomb-style problems actually pinnable down rather than just argued about in circles. The old tangle of sections on causal-versus-evidential decision theory has been cut and sent off to its own page — that’s the Cosmic decision theories post a couple above, if you want to see those two camps properly argue it out.

2.5 Computational complexity of Bayesian inference

Bayesian inference is just the business of updating what you believe once the evidence comes in — but working out the actual number turns out to be a stone-cold nightmare. Exact answers are NP-hard, which is the maths way of saying “probably impossible in any useful amount of time.” Right, you think, just use a close enough approximation. Also NP-hard. A relative approximation? That lands in something called #P-hard, which is a whole level worse again — think of it as the same problem but someone’s added a locked gate at the front. The one escape route is keeping your probabilities away from the extremes — nothing too close to zero or one. The catch is that extreme probabilities are exactly how hard problems sneak inside a model and hide in the first place. So the only exit is also where the trouble gets in. Dan seems a bit too cheerful about all this for my liking, but that’s Dan.

2.6 Nearly sufficient statistics and information bottlenecks

Turns out there’s always a perfect summary of your data — it just has to be a full probability distribution, not a tidy number. Fine in theory, bloody useless if you haven’t got infinite memory. The funny bit is that statisticians, control engineers, and the machine learning lot all figured this out separately, gave it different names, and spent decades acting like they’d invented something new — Dan’s stuck a table in there showing they’re all talking about the same thing. If you read the Markov decision post above, those belief states the beer warehouse keeps track of? Exactly this object, showing up under yet another name. The information bottleneck stuff that follows is basically asking: if you can’t carry the whole distribution around with you, how close can you get without losing the plot? Worth a look if you’ve ever wondered why your model seems to forget things it should’ve kept hold of.

3 Everyone who can’t answer back

3.1 Intergenerational game theory

Standard game theory keeps people honest the same way a grudge does — you screw me, I’ll remember. But future generations can’t do that. They weren’t alive for the decisions, they can’t refuse to inherit whatever mess we leave, and they’ve got no seat at any table. So the whole mechanism that makes cooperation work just bloody falls apart when the people most affected can’t answer back at all. Dan’s gone through the various patches people have tried: chaining generations together like links in a fence so at least your grandkids’ generation can push back on yours, asking what future people would veto if you gave them a hypothetical say, or just appointing some poor bastard whose actual job is to sit in a room and pretend to speak for the unborn. None of it’s entirely satisfying. Turns out it’s hard to represent people who don’t exist yet, which — now that I say it out loud — probably should’ve been obvious from the start.

3.2 Civic technology

Most of what we call democracy is just counting hands — which is fine right up until half the room can’t agree on what’s actually happening, let alone what to do about it. Dan’s drawn a cleaner line here between three different problems: what people want, what people think is true, and the flat-out hostility that makes either kind of conversation impossible before it starts. The bit worth actually sitting with is the stuff on Community Notes and Polis — instead of asking “do most people agree?”, they look for “do people who normally can’t stand each other agree?” That’s a much smarter question. The majority saying something and people-across-all-the-usual-divides saying something turn out to be very different things, and we’ve been treating them as the same for decades. Democracy’s got a maths problem, is the thing. Shame the people running it generally can’t count. This fits pretty neatly with the intergenerational post above — both are really about who gets left outside the room when the decisions get made.

3.3 Institutions for angels

Dan’s been thinking about why open-membership groups — the kind where anyone can rock up and join — keep getting taken over by their most extreme members. You’ve seen it: five years after founding, the committee’s a different species from the general membership. Turns out it’s not bad luck. Open groups attract exactly the people who care most, the regulars gradually crowd out the casuals, and your public spokesperson ends up being whoever frightens journalists the most. The new bit adds the reasons why groups stay open in the first place: it’s easier than vetting people, you need the raw numbers, and there’s this touching faith that anyone who shares your values will also behave themselves. That last assumption does a lot of damage. Dan’s also fair enough to note that whether this looks like a disaster or a triumph depends entirely on who you ask — the mob who just took over think it’s going beautifully.

3.4 Metis and .*-rationality

The German foresters thought they had it sorted. Strip out the messy undergrowth, plant the trees in neat rows, count everything that could be counted — first-generation timber yields were excellent. Second generation, the whole thing fell over. Turns out forests run on all sorts of complicated stuff that doesn’t fit in a ledger: fungi, leaf litter, tangled relationships between things nobody’d bothered to name. Dan’s been reading James C. Scott on exactly this — the idea that once you tidy reality into something a bureaucrat can measure and manage, you’ve already thrown out half of what made it work. Scott calls that lost knowledge metis: practical, local, hard-won, and completely impossible to write up in a policy manual. The bit Dan’s added: it’s not just top-down planning that does this. Markets probably pull the same trick, just with a different kind of clipboard.

4 The slush pile

4.1 Vibecoding Apple apps

Here’s a thing I didn’t expect to be interesting: Dan’s set up a way to build real iPhone apps without writing a single line of Swift himself. He just tells Claude what he wants, and Claude does the actual coding. Two tools handle the plumbing — one runs the build, checks if it compiles, and tells you when it’s broken, while the other peers into the open Xcode window and grabs live error messages and screenshots of what the interface actually looks like. That means Claude can see what went wrong and fix it, rather than relying on the human to ferry error messages back and forth like some kind of unpaid secretary. He’s also added a note about Xcode eating your hard drive whole — which, fair enough, that’s the sort of thing you want to know before it’s too late. Whether this counts as programming or just very bossy dictation, I genuinely can’t tell you.

4.2 Generative AI workflows and hacks 2026

The bit here that’d make me sit up: if you run the same AI model through Ollama versus the Hugging Face setup, you don’t always get the same results back. Turns out the two systems have their own separate ways of chopping your text into pieces before it ever reaches the model — and when your input’s got markdown or anything a bit untidy in it, they go their own ways. Quietly. No error, no warning, nothing. Same model, same input, different answer depending on which door you walked in. That’s the kind of thing that’ll send you absolutely spare for a week before you even think to look there. Dan’s also rounded out the page with notes on PDF tools, a workaround for running Claude Code without Google Chrome — Anthropic decided Chrome is the only browser that exists, which is very much their problem becoming yours, and there’s more on getting around that in the next post — and a reminder to stop your AI model files from eating your entire backup storage. That last one sounds obvious until it isn’t.

4.3 Chromium browsers

Anthropic’s decided Claude’s browser integration only works in Google Chrome, full stop, and they’re not budging on that. Dan’s added a section on how to get around it if you’re using one of those Chrome-adjacent browsers — Vivaldi, Brave, and that lot — because apparently the official answer of “just use Chrome” wasn’t cutting it. There’s a community workaround, but he reckons nobody’s vouching for it with their hand on their heart. He’s also written more about Vivaldi itself — which he clearly has a soft spot for despite it being, by his own admission, a bit bloody buggy — and the tab-stacking feature that sounds clever in theory and drives him up the wall in practice. Classic Dan: finds something with good bones, falls for it anyway, then writes three paragraphs about why it frustrates him.

4.5 Care and feeding of macOS filesystems

Dan’s tidied up his notes on why copying files between a Mac and anything else is a special kind of headache. The short version: Apple’s filesystem doesn’t care whether your filename is uppercase or lowercase — “Budget.xlsx” and “budget.xlsx” are the same file, far as it’s concerned. Most other systems aren’t so relaxed, so things go sideways when you try to move files across. He’s also clocked that the rsync Apple ships is a neutered version — like getting a ute with the tray welded shut — and the fix is to install the real one yourself. It’s not glamorous stuff, but if you’ve ever had files vanish or duplicate themselves on transfer, now you know it wasn’t you being an idiot.

4.6 Quarto

Quarto’s what Dan uses to write all this stuff up — mixes code and text in the one document, spits out a proper webpage. He’s been using it every single day for four years, which for Dan is basically a world record. The update worth knowing: he’s finally written down that if you try to fiddle with how Quarto does things, it’s a bigger pain in the arse than it looks. Took him four years of daily use to mention that, but here we are. There’s also a note that Quarto now experimentally lets you plug in your own custom engines for running code — if that means anything to you, there are links to the developer docs; if it doesn’t, don’t worry about it.

4.7 TikZ/PGFplots etx

Someone’s gone and compiled TikZ to run in a web browser — the program Dan uses for all those neat boxes-and-arrows diagrams — which means you no longer need a full LaTeX installation just to draw a line with an arrowhead. Used to be a solid hour of faffing before you’d drawn a single thing; now the tradeoff is your page has to haul in a few extra megabytes when it loads. So you’ve traded a complicated install for a slightly chunky download. Same diagrams, less ceremony.

4.8 Ageing

Dan’s sorted the anti-ageing supplements into three piles: worth it, borderline, and don’t bother. Someone’s gotta do it, now that everyone’s flogging you something on their podcast to make you live forever without your soft bits sagging. The sad truth is that exercise beats every supplement on the planet by such a margin it’s almost embarrassing. But creatine, vitamin D, and a couple of cheap others make the “worth it” cut because the evidence is decent enough and they won’t cost you much to find out. He’s also dialled back a few of his bolder claims — metformin now “probably” cuts into your exercise benefits rather than definitely does, and one supplement’s animal research got demoted from solid to “credible-ish.” Good. Knowing what you’re not sure about is half the job, and it’s a damn sight more honest than most people who write about this stuff.